[{"content":"","date":"21 January 2027","externalUrl":null,"permalink":"/en/","section":"GETALP","summary":"","title":"GETALP","type":"page"},{"content":"Unless stated otherwise, seminars take place on Thursdays at 2 pm in the IMAG building. This page also announces the team’s PhD defenses.\n","date":"21 January 2027","externalUrl":null,"permalink":"/en/seminars/","section":"Team seminars","summary":"","title":"Team seminars","type":"seminars"},{"content":"Le titre et le résumé seront annoncés prochainement.\n","date":"21 January 2027","externalUrl":null,"permalink":"/en/seminars/2027-01-21-herron/","section":"Team seminars","summary":"","title":"Titre à venir","type":"seminars"},{"content":"The title and abstract will be announced soon.\n","date":"14 January 2027","externalUrl":null,"permalink":"/en/seminars/2027-01-14-masson/","section":"Team seminars","summary":"","title":"Title to be announced","type":"seminars"},{"content":"Thesis conducted at the Laboratoire d’Informatique de Grenoble (GETALP and SIGMA teams), supervised by François Portet, Cyril Labbé and Frédérique Bordignon. The defense will take place in English; the presentation will take 45 minutes. It will be followed by a reception in the Magellan room, in the same building.\nJury # François Portet, Professor, Université Grenoble Alpes, Thesis Director Cyril Labbé, Professor, Université Grenoble Alpes, Thesis Co-director Konstantin Todorov, Associate Professor, Université de Montpellier, Reviewer Iana Atanassova, Professor, Université Marie et Louis Pasteur, Reviewer Jodi Schneider, Associate Professor, University of Wisconsin–Madison iSchool, Examiner Boris Barbour, Research Director, CNRS, Délégation Paris Centre, Examiner Didier Schwab, Professor, Université Grenoble Alpes, Examiner Guests:\nJennifer Wright, PhD, Cambridge University Press Frédérique Bordignon, Research Engineer, École nationale des ponts et chaussées, Thesis Co-supervisor Abstract # The scientific record is dynamic, constantly growing and expected to be corrected when errors are found. However, the volume of problematic scientific publications is growing faster than formal correction mechanisms (e.g. corrections and retractions) or community-driven efforts, such as post-publication peer review (PPPR), can keep pace with. We address this problem by assisting PPPR with natural language processing (NLP), investigations of the scientific record and standardized guidelines. We pursue these fronts through three contributions. First, we investigate how problematic content enters the record in the first place. One such example of this is paper mills, businesses that sell authorship of scientific papers, often advertised through social media. Through a systematic investigation, we trace these offers to their resulting publications in conference proceedings and document recurring red flags in the literature. In doing so, we advance the knowledge on paper mill activity in conference proceedings, which is a venue that remains under-studied compared to journals. Second, we address a gap in existing PPPR tools, which largely rely on bibliographic metadata or pre-defined problematic textual content alone when re-assessing publications. We present CiteScreener, a decision support system that combines bibliographic and retraction metadata with NLP-based citation-context classification to help PPPR practitioners (regardless of their expertise level) identify problematic citations. CiteScreener\u0026rsquo;s web interface has been evaluated by PPPR practitioners with varying levels of experience, who found it useful not only for surfacing problematic citations but also for identifying issues beyond its intended scope. Third, motivated by a recurring finding that practitioners need more guidance in PPPR, we present the Collection of Open Science Integrity Guides (COSIG). COSIG is an open-source project that curates and standardizes expert PPPR knowledge into an accessible source. Its aim is to reduce the barriers to entry that currently keep effective PPPR participation concentrated among a small number of experienced practitioners. Together, these contributions address the scientific record\u0026rsquo;s correction problem from three complementary angles: understanding the problematic content, equipping practitioners with tools to identify such content, and curating the knowledge needed to act on it.\n","date":"19 November 2026","externalUrl":null,"permalink":"/en/seminars/2026-11-19-soutenance-yagmur-ozturk/","section":"Team seminars","summary":"","title":"Supporting post-publication peer review: Empirical investigations, computational methods and guidelines","type":"seminars"},{"content":"Original title: L\u0026rsquo;ordre genré du journalisme. Une approche mixte et computationnelle des inégalités de genre dans les contenus de presse écrite française.\nPhD thesis in sociology, supervised by Gilles Bastin (PACTE laboratory) and François Portet (Laboratoire d\u0026rsquo;Informatique de Grenoble). The defense will be held in French. A reception will follow in the building\u0026rsquo;s hall.\nThe defense will also be streamed online. To get the link, and to confirm your attendance (on site or remotely) and/or at the reception, please fill in this form.\nJury # Florence Le Cam, Professor, Université libre de Bruxelles (Reviewer) Étienne Ollion, CNRS Research Director, École polytechnique (Reviewer) Roselyne Ringoot, Professor, Université Grenoble Alpes Aurélie Olivesi, Associate Professor, Université Lyon 1 David Doukhan, Researcher, Institut national de l\u0026rsquo;audiovisuel Gilles Bastin, Professor, Université Grenoble Alpes (Thesis Director) François Portet, Professor, Université Grenoble Alpes (Invited Thesis Director) Abstract # Abstract translated from French.\nThis thesis examines how gender inequalities are produced in press discourse, from the point of view of content and of its producers, journalists. The analysis draws on two prestigious French newsrooms that have introduced editorial policies against gender inequalities in their output: the Agency, a major news agency, and the Pureplayer, an online news outlet.\nOne of the thesis\u0026rsquo;s main contributions lies in its mixed and computational methodology. It combines a qualitative study of 30 interviews conducted in 2021 and 2022 within the two case studies with the application of natural language processing tools to a corpus of about 2.2 million articles produced by the two newsrooms. To analyse these data, I propose three indicators of inequality in articles: the masculinity rate of mentions, the masculinity rate of quotations, and the stereotyped use of reporting verbs. These indicators are computed on the corpora with machine-learning-based algorithms developed during the thesis.\nThe analysis first reveals a tension: the profession\u0026rsquo;s recognition of gender inequalities in media content has not led to a lasting change in imbalances observed for several decades. The first part of the manuscript analyses the organisational and editorial changes made by the two case studies between 2016 and 2021, then the evolution of gender inequalities in the content they produced since 1984 (for the Agency) and 2008 (for the Pureplayer). I describe how gender issues became a \u0026ldquo;professional problem\u0026rdquo; within journalism, giving rise to \u0026ldquo;feminist interventions\u0026rdquo; in newsrooms. Despite this, I show the regularity of masculinity and the gendered assignment phenomena that persist in journalistic discourse.\nThe second part of the thesis seeks to understand why these inequalities are so constant and evolve so slowly. I explain this inertia by the persistence of professional routines that (re)produce a gendered \u0026ldquo;system of exclusion\u0026rdquo; (Foucault, 1971) in media discourse. I show that adopting practices that reduce gender imbalances depends on professional routines specific to journalistic work and to the organisation of work in each newsroom. The Agency and the Pureplayer are two cases more or less favourable to changing journalistic routines, although both adhere to routines typical of the profession, such as the pursuit of objectivity. Finally, I show that the profession\u0026rsquo;s attachment to a dominant paradigm of objectivity reproduces and maintains a gendered order of media discourse. The thesis thus makes an empirical contribution to the feminist critique of objectivity by applying it to French journalism. Through its methodology, it also opens an epistemological reflection on the use of algorithms in the study of gender relations.\n","date":"13 November 2026","externalUrl":null,"permalink":"/en/seminars/2026-11-13-soutenance-ange-richard/","section":"Team seminars","summary":"","title":"The gendered order of journalism: a mixed and computational approach to gender inequalities in French print media content","type":"seminars"},{"content":"Often introduced as a generalization of proper nouns, Named Entities seem as hard to define precisely as they are useful to fields such as Information Extraction or Semantic Annotation. In the Digital Humanities context of my PhD, they shed an interesting light on the study of geographical discourses within two historical encyclopedias from the 18th and 19th centuries.\nAfter introducing key concepts from my doctoral work, this presentation will center on Named Entities and a useful extension to them. I will show how, along with other methods from machine learning and corpus linguistics, Named Entities Recognition both helped me identify important aspects of geography articles from my corpus and benefitted from them. Then, I will focus on the harder parts of the task, in particular the study of events, and how they relate to my current work within project RIPOSTE.\n","date":"12 November 2026","externalUrl":null,"permalink":"/en/seminars/2026-11-12-brenon/","section":"Team seminars","summary":"","title":"Of Named Entities and their applications","type":"seminars"},{"content":"Thesis conducted at LPNC (Body \u0026amp; Space team), GIPSA-lab (COPERNIC team) and LIG (GETALP team) laboratories, supervised by Christian Graff, Solange Rossato, Sylvain Huet and Jose-Ernesto Gomez-Balderas. Please note that the defense will be presented in French.\nThesis supervision # Christian Graff, Thesis Director, Université Grenoble Alpes Solange Rossato, Thesis Co-supervisor, Université Grenoble Alpes Sylvain Huet, Thesis Co-supervisor, Université Grenoble Alpes Jose-Ernesto Gomez-Balderas, Thesis Co-supervisor, Université Grenoble Alpes Defense jury # Christian Graff, Thesis Director, Associate Professor (HDR), Université Grenoble Alpes Marie Tahon, Reviewer, Professor, Le Mans Université Mohamed Chetouani, Reviewer, Professor, Sorbonne Université Adriana Tapus, Examiner, Professor, ENSTA Paris Dominique Vaufreydaz, Examiner, Professor, Université Grenoble Alpes Abstract # Steering a drone by voice requires understanding what a user is saying. Existing voice-control systems rely on predefined lexicons, and no resources document unconstrained speaker speech. This thesis starts from this observation and addresses it in three steps: characterizing spontaneous vocal guidance, producing a corpus aligned with flight commands, and designing an architecture capable of interpreting speech in real time to generate drone movements.\nTwo requirements structure the problem. Making piloting accessible to untrained users implies accepting unconstrained speech. Maintaining control stability imposes a fast response: beyond a 200 ms delay between intention and reaction, the operator is correcting a situation that has already passed, leading to instability. After deducting fixed acquisition and transmission delays, the speech understanding budget is approximately 100 ms. Literature addresses one or the other of these requirements, but never both.\nThree experiments were conducted using a shared dyadic protocol, where a guide vocally directs a pilot toward targets the pilot cannot see. In the first, the pilot is a human moving blindfolded, in the second, they control a drone via a joystick, in the third, an automated system replaces the pilot without the guide\u0026rsquo;s knowledge. Corpus analysis reveals four properties of spontaneous guidance: less than 2% of utterances combine multiple directions (guides break navigation down into elementary instructions), just over half carry an explicit directional marker (immediate context brings this to 61%, with the remainder covering flow control), movement amplitude, rarely quantified, is modulated by mitigators (e.g., \u0026ldquo;a little\u0026rdquo;) correlated with distance to target rather than movement duration, finally, guidance unfolds in three phases: Orientation, Approach, and Adjustment. It remains identical across both human configurations despite very different motor interfaces.\nThis resulted in the VoiceStick corpus, the first French spontaneous speech resource aligned with flight control commands: 4,219 utterances produced by 29 dyads, each paired with two labels (the guide\u0026rsquo;s intention and the executed action). This duality allows measuring the gap between what is asked and what is performed.\nArchitecturally, a cascaded approach is ruled out: 202 ms inference with 81% of errors attributable to transcription. The proposed end-to-end architecture, based on a French self-supervised encoder with weighted layers and attention aggregation, achieves 93% intention accuracy in 7 ms. A duration regression model based solely on temporal metadata leverages a novel finding: silence between utterances predicts movement amount better than any prosodic cue. Adapted for continuous streaming, the system processes 300 ms frames in 60 ms and reaches 70% accuracy in deployment, with degradation affecting direction rather than movement.\nFlight validation, where ten participants alternately direct a human pilot or the system without knowing which, reveals no difference across nine measured user experience dimensions, with half of the evaluations assigning identical scores to both pilots. This work demonstrates that a system trained on unconstrained speech can operate in real time, and that critical design constraints are not derived from intuition: they can only be observed in data that had never been collected before.\n","date":"6 November 2026","externalUrl":null,"permalink":"/en/seminars/2026-11-06-soutenance-allan-henry/","section":"Team seminars","summary":"","title":"Speaking to Fly: spontaneous speech understanding for drone guidance by naive operators","type":"seminars"},{"content":"Speech production is a complex process that relies on the coordinated interaction of several articulators, including the tongue, lips, and jaw. These interactions are partly governed by the physical, biomechanical, and dynamic properties of the structures of the vocal tract. This raises the question of how biomechanical modeling of the vocal tract, and particularly of the tongue, can contribute to a better understanding of the physical and physiological mechanisms underlying speech production.\nThis two-part presentation will first focus on the development and use of biomechanical models to investigate speech production and the emergence of speech. The second part will explore the potential integration of biomechanical models into automatic speech recognition systems, with the aim of investigating new approaches that could address some of the limitations of current recognition systems.\nTo attend remotely, contact the seminar organiser, who will create a video-conference link.\n","date":"5 November 2026","externalUrl":null,"permalink":"/en/seminars/2026-11-05-calka/","section":"Team seminars","summary":"","title":"Exploring speech production through biomechanics: From modeling to automatic speech recognition","type":"seminars"},{"content":"","date":"2 November 2026","externalUrl":null,"permalink":"/en/news/","section":"News","summary":"","title":"News","type":"news"},{"content":"Le site web de l\u0026rsquo;équipe évolue et devient un site web statique. Fini Php, fini Wordpress, fini les trous de sécurité ?\n","date":"2 November 2026","externalUrl":null,"permalink":"/en/news/2026/2026-11-02-nouveau-site/","section":"News","summary":"Le site du GETALP passe à Hugo : publications HAL, séminaires et actualités Bluesky intégrés.","title":"Nouveau site web de l'équipe","type":"news"},{"content":"","date":"2 November 2026","externalUrl":null,"permalink":"/en/tags/site/","section":"Tags","summary":"","title":"Site","type":"tags"},{"content":"","date":"2 November 2026","externalUrl":null,"permalink":"/en/tags/","section":"Tags","summary":"","title":"Tags","type":"tags"},{"content":"","date":"1 November 2026","externalUrl":null,"permalink":"/en/tags/arrival/","section":"Tags","summary":"","title":"Arrival","type":"tags"},{"content":"","date":"1 November 2026","externalUrl":null,"permalink":"/fr/tags/arriv%C3%A9e/","section":"Tags","summary":"","title":"Arrivée","type":"tags"},{"content":"We are delighted to welcome Chuyuan Li, who joined GETALP on 1 November 2026 as a permanent researcher (chargée de recherche) at CNRS.\nChuyuan has been with the team since May 2026 as a postdoctoral researcher (PARTAGES project, efficient training of specialised and multimodal French language models). She did her PhD at LORIA (Université de Lorraine) on discourse analysis in dialogue, then worked at the University of British Columbia (Canada).\nHer research covers:\nlarge language and multimodal models: interpretability, in-context learning, post-training, hallucination; AI for health: specialised models, detection of language cues in mental and neurodegenerative diseases; discourse processing: discourse structure and relations, pragmatic reasoning, long-document understanding and generation. Welcome, Chuyuan!\n","date":"1 November 2026","externalUrl":null,"permalink":"/en/news/2026/2026-11-01-bienvenue-chuyuan-li/","section":"News","summary":"Chuyuan Li joins GETALP as a CNRS researcher: language models, discourse processing and AI for health.","title":"Welcome to Chuyuan Li, CNRS researcher","type":"news"},{"content":"AI alignment is a central challenge in developing safe and reliable AI systems. In the medical domain, this raises important questions about how well AI models’ capabilities translate into clinically meaningful understanding and reliable outputs. In this talk, I will explore these questions through research on medical conceptual alignment, drug-drug interactions, and domain-specific hallucination detection. Together, these perspectives will shed light on the current state of clinical AI systems, their alignment in clinical settings, and their implications for medical AI safety.\nTo attend remotely, contact the seminar organiser, who will create a video-conference link.\n","date":"22 October 2026","externalUrl":null,"permalink":"/en/seminars/2026-10-22-sinha/","section":"Team seminars","summary":"","title":"How safe is Medical AI?","type":"seminars"},{"content":"Automatic toxicity detection in speech remains a challenge, particularly in French and for harmful utterances with implicit or context-dependent language. As part of the Pantagruel project, we first provide an overview of existing toxicity detection approaches. We then evaluate several of these tools across different French corpora, ranging from read to online spontaneous speech, comparing text and audio modalities. We then focus on LGBTphobic speech as a case study. We build a dataset of 2,084 French audio sentences with LGBT-related content, extracted from Yodas2, a large-scale corpus of YouTube videos. The sentences are being manually annotated by 9 annotators, including members of the LGBTQIA+ community. As annotation is still ongoing, we present preliminary results on a subset of the dataset, including inter-annotator agreement and annotators feedback on how audio contributes to judge the toxicity of an utterance.\nTo attend remotely, contact the seminar organiser, who will provide a video-conference link.\n","date":"15 October 2026","externalUrl":null,"permalink":"/en/seminars/2026-10-15-fabre/","section":"Team seminars","summary":"","title":"Toxicity evaluation in French audio data with a focus on LGBTphobia","type":"seminars"},{"content":"Transformer-based models have seen growing popularity in modern society, academia and industry, with ever increasing adoption. However, current models are hindered by their lack of support for Low Resourced Languages (LRLs), roughly corresponding to minority languages in a linguistic context. Their lack of text corpora means that they have been excluded from model pre-training —or its impact on it has been negligible. In my PhD, I research ways via which LRLs that are typologically related to other high-resourced languages –widely supported in LLMs– can serve as a scaffold to enhance NLU in these models. My thesis deals specifically with Aragonese, a Romance language that is extremely similar to Spanish, with almost identical syntax and close morphology, but differing lexical roots that render the language closer to Catalan and French.\nI first delve into my attempts at linguistic interpretability (the “reverse engineering” of linguistic abilities in LLMs), as an initial attempt to discover potential modifiable mechanisms of LLMs for precise linguistic intervention, followed by experiments on knowledge injection of structured linguistic knowledge as well as precise, carefully balanced continuous training of models with small amounts of text. I also present my attempts at bringing Aragonese closer to the digital world, via the creation of evaluation benchmarks for this language, as well as my participation in the CLASiK project, in which GETALP is involved, and from which my research stay at Grenoble is steered at, via the research on methods for model domain adaptation that could benefit language adaptation as well.\n","date":"24 September 2026","externalUrl":null,"permalink":"/en/seminars/2026-09-24-lopez-otal/","section":"Team seminars","summary":"","title":"Adaptation of Transformer-based models to low-resourced languages via linguistic proximity and of specific domains","type":"seminars"},{"content":"We are delighted to welcome Maxime Calka, who joined GETALP on 1 September 2026 as an associate professor (maître de conférences).\nMaxime defended his PhD in 2023 at the TIMC laboratory (Université Grenoble Alpes), supervised by Yohan Payan and Pascal Perrier: Finite element biomechanical modeling of the tongue: evaluation, speech production, and perspectives for computer-assisted tongue surgery. He then continued this work at GIPSA-lab in Grenoble: finite element tongue models (including the open-source G-OBIM model), machine-learning-based model order reduction for real-time simulation, somatosensory control of speech, and clinical applications.\nWelcome, Maxime!\n","date":"1 September 2026","externalUrl":null,"permalink":"/en/news/2026/2026-09-01-bienvenue-maxime-calka/","section":"News","summary":"Maxime Calka joins GETALP as an associate professor.","title":"Welcome to Maxime Calka, associate professor","type":"news"},{"content":"Comment annoter efficacement une langue peu dotée à l\u0026rsquo;oral quand les ressources manquent ? À partir du corpus Radio Haïti-Inter, je présenterai nos explorations sur l\u0026rsquo;apport de l\u0026rsquo;adaptation au domaine, de différentes stratégies d\u0026rsquo;entrainement et de l\u0026rsquo;apprentissage actif pour l\u0026rsquo;étiquetage morphosyntaxique du créole haïtien transcrit. Je mettrai en discussion nos résultats qui pointent vers une hiérarchie claire entre les différents facteurs expérimentés.\n","date":"22 June 2026","externalUrl":null,"permalink":"/en/seminars/2026-06-22-ziane/","section":"Team seminars","summary":"","title":"Pistes pour l'annotation morphosyntaxique du créole haïtien","type":"seminars"},{"content":"Séminaire interne RAVE, avec un invité extérieur : Quentin Macé, doctorant chez ILLUIN.\nOrchestrer plusieurs agents LLM expose le système à des vulnérabilités en cascade : une API défaillante ou une information corrompue sur un seul nœud peut faire s\u0026rsquo;effondrer l\u0026rsquo;ensemble. Pourtant, il n’y a pas encore de méthode systématique des dynamiques d’échecs de ces système et encore moins de leur optimisation. Cette présentation abordera deux questions liées à ce sujet : d’une part comment évaluer concrètement la robustesse d’un système multi-agents, au delà de la simple évaluation de la réponse finale. D’autre part comment optimiser ces systèmes pour les rendre plus robustes aux perturbations.\n","date":"13 May 2026","externalUrl":null,"permalink":"/en/seminars/2026-05-13-mace/","section":"Team seminars","summary":"","title":"Robustesse des systèmes multi-agents LLM : évaluation et optimisation","type":"seminars"},{"content":"Le titre et le résumé seront annoncés prochainement.\n","date":"12 March 2026","externalUrl":null,"permalink":"/en/seminars/2026-03-12-li/","section":"Team seminars","summary":"","title":"Titre à venir","type":"seminars"},{"content":"🚀 Deux postes de Maître·sse·s de Conférences ouverts – \u0026#34;IA et ses applications\u0026#34; (IUT-2) et \u0026#34;Modélisation, simulation, traitement et analyse du comportement humain\u0026#34; (UFR SHS) – sont associés au GETALP du Laboratoire d’Informatique de Grenoble (LIG).\n📧 contactez : didier.schwab@univ-grenoble-alpes.fr\n","date":"23 January 2026","externalUrl":null,"permalink":"/en/news/2026/bluesky-3md36ufipx22m/","section":"News","summary":"","title":"🚀 Deux postes de Maître·sse·s de Conférences ouverts – \"IA et ses applications\" (IUT-2) et …","type":"news"},{"content":"Speech and Language as Biomarkers are powerful, low-cost and completely non-invasive. They can tell us a lot about a person, allowing us to diagnose and monitor a range of conditions, from mental health to neurodegeneration. I will highlight how useful they can be in a few specific cases and the research I am developing around them, especially in the context of neurodegenerative diseases.\n","date":"15 January 2026","externalUrl":null,"permalink":"/en/seminars/2026-01-15-jouaiti/","section":"Team seminars","summary":"","title":"Speech and Language as Digital Biomarkers","type":"seminars"},{"content":"It is something of a trope to state that neural NLP has a problem of truthfulness. Hallucinations, for instance, are a problem so widespread in LLM technology that commercial providers often include disclaimers as to the truthfulness of what their systems can output. We often assume that such models, because they are derived from distributional data, have at best a loose relationship with truth. In this talk, I hope to bring some nuance to this perspective. On many levels, this is a more pervasive problem than we might think if we focus solely on hallucinations. While hallucinations are easy to come by, we can also highlight that annotators often do not agree as to what counts as a hallucination — and models routinely struggle with human label variation. Conversely, we can even highlight cases where LLMs are biased towards producing the expected true answer. In many instances, tackling this problem of truthfulness requires rethinking where the field of NLP is at: how do we train models that capture linguistic ambiguity? What are our metrics actually sensitive to? What counts as a true label?\n","date":"11 December 2025","externalUrl":null,"permalink":"/en/seminars/2025-12-11-mickus/","section":"Team seminars","summary":"","title":"Two truths and a lie: Modern NLP vs ground truth","type":"seminars"},{"content":"Congratulations to Yongxin Zhou for winning the Best Evaluation Award at INLG 2025 (2025.inlgmeeting.org) for the paper:\n\u0026#34;Can GPT models Follow Human Summarization Guidelines? A Study for Targeted Communication Goals\u0026#34; (with Fabien Ringeval and François Portet)\narxiv.org/abs/2310.16810\nINLG2025\nThe 18th International Natural Language Generation Conference is scheduled to be held in Hanoi, Vietnam from October 29 to November 2, 2025.\n","date":"5 November 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3m4um7i3svk2y/","section":"News","summary":"","title":"Congratulations to Yongxin Zhou for winning the Best Evaluation Award at INLG 2025 …","type":"news"},{"content":"À travers plusieurs exemples, je discuterai de l’importance de la diversité des ressources d’évaluation pour l’appréciation de la variation et de la performance des systèmes en TAL.\nSur une tâche comme la reconnaissance des entités nommées en français, on observe chez la plupart des systèmes un biais de performance en fonction du « genre textuel » annoté. Si celui-ci semble être à l’origine de la variation de performance, il ne l’explique pas d’un point de vue linguistique. Comment la variation se manifeste-t-elle et impacte-t-elle concrètement les systèmes ? Nous proposons pour y répondre une étude de la variation basée sur sa modélisation sous forme de traits linguistiques proposée par Douglas Biber.\nNous verrons par un second exemple, celui de la transcription automatique du breton, la problématique sous l’angle du coût de production des ressources linguistiques. Quelle stratégie adopter dans le choix des ressources à (faire) produire dans le cadre de langues peu dotées et présentant une forte variation ?\n","date":"30 October 2025","externalUrl":null,"permalink":"/en/seminars/2025-10-30-millour/","section":"Team seminars","summary":"","title":"Variation(s) : produire des ressources au service de l’évaluation en TAL","type":"seminars"},{"content":"Join 🏔️ ALPS 2026 for a week-long winter school in the French Alps at Aussois (1500m) from 29th March to 3th April! Open to PhD and master students, and academic/industrial professionals.\nAccommodation and full board are included in the fees!\nLearn more and apply by October 12th: lig-alps.imag.fr\nAdvanced Language Processing Winter School – ALPS, 29th March – 3th April 2026\n","date":"19 September 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lz7jil4gds2m/","section":"News","summary":"","title":"Join 🏔️ ALPS 2026 for a week-long winter school in the French Alps at Aussois (1500m) from …","type":"news"},{"content":"Congratulations to William N. Havard, Renault Govain, Benjamin Lecouteux and Emmanuel Schang for the best paper award at CORIA-TALN 2025\nGETALP (Laboratoire d\u0026#39;Informatique de Grenoble) is proud to have been awarded the prize for the 2nd year running. www.atala.org/content/prix...\nPrix TALN-RECITAL | ATALA\n","date":"7 July 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3ltelxhdltc23/","section":"News","summary":"","title":"Congratulations to William N. Havard, Renault Govain, Benjamin Lecouteux and Emmanuel …","type":"news"},{"content":"L’adaptation des grands modèles de langue (LLMs) à des domaines spécialisés, comme le domaine médical en français, soulève des questions méthodologiques importantes. Une évaluation comparative de trois stratégies — pré-entraînement continu (CPT), affinage supervisé (SFT) et combinaison des deux — permet d’en analyser la pertinence en contexte francophone.\n","date":"18 June 2025","externalUrl":null,"permalink":"/en/seminars/2025-06-18-belmadani/","section":"Team seminars","summary":"","title":"Adaptation des connaissances médicales aux grands modèles de langue pour le question-answering : stratégies et analyse comparative","type":"seminars"},{"content":"Correct causal reasoning requires domain knowledge beyond observed data. In this talk, we will delve into the causal capabilities of large language models (LLMs), discussing recent studies and benchmarks of their ability to retrieve and apply causal knowledge, as well as the limitations of their causal reasoning capabilities.\n","date":"5 June 2025","externalUrl":null,"permalink":"/en/seminars/2025-06-05-kiciman/","section":"Team seminars","summary":"","title":"A New Frontier at the Intersection of Causality and LLMs","type":"seminars"},{"content":"Emmanuel Schang, professeur au Laboratoire Ligérien de Linguistique (Orléans), et William Havard, post-doctorant dans le même laboratoire et ancien doctorant de l\u0026rsquo;équipe GETALP, présentent leurs travaux autour de la documentation des langues créoles.\nDans le cadre du projet ANR CREAM, nous explorons les enjeux liés à la documentation des langues créoles, souvent marginalisées dans les ressources numériques. Nous présentons des expérimentations sur la reconnaissance automatique de la parole (ASR) pour le créole haïtien (kreyòl ayisyen), ainsi que des mesures de distance linguistique entre le morisien, le français et l’anglais, afin de mieux comprendre les dynamiques de proximité et de divergence entre ces langues.\n","date":"5 June 2025","externalUrl":null,"permalink":"/en/seminars/2025-06-05-schang/","section":"Team seminars","summary":"","title":"Documentation des langues créoles","type":"seminars"},{"content":"Advancements in large language models (LLMs) make it easier to develop interactive information systems. However, LLMs have a problem of generating hallucinations, especially for closed information of organizations and frequently renewed information. We prototyped Doshisha University\u0026rsquo;s interactive information system (text-chat system) in which we focused on the correctness of the information supplied by retrieval-augmented generation using knowledge graph (KG-RAG) and bilingualism. We built a knowledge graph (KG) on the Neo4j database and developed a chat system that used the KG and OpenAI API for GPTs. The KG enables three types of search, i.e., KG search, full-text search, and vector search, by registering text information and its embedding vector into the nodes. The chat system calls the API for GPTs four times: 1) preprocessing, i.e., translating questions in English to those in Japanese and converting colloquial Japanese expressions to formal Japanese texts; 2) conversion from natural language to Cypher queries for KG search; 3) response generation based on KG-RAG; and 4) additional suggestion of next question for continuing the dialogue. We further developed a smart responding mechanism to inaccurate user inputs that contain wrong proper nouns, etc. We quantitatively evaluated the system from two points of view: 1) general accuracy of information provided by the responses in two languages and 2) accuracies of responses to inaccurate inputs. The addition of the first LLM for preprocessing and the smart responding mechanism significantly improved the general accuracy and the accuracy of responses to inaccurate inputs, respectively.\n","date":"9 April 2025","externalUrl":null,"permalink":"/en/seminars/2025-04-09-kato/","section":"Team seminars","summary":"","title":"Retrieval-Augmented Generation based on Knowledge Graph (KG-RAG) for Providing Accurate Information to Dialogue Systems","type":"seminars"},{"content":"Cet article regarde si les consommateurs et consommatrices de musique se détournent des artistes à la suite d\u0026rsquo;accusations publiques de violences sexistes et sexuelles (VSS). A l\u0026rsquo;aide des données du marché de la musique enregistrée en France de 2006 à 2022, ainsi que d\u0026rsquo;une liste d\u0026rsquo;artistes accusés publiquement de VSS construite à l\u0026rsquo;aide d\u0026rsquo;expressions régulières, d\u0026rsquo;une détection d\u0026rsquo;entités nommées et validée manuellement, l\u0026rsquo;analyse se concentrera sur 4 études de cas. Les effets sur les ventes d\u0026rsquo;albums et les écoutes en ligne (streaming) seront discutés. La présentation s\u0026rsquo;attachera surtout à discuter d\u0026rsquo;une autre source de données dont l\u0026rsquo;analyse pourrait compléter le travail: les commentaires Youtube.\n","date":"21 February 2025","externalUrl":null,"permalink":"/en/seminars/2025-02-21-chatelain/","section":"Team seminars","summary":"","title":"Il n'y a pas de mauvaise publicité ? De l'effet des scandales sexuels sur le succès commercial des artistes","type":"seminars"},{"content":"Angela Potochnik, a philosopher of science, joins us remotely for a presentation that will make us take a step back and reflect on the foundations of \u0026ldquo;Science\u0026rdquo;. While we practice science on a daily basis, we rarely pause to consider the principles that make it work. In particular, fields like machine learning and explainable AI often aim, explicitly or implicitly, to automate the discovery of scientific explanations. However, many of this approaches are built on our superficial understanding of what truly makes science work. This seminar will be an opportunity for all of us to reflect on these questions.\n","date":"6 February 2025","externalUrl":null,"permalink":"/en/seminars/2025-02-06-potochnik/","section":"Team seminars","summary":"","title":"Reflecting on the foundations of science","type":"seminars"},{"content":"Congratulations to Mariam Nakhlé and all the other co-authors! Their paper, \u0026#34;Document-Level Financial Test-Set for Machine Translation (DOLFIN)\u0026#34; (a collaboration with Lingua Custodia), has been accepted at NAACL 2025!\n","date":"28 January 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lgsru5rwuc2m/","section":"News","summary":"","title":"Congratulations to Mariam Nakhlé and all the other co-authors! Their paper, …","type":"news"},{"content":"Le GETALP est présent aux journées R\u0026amp;T Cognition de l\u0026#39;Institut Carnot Cognition ! Nos chercheur(e)s ont l\u0026#39;opportunité d\u0026#39;échanger avec plusieurs centaines de participants autour de deux démonstrateurs illustrant le savoir-faire de notre équipe.\nwww.institut-cognition.com/journees-rt-...\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"28 January 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lgsrfa52nk2m/","section":"News","summary":"","title":"Le GETALP est présent aux journées R\u0026T Cognition de l'Institut Carnot Cognition ! Nos …","type":"news"},{"content":"Congratulations to Sannara EK for successfully defending his thesis “Personalized Federated Learning for Sensor-based Human Activity Recognition in Pervasive Heterogeneous Environments” on November 26, 2024. theses.fr/s296447\nApprentissage fédéré personnalisé pour la reconnaissance de l\u0026#39;activité humaine à partir de capteurs dans des environnements pervasifs hétérogènes | Theses.fr\nLes avancées récentes dans la technologie des capteurs et l\u0026#39;informatique mobile ont considérablement amélioré les applications de l\u0026#39;informatique omniprésente, intégrant des dispositifs intelligents da...\n","date":"22 January 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lgcvd46ff22u/","section":"News","summary":"","title":"Congratulations to Sannara EK for successfully defending his thesis “Personalized …","type":"news"},{"content":"Congratulations to Yongxin ZHOU for successfully defending her thesis “Affect-aware Natural Language Generation: application to dialogue and cognitive remediation session summarization in low-resource settings” on November 6, 2024. theses.fr/s246878\nApproches sensibles à l\u0026#39;affect pour le résumé automatique de textes : application au dialogue et à la remédiation cognitive dans un contexte à faible ressources | Theses.fr\nCette thèse explore la génération automatique de texte (GAT) pour la thérapie numérique, en se concentrant sur la génération de rapports qui résument les sessions d\u0026#39;entraînement cognitif menées par le...\n","date":"22 January 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lgcv5vlx622u/","section":"News","summary":"","title":"Congratulations to Yongxin ZHOU for successfully defending her thesis “Affect-aware …","type":"news"},{"content":"Félicitations à Yongxin ZHOU pour avoir soutenu avec succès sa thèse \u0026#34;Approches sensibles à l\u0026#39;affect pour le résumé automatique de textes : application au dialogue et à la remédiation cognitive dans un contexte à faible ressources\u0026#34; le 6 novembre 2024.\ntheses.fr/s246878\nApproches sensibles à l\u0026#39;affect pour le résumé automatique de textes : application au dialogue et à la remédiation cognitive dans un contexte à faible ressources | Theses.fr\nCette thèse explore la génération automatique de texte (GAT) pour la thérapie numérique, en se concentrant sur la génération de rapports qui résument les sessions d\u0026#39;entraînement cognitif menées par le...\n","date":"22 January 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lgcv5cw5bs2u/","section":"News","summary":"","title":"Félicitations à Yongxin ZHOU pour avoir soutenu avec succès sa thèse \"Approches sensibles …","type":"news"},{"content":"Félicitations à Solène Evain pour avoir soutenu avec succès sa thèse \u0026#34;Dimensions de variation de la parole spontanée pour l’étude inter-corpus des performances de systèmes de reconnaissance automatique de la parole\u0026#34; le 8 octobre 2024 theses.fr/s252448\nDimensions de variation de la parole spontanée pour l\u0026#39;étude inter-corpus des performances de systèmes de reconnaissance automatique de la parole | Theses.fr\nCes dernières années, la reconnaissance automatique de la parole a beaucoup progressé grâce au développement du deep learning et des modèles pré-appris. Toutefois, les performances sur la parole spont...\n","date":"22 January 2025","externalUrl":null,"permalink":"/en/news/2025/bluesky-3lgcux5lb4c2u/","section":"News","summary":"","title":"Félicitations à Solène Evain pour avoir soutenu avec succès sa thèse \"Dimensions de …","type":"news"},{"content":"Instruct-tuned language models have changed the way we interact with computers through language. A lot of work is now devoted to outlining the envelope of their capabilities and their limitations. In this talk, I will describe my tentative to tackle a subset of this problem in the context of multimodal learning. The first line of work I will present is devoted to understanding how vision-language models (VLM) represent fine-grained relationships between objects. Previously, we had shown that VLMs trained on image-text matching fail to extract reliable relationships between objects. I will report on an updated set of experiments to see how those capabilities have evolved in recent instruction-tuned multimodal models. The second line of work aims at better understanding the space of tasks that one may ask from a LLM, in order to select better training mixes and adequately evaluate them. I will evoke past work where we tried to decompose the task space into unitary skills from an introspective and benchmark availability perspective, and more recent work which tries to address this problem from a statistical point of view, allowing to build empirically a hierarchy of task solving capabilities. Finally, I will describe my ongoing project on building foundation models of multimodal conversation. Drawing from video generation techniques, this project aims at assessing the representation capabilities of models able to continue a human-human conversation, both in the audio and video modality. I will describe a framework for studying those capabilities through probing underlying representations according to various tasks. This project has applications in studying human conversation behavior, as well as simulating this behavior in the context of robotic interactions.\n","date":"16 January 2025","externalUrl":null,"permalink":"/en/seminars/2025-01-16-favre/","section":"Team seminars","summary":"","title":"What do multimodal LLMs learn?","type":"seminars"},{"content":"Individuals with severe to profound disabilities (SPD) represent a particularly vulnerable and often marginalized minority. Long considered non-communicative, recent advancements in legislation concerning the rights of people with disabilities—especially the right to communication and inclusion—compel us to reconsider their communicative potential. However, due to their historical exclusion from communication research and (re)educational programs, we lack methods and models to describe, understand, and assess their communication.\nOver the past two decades, clinical research has begun addressing this gap, but it often progresses independently of fundamental research on communication, rarely integrating its theoretical and methodological advancements. Moreover, the growing development and use of augmentative and alternative communication (AAC) tools, such as communication binders or software with voice synthesis using pictogram boards, suggest that individuals with SPD might achieve symbolic communication far more sophisticated than previously thought—or than is still commonly believed. To better understand these capacities, it is essential to engage in meticulous observation and reconsider existing concepts and methods, or even invent new ones.\nIn this context, the ParticipAACtion research project proposes an approach at the intersection of fundamental research, clinical research, and the expertise of professional or family caregivers. The objective is to integrate these perspectives to establish inclusive standards for describing behaviors and to create a \u0026ldquo;dictionary\u0026rdquo; of communicative behaviors of individuals with SPD, including their use of AAC. Our methodology is based on detailed observation and manual annotations of individuals\u0026rsquo; behaviors in interactive situations. However, this approach faces significant time constraints, highlighting the need to develop automated methods, potentially leveraging artificial intelligence (AI), to accelerate and enhance these analyses.\n","date":"12 December 2024","externalUrl":null,"permalink":"/en/seminars/2024-12-12-rochet-capellan/","section":"Team seminars","summary":"","title":"How to Observe and Describe Communication in Individuals with Severe to Profound Disabilities?","type":"seminars"},{"content":"The cyber threat landscape is constantly evolving, with new tools and technological advancements changing how adversaries conduct cyberattacks. While Artificial Intelligence (AI) has the potential to offer advantages to our society, adversaries are not asleep at the wheel. More and more evidence appears that AI is also used for offensive purposes. In this talk, we explore the innovation potential of hackers, focusing on the specific case of (i) how hackers can leverage AI and (ii) how we (as researchers and defenders) can gain insights into the operations of cybercriminals. For the latter, we discuss the purpose and context of cyber threat intelligence extraction from various data sources, such as social network data, underground forums (from the clearnet and darknet), cyber-crime-related chat channels, and darknet websites. In this regard, I will deep dive into the role of NLP, and specifically NLU, for cyber threat intelligence extraction, as well as discuss the challenges and open problems in the context of cybersecurity.\n","date":"10 October 2024","externalUrl":null,"permalink":"/en/seminars/2024-10-10-schroer/","section":"Team seminars","summary":"","title":"Hackers, AI and NLP for cyber threat intelligence","type":"seminars"},{"content":"Language change over time has been researched for decades by historical linguists and, more recently, by computational linguists. Computational modeling of language change enables cross-language comparison of multifaceted changes at a large scale. In this talk, I will start by presenting a method to capture syntactic changes and investigate how similar these changes are in German and English over the past hundred years. Additionally, I will present a method to capture semantic changes, particularly to detect the gained or lost meanings with low frequency over time. I will also showcase the use of our method as a visualization tool to compare cross-language semantic changes in German and English. Lastly, I will present a pathway from computational modeling of language change to computational lexicography by comparing detected word meanings with dictionary sense inventories to identify unrecorded meanings in dictionaries.\n","date":"19 June 2024","externalUrl":null,"permalink":"/en/seminars/2024-06-19-zhao/","section":"Team seminars","summary":"","title":"A Pathway from Language Change to Computational Lexicography","type":"seminars"},{"content":"NLP systems promise to disrupt society through applications in high-stakes social domains. However, current evaluation and development focus on tasks that are not grounded in specific societal implications, which can lead to societal harm. There is a need to evaluate and mitigate the societal harms and, in doing so, bridge the gap between the realities of application and how models are currently developed.\nIn this talk, I will present recent work addressing these issues in the domain of online content moderation. In the first part, I will discuss online content moderation to enable constructive conversations about race. Content moderation practices on social media risk silencing the voices of historically marginalized groups. We find that both the most recent models and humans disproportionately flag posts in which users share personal experiences of racism. Not only does this censorship hinder the potential of social media to give voice to marginalized communities, but we also find that witnessing such censorship exacerbates feelings of isolation. We offer a path to reduce censorship through a psychologically informed reframing of moderation guidelines. These findings reveal how automated content moderation practices can help or hinder this effort in an increasingly diverse nation where online interactions are commonplace.\nIn the second part, I will discuss how identified biases in models can be traced to the use-mention distinction, which is the difference between the use of words to convey a speaker\u0026rsquo;s intent and mention of words for quoting what someone said or pointing out properties of a word. Computationally modeling the use-mention distinction is crucial for enabling counterspeech to hate and misinformation. Counterspeech that refutes problematic content mentions harmful language but is not harmful itself. We show that even recent language models fail at distinguishing use from mention and that this failure propagates to downstream tasks. We introduce prompting mitigations that teach the use-mention distinction and show that they reduce these errors.\nFinally, I will discuss the big picture and other recent efforts to address these issues in different domains beyond content moderation, including education, emotional support, and public discourse about AI. I will reflect on how, by doing so, we can minimize the harms and develop and apply NLP systems for social good.\n","date":"2 May 2024","externalUrl":null,"permalink":"/en/seminars/2024-05-02-gligoric/","section":"Team seminars","summary":"","title":"Large language models for enabling constructive online conversations","type":"seminars"},{"content":"In the past 10 years, ethics has become a major topic in NLP, both in terms of community management, with ethics committees for most international conferences, and in research work, with specialized tracks in the most renowned conferences and a very active field of research. I\u0026rsquo;ll present here my latest work on ethics in NLP, focusing on the evaluation of stereotyped biases in LLMs. I\u0026rsquo;ll show why we need to improve our models and how ethics can help us do so. I\u0026rsquo;ll present new resources and tools that we developed to evaluate LLMs, especially adapted for languages other than English. I encourage all the attendees to come with their concerns and questions, so that this time is devoted to open discussion.\n","date":"7 March 2024","externalUrl":null,"permalink":"/en/seminars/2024-03-07-fort/","section":"Team seminars","summary":"","title":"How Ethics can help us build better systems: Towards a comprehensive evaluation of stereotyped biases in LLMs","type":"seminars"},{"content":"Les graphes de connaissances scientifiques (GCS), c’est-à-dire des graphes de connaissances représentant des articles scientifiques, leurs métadonnées, leur contenu, ainsi des connaissances structurées extraites de ce contenu, connaissent un incroyable essor. En France, les initiatives à la fois techniques (graphe de connaissances HAL) et scientifiques (nombreux projets en informatique médicale, Covid-on-the-web, ISSA) se multiplient, suivant de près une tendance déjà très ancrée au niveau Européen. Les GCS peuvent tout autant servir les institutions pour évaluer et analyser la production de la recherche, les décideurs à orienter les politiques publiques de recherche et les scientifiques à découvrir des tendances émergentes et à mener des revues systématiques d\u0026rsquo;envergure. Par ailleurs, devant l\u0026rsquo;émergence de modèles d\u0026rsquo;IA générative, et de techniques de plus en plus sophistiquées de falsification, s\u0026rsquo;ajoutant à une quantité déjà très importante de publications frauduleuses, l\u0026rsquo;utilisation des GSC conjointement aux techniques de vérification de faits computationnelles, est un outil clé dans la lutte contre ce fléau.\nCette présentation présentera d\u0026rsquo;abord les bases des graphes de connaissances et du web sémantique, suivies d\u0026rsquo;une présentation des spécificités des GSC, des techniques permettant de les construire et d\u0026rsquo;une revue d\u0026rsquo;initiatives et de cas d\u0026rsquo;usage existants. Ensuite nous nous intéresseront spécifiquement à l\u0026rsquo;utilisation potentielle des GC (et particulièrement les GCS) et de modèles de vérification computationnelle pour l\u0026rsquo;identification d\u0026rsquo;informations erronées ou contrefactuelles.\n","date":"15 February 2024","externalUrl":null,"permalink":"/en/seminars/2024-02-15-tchechmedjiev/","section":"Team seminars","summary":"","title":"Les graphes de connaissances scientifiques dans tous leurs états : de la construction à la vérification de factualité","type":"seminars"},{"content":"As a new GetAlp member, I\u0026rsquo;ll use this presentation to provide a shallow but broad overview of my recent research interests: (i) causal machine learning, a sub-field of ML aiming to understand and rectify issues arising from traditional statistical approaches; (ii) LLM interpretability, which I will argue should be rooted in causal reasoning, the rigorous framework to answer \u0026ldquo;why\u0026rdquo; questions; (iii) introducing Thought-Decoding, an emerging paradigm viewing LLMs as producers of semantically meaningful messages (thoughts), where the goal is to use LLMs in interactions with other AI systems, tools, or humans in control flows to iteratively refine and improve the thoughts until a robust answer is found.\n","date":"7 December 2023","externalUrl":null,"permalink":"/en/seminars/2023-12-07-peyrard/","section":"Team seminars","summary":"","title":"Towards Understanding, Interpreting, and Using Language Models","type":"seminars"},{"content":"Notre équipe représentée à #TC45! Our team represented at #TC45.\nx.com/E_Esperanca/st…\n","date":"21 November 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3mv7rj22k/","section":"News","summary":"","title":"Notre équipe représentée à #TC45! Our team represented at #TC45. x.com/E_Esperanca/st…","type":"news"},{"content":"Whisper (Radford, 2023) is a multilingual audio Large Language Model (LLM) developed by OpenAI and trained with 680,000 hours of labeled audio data to do transcription and translation into English (as to the free models released on HuggingFace) for 96 languages. This talk will present results and current research based on a customised version of Whisper’s C++ implementation (Gerganov, 2022) developed at Université Paris Cité (Younès 2023). We were able to probe the LLM outputs for French, English and Persian and investigate Whisper’s dictionary of subtokens.\nFirst, reporting on a paper presented at SummitMT, I will present the results of a comparison of the six models designed to translate from French and Persian into English (Ballier et al, 2023).\nSecond I will show that the transcriptions in English produced by the multilingual medium model are more robust to non-native speech variability than the tiny model transcriptions. In turn, the tiny model transcriptions provide plausible correlates of non-native pronunciation errors (Ballier et al., submitted). The audio LLM might be used as a plausible simulation of misunderstanding scenarios with native speakers. Using the probability of the LLM prediction, we can try to score non-native productions (Ballier et al., in prep).\nThen, presenting ongoing experiments with the translation and transcription of Persian data, I will showcase some architectural biases of the multilingual model: hallucinations can be observed in the transcriptions of Persian that can be explained by the absence of some graphemes of Persian in the vocabulary of subtokens. I will also present some of the Americancentered biases of the Whisper models for English and discuss some of the consequences for educational uses of these models).\nMore generally, I will discuss the reverse engineering strategy that we have adopted. With great help from data scientists and phoneticians, we have created scripts to map back the LLM predictions onto the speech signal. We aim to contribute to XAI (Explainable Artificial Intelligence) by trying to match the out of vocabulary subtokens produced by Whisper to discrete prosodic events. We also try to model how some of the LLM transcriptions may or may not correlate with human perception tests.\n","date":"16 November 2023","externalUrl":null,"permalink":"/en/seminars/2023-11-16-ballier/","section":"Team seminars","summary":"","title":"Using Whisper for translation and transcription: an XAI roadmap for audio multilingual LLMs","type":"seminars"},{"content":"La saison des soutenances continue avec la thèse de @SinaAlisamir intitulée \u0026#34;Machine Learning of Emotional Expressions In the Wild from Acoustic Signals and Text\u0026#34; soutenue avec brio ! #MachineLearning #SpeechSynthesis\n","date":"10 November 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3msoitl2h/","section":"News","summary":"","title":"La saison des soutenances continue avec la thèse de @SinaAlisamir intitulée \"Machine …","type":"news"},{"content":"C’est la saison des soutenances ! Ouverture avec notre collègue Brooke Stephenson qui a brillamment soutenu sa thèse de doctorat intitulée « Beyond words: leveraging language models for incremental and context-aware text-to-speech synthesis » #SpeechSynthesis\n","date":"26 October 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3mq4jts2f/","section":"News","summary":"","title":"C’est la saison des soutenances ! Ouverture avec notre collègue Brooke Stephenson qui a …","type":"news"},{"content":"Notre équipe mise en avant par le CNRS grace à @didier_schwab porteur du projet Pantagruel 👏\nx.com/CNRSinformatic…\n","date":"24 October 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3mnm2uu2y/","section":"News","summary":"","title":"Notre équipe mise en avant par le CNRS grace à @didier_schwab porteur du projet Pantagruel …","type":"news"},{"content":"We are doing🏔️ALPS again, 4⃣th edition!\nA week full of insightful lectures on Advanced Natural Language Processing.\nThis time: we will meet for real in the Alps at 1500m, surrounded by mountains.\nApplication deadline: Oct 1st\nlig-alps.imag.fr/index.php/appl…\nApplication\nThe intended audience of ALPS 2025 are graduate students (PhD students or advanced master students) in natural language processing or related fields. We also welcome other NLP practitioners working in industry or academia.\n","date":"11 July 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3ml2nsj2u/","section":"News","summary":"","title":"We are doing🏔️ALPS again, 4⃣th edition! A week full of insightful lectures on Advanced …","type":"news"},{"content":"We are recruiting a PhD student to work on the prediction of user satisfaction from massive conversational telephonic speech. More details here:\nsites.google.com/site/fabienrin…\nFabien Ringeval - Jobs\nOpened positions will be listed here when available\n","date":"23 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3mffghb2n/","section":"News","summary":"","title":"We are recruiting a PhD student to work on the prediction of user satisfaction from …","type":"news"},{"content":"Fin de @coriataln2023, avec pour le dernier jour un exemple de projet transdisciplinaire international : PROPICTO, Développer des systèmes de traduction de la parole vers des séquences de pictogrammes pour améliorer l’accessibilité de la communication\nlnkd.in/eeaUtw-6\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"9 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3mb73cz2z/","section":"News","summary":"","title":"Fin de @coriataln2023, avec pour le dernier jour un exemple de projet transdisciplinaire …","type":"news"},{"content":"2ème journée @coriataln2023 Un poster sur un exemple de projet jeune chercheur financé par l’Agence Nationale de la Recherche présenté depuis hier par Fabien Lopez, État de l’art sur la coréférence\nlnkd.in/eTK9dmGA\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"7 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3m574ql2h/","section":"News","summary":"","title":"2ème journée @coriataln2023 Un poster sur un exemple de projet jeune chercheur financé par …","type":"news"},{"content":"2ème journée @coriataln2023 avec deux démonstrations de @cecilemacaire : - Plateformes pour la création de données en pictogrammes- Voice2Picto : un système de traduction automatique de la parole vers des pictogrammes\nlnkd.in/eVwDdj6b lnkd.in/e3iKKpDp\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"7 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3lz3npl2h/","section":"News","summary":"","title":"2ème journée @coriataln2023 avec deux démonstrations de @cecilemacaire : - Plateformes …","type":"news"},{"content":"2ème journée @coriataln2023 avec une présentation de poster de Jeongwoo Kang, doctorante cifre chez Emvista, Analyse sémantique AMR pour le français par transfert translingue\nlnkd.in/e4ctAaFW\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"7 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3ltz7sb2j/","section":"News","summary":"","title":"2ème journée @coriataln2023 avec une présentation de poster de Jeongwoo Kang, doctorante …","type":"news"},{"content":"2ème journée @coriataln2023 avec une présentation orale de Mariam Nakhlé, doctorante cifre chez Lingua Custodia, l’évaluation de la traduction automatique du caractère au document : un état de l’art\nlnkd.in/eGdR3pmi\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"7 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3lqa3kc2f/","section":"News","summary":"","title":"2ème journée @coriataln2023 avec une présentation orale de Mariam Nakhlé, doctorante cifre …","type":"news"},{"content":"Aidan Manion, doctorant cifre chez Epos/SOS médecins, présente maintenant @coriataln2023 son travail exploratoire sur l’Augmentation des modèles de langage français par graphes de connaissances pour la reconnaissance des entités biomédicales\nlnkd.in/eBJh8jX7\nLinkedIn\nThis link will take you to a page that’s not on LinkedIn\n","date":"6 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3lm42sc25/","section":"News","summary":"","title":"Aidan Manion, doctorant cifre chez Epos/SOS médecins, présente maintenant @coriataln2023 …","type":"news"},{"content":"L’équipe GETALP (Groupe d’Étude en Traduction Automatique et en Traitement Automatique des langues et de la Parole) du Laboratoire d’Informatique de Grenoble est encore une fois présente pour la conférence francophone de référence du domaine, @coriataln2023. 1/7\n","date":"6 June 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3kosmbi2z/","section":"News","summary":"","title":"L’équipe GETALP (Groupe d’Étude en Traduction Automatique et en Traitement Automatique des …","type":"news"},{"content":"GETALP co-organise avec enthousiasme les JLC2023 , du 3 au 7 juillet 2023, à la MACI sur le campus de St Martin D’Hères, Grenoble. Les inscriptions sont ouvertestarif réduit jusqu’au 1er juin !\nbit.ly/3pwWjr9 bit.ly/3MrQ7tD\nJournées Linguistique de Corpus 2023 - Sciencesconf.org\nPour les membres des laboratoires organisateurs, merci de contacter un membre du comité d\u0026#39;organisation de votre laboratoire pour connaitre les tarifs accordés.\n","date":"11 May 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3km7wb422/","section":"News","summary":"","title":"GETALP co-organise avec enthousiasme les JLC2023 , du 3 au 7 juillet 2023, à la MACI sur …","type":"news"},{"content":"GETALP sera présent à CORIA-TALN 2023 via l’atelier hOUPSh organisé par @E_Esperanca \u0026amp; Vincent Segonne, en collaboration avec José G. Moreno (IRIT). Venez participer le lundi 5 juin !\nhoupsh2023.sciencesconf.org coria-taln-2023.sciencesconf.org\nCORIA TALN RJCRI RECITAL 2023 - Sciencesconf.org\nOrganisée conjointement par les laboratoires franciliens sous l’égide de l\u0026#39;Association francophone de Recherche d’Information et Applications (ARIA) et l’Association pour le Traitement Automatique des Langues (ATALA), l’édition 2023 regroupe :\n","date":"27 April 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3juwp7m2y/","section":"News","summary":"","title":"GETALP sera présent à CORIA-TALN 2023 via l’atelier hOUPSh organisé par @E_Esperanca \u0026 …","type":"news"},{"content":"[Appel à volontaires]\nRecherchons des volontaires pour enregistrer 50 phrases en français, provenant d\u0026#39;histoires pour enfant, de poèmes, chansons, et du domaine médical.\nTout se passe en ligne et prend 10 minutes maximum.\nÉcrivez à propicto-corpus@univ-grenoble-alpes.fr\n","date":"21 April 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3jphfhz2n/","section":"News","summary":"","title":"[Appel à volontaires] Recherchons des volontaires pour enregistrer 50 phrases en français, …","type":"news"},{"content":"Discussion avec Solange Rossato autour de l’axe 4 Cogniton et Langage puis Conclusion par @SedogboC de cette Convention Scientifique de l’@inst_cognition #IA #cognition\n","date":"30 March 2023","externalUrl":null,"permalink":"/en/news/2023/bluesky-3lfx3jke76m2y/","section":"News","summary":"","title":"Discussion avec Solange Rossato autour de l’axe 4 Cogniton et Langage puis Conclusion par …","type":"news"},{"content":"From the very beginning of natural language processing, one of the most important issues has been ambiguity, when words, phrases or sentences have multiple meanings. For Machine Translation (MT), ambiguity is an issue when those meanings result in different translations. Is French trombone a paperclip or a musical instrument? If the fans are out of order, is there a problem at the football (supporteurs) or a ventilation problem (ventilateurs). Traditionally, MT was carried out sentence by sentence, and such sentences could be lacking the context necessary to distinguish between correct and incorrect translations. Since then, contextual MT has been growing in popularity, with the inclusion of both linguistic and extra-linguistic context. New evaluation strategies have also been developed to target context-dependent phenomena, since the usual metrics are ill-adapted. In this talk, I will start by introducing contextual MT using linguistic context, including some of my earlier work carried out in evaluation and more recent work for the evaluation of the multilingual language model BLOOM. I will then describe recent work in the inclusion of visual context for disambiguation in MT, including a new test set and an adapted multimodal MT approach.\n","date":"28 March 2023","externalUrl":null,"permalink":"/en/seminars/2023-03-28-bawden/","section":"Team seminars","summary":"","title":"From Linguistic to Visual Context in Machine Translation","type":"seminars"},{"content":"L\u0026rsquo;évaluation des compétences de lecture est une question abordée dans le domaine de la lisibilité. Les formules de lisibilité se sont développées au cours du 20ème siècle, passant d\u0026rsquo;un calcul manuel à une automatisation par des techniques de traitement automatique du langage naturel (TAL) et d’apprentissage automatique. Ce changement de paradigme a amélioré la précision des modèles de lisibilité, facilitant les applications dans d\u0026rsquo;autres domaines tels que l\u0026rsquo;évaluation de la production écrite et la simplification automatique des textes. Plusieurs outils sont disponibles pour l\u0026rsquo;anglais afin d\u0026rsquo;aider les chercheurs à calculer automatiquement les caractéristiques textuelles. Cependant, les outils similaires pour la langue française sont rares. Cet atelier présentera FABRA, un outil pour l\u0026rsquo;annotation de textes, qui fournit une solution pour ce défi en obtenant et en calculant automatiquement plus de 5000 caractéristiques textuelles pertinentes liées à la lisibilité. Ensuite, un outil (Dmesure) qui allie la connaissance linguistique et le deep learning pour soutenir et évaluer la lisibilité d\u0026rsquo;un texte écrit sera présenté et démontré. L\u0026rsquo;exposé explorera également le défi que représente la combinaison de la connaissance linguistique et de l\u0026rsquo;apprentissage profond, en présentant de nouvelles façons d\u0026rsquo;améliorer la précision des outils de lisibilité.\n","date":"23 January 2023","externalUrl":null,"permalink":"/en/seminars/2023-01-23-wilkens/","section":"Team seminars","summary":"","title":"Améliorer la précision des outils de lisibilité : présentation de FABRA et Dmesure pour l'annotation et l'évaluation de la lisibilité des textes en français","type":"seminars"},{"content":"Martin Rodrigue A. Ongolo : Plongement multilingue basé sur le mapping monolingue pour la traduction automatique neuronale des langues peu dotées\nLow-resource languages pose a performance problem for neural machine translation because they have scarce data resources. In this work, we propose to use multilingual embedding based on monolingual mapping as a method of representing words as input to a neural automatic translator. This approach is then compared to word representation approaches in the monolingual context used in the literature for neural machine translation. The results we obtain with multilingual embedding on a dataset of 7942 French-Ewondo parallel sentences are promising because they are better than those obtained with the representation approaches, in particular, the Onehot and Word embedding methods used so far. Also, we highlight the phenomenon of agglutination which is a characteristic of several African languages and more particularly Cameroonian and we try to bring as a solution to this phenomenon a grouping of words before the tokenization stage.\nYannick Yomie Nzeuhang : Acoustic modeling by self-supervised learning for low ressources languages\nAcoustic modeling is the task of determining a mapping function between an audio signal and the corresponding linguistic unit transcription. Initially a distinct part of a global speech recognition architecture (Hidden Markov Model), the acoustic model has been merged into the recent neural network-based architectures. We present the main lines of development of this research intending to find an exploitable path for low resources languages.\n","date":"12 January 2023","externalUrl":null,"permalink":"/en/seminars/2023-01-12-ongolo/","section":"Team seminars","summary":"","title":"Plongement multilingue pour la traduction neuronale des langues peu dotées / Acoustic modeling by self-supervised learning for low resource languages","type":"seminars"},{"content":"Nicolas Ballier : Travaux en cours sur la traduction neuronale\nDans un premier temps, je présenterai quelques résultats du projet Neuroviz sur l\u0026rsquo;analyse des flux d\u0026rsquo;informations et ses effets pour les biais de genre dans la traduction français/anglais, en particulier liés à la phase de sous-tokenisation.\nPuis, je présenterai les premiers résultats de travaux en cours sur la traduction en anglais des dislocations du français. Je montrerai que les systèmes continuent à dupliquer les sujets dans les traductions en anglais dès que l\u0026rsquo;on s\u0026rsquo;écarte de la forme canonique de la dislocation à gauche très présente dans les corpus (type \u0026ldquo;X, c\u0026rsquo;est \u0026hellip;\u0026rdquo;), en particulier dans le cas des dislocations multiples. J\u0026rsquo;analyserai des traductions d\u0026rsquo;exemples authentiques tirés du Corpus de Français Parlé Parisien (CFPP) et montrerai les limites des systèmes testés (Deepl, Google translate, MBART-50). Je détaillerai une analyse en cours qui vise à l\u0026rsquo;opérationnalisation de la distance à l\u0026rsquo;élément disloqué. Si la variable à prédire est le succès de la traduction (au moins en termes d\u0026rsquo;adéquation syntaxique, grammaticalité des traductions produites), je questionnerai les variables susceptibles d\u0026rsquo;en rendre compte, qu\u0026rsquo;il s\u0026rsquo;agisse de variables qualitatives plus classiques en linguistique de corpus ou du paramètre de la distance entre l\u0026rsquo;élément disloqué et le prédicat de la principale. Je questionnerai la représentation de la distance en structure de surface pour le linguiste (nombre de tokens, nombre de caractères) mais également dans l\u0026rsquo;input réel de la machine (nombre de sous-tokens après le prétraitement en byte pair encoding).\nEnfin, j\u0026rsquo;exposerai les idées sous-jacentes au projet MAKE-NMTViz en explicitant les attentes d\u0026rsquo;une telle démarche : visualiser pour comprendre. Je présenterai des systèmes de visualisation existants ou ayant existé et esquisserai les souhaits de visualisation pour les différentes étapes de la traduction neuronale.\nTravaux en collaboration avec Behnoosh Namdarzareh, Guillaume Wisniewski, Jean-Baptiste Yunès, François Yvon, Maria Zimina-Poirot et Lichao Zhu.\nLaurianne Sitbon : Co-designing intelligent systems that can enrich inclusive pictorial communication\nBridging the gap between advances in artificial intelligence and assistive technologies, I will discuss why and how we can co-design pictorial visual communication applications with people with intellectual disability or autism and community members. There is little research that explores the opportunities that multi-modal conversational systems present for people who are not most comfortable communicating linguistically, such as people with intellectual disability. Yet, access to online materials offers unlimited resources to support communication and connection with others. I will present my team’s research on how computers can facilitate rich pictorial communication environments, thereby supporting self-expression and inclusion.\n","date":"8 December 2022","externalUrl":null,"permalink":"/en/seminars/2022-12-08-ballier/","section":"Team seminars","summary":"","title":"Travaux en cours sur la traduction neuronale / Co-designing intelligent systems that can enrich inclusive pictorial communication","type":"seminars"},{"content":"Recent works showed that end-to-end neural approaches tend to become very popular for spoken language understanding (SLU). Through the term end-to-end, one considers the use of a single model optimized to extract semantic information directly from the speech signal. A major issue for such models is the lack of paired audio and textual data with semantic annotation. In this work, we propose an approach to build an end-to-end neural model to extract semantic information in a scenario in which zero paired audio data is available. Our approach is based on the use of an external model trained to generate a sequence of vectorial representations from text. These representations mimic the hidden representations that could be generated inside an end-to-end automatic speech recognition (ASR) model by processing a speech signal. An SLU neural module is then trained using these representations as input and the annotated text as output. Last, the SLU module replaces the top layers of the ASR model to achieve the construction of the end-to-end model. Our experiments on named entity recognition, carried out on the QUAERO corpus, show that this approach is very promising, getting better results than a comparable cascade approach or than the use of synthetic voices.\n","date":"1 December 2022","externalUrl":null,"permalink":"/en/seminars/2022-12-01-mdhaffar/","section":"Team seminars","summary":"","title":"End-to-end model for named entity recognition from speech without paired training data","type":"seminars"},{"content":"Amira Barhoumi : Sentiment analysis: document representation, machine learning methods and corpora\nSentiment Analysis (SA) is a growing field in both industry and academia. It consists of determining the polarity (or sentiment) of a reviewer towards an entity or specific entity aspects in reviews. I will present an overview of the SA field. Firstly, I will show a specific protocol to fix the length of reviews. Secondly, I will present some deep learning methods and highlight how to take into account the specificities of the language in which reviews are written (Arabic language as a use case). Then, I will discuss some limitations of available corpora. Finally, I will present how to explore citation polarities in scientific papers within the framework of the NanoBubbles project.\nMartin Lentschat : n-Ary relations extraction and claims in scientific documents\nThe automatic extraction of knowledge in scientific documents is an important step that could help in the cross-checking, diffusion and exploitation of research data. However, this raises a series of questions related to the language used in each domain, the targeting of specific information, and the representation of knowledge. I will present the challenges specific to those tasks in experimental fields, and a method that uses semantic and lexical resources to extract and represent knowledge related to packaging permeability as n-Ary relations. I will then discuss how it can be transferred to a different domain and my future contributions to the NanoBubbles project.\n","date":"15 September 2022","externalUrl":null,"permalink":"/en/seminars/2022-09-15-barhoumi/","section":"Team seminars","summary":"","title":"Sentiment analysis: document representation, machine learning methods and corpora / n-Ary relations extraction and claims in scientific documents","type":"seminars"},{"content":"Sujet de thèse ou de postdoctorat dans le cadre du projet Popcorn (projet collaboratif avec deux entreprises)\nPeuplement OPérationnel de bases de COnnaissances et Réseaux Neuronaux # Le projet aborde le problème de l’enrichissement semi-automatisé d’une base de connaissance au travers de l’analyse automatique de textes. Afin d’obtenir une innovation de rupture dans le domaine du Traitement Automatique du Langage Naturel (TALN) pour les clients sécurité et défense, le projet se focalise sur le traitement du français (même si les approches retenues seront par la suite généralisables à d’autres langues). Les travaux aborderont différents aspects :\nL’annotation automatique de documents textuels par la détection de mentions d’entités présentes dans la base de connaissance et leurs désambiguïsation sémantique (polysémie, homonymie) ; La découverte de nouvelles entités (personnes, organisations, équipements, événements, lieux), de leurs attributs (âge d’une personne, numéro de référence d’un équipement, etc.), et des relations entre entités (une personne travaille pour une organisation, des personnes impliquées dans un événement, \u0026hellip;). Une attention particulière sera donnée au fait de pouvoir s’adapter souplement à des évolutions de l’ontologie, la prise en compte de la place de l’utilisateur et de l’analyste pour la validation/capitalisation des extractions effectuées. Le projet se focalise autour des trois axes de recherches suivants :\nGénération de données synthétiques textuelles à partir de textes de référence ; La reconnaissance des entités d’intérêt, des attributs associés et des relations entre les entités. La désambiguisation sémantique des entités (en cas d’homonymie par exemple) Profil recherché: # Solide expérience en programmation \u0026amp; machine learning pour le Traitement Automatique de Langues (TAL), notamment l’apprentissage profond Master/Doctorat Machine Learning ou informatique, une composante TAL ou linguistique computationnelle sera un plus apprécié Bonne connaissance du français Détails pratiques: # Début de la thèse ou du PostDoc dès que possible Contrat doctoral à temps plein au LIG (équipe Getalp) pour 3 ans (salaire: min 1768€ brut mensuel) ou Contrat postdoctoral à temps plein au LIG (équipe Getalp) pour 24 mois (salaire: min 2395€ brut mensuel) Environnement scientifique: # Le doctorat ou le postdoctorat sera mené au sein de l\u0026rsquo;équipe Getalp du laboratoire LIG (http://lig-getalp-new.imag.fr/). La personne recrutée sera accueillie au sein de l’équipe qui offre un cadre de travail stimulant, multinational et agréable. Les moyens pour mener à bien le (post)doctorat seront assurés tant en ce qui concerne les missions en France et à l’étranger qu’en ce qui concerne le matériel (ordinateur personnel, accès aux serveurs GPU du LIG, Grille de calcul Jean Zay du CNRS). Comment postuler ? # Pour postuler sur une thèse de doctorat, les candidats doivent être titulaires d\u0026rsquo;un Master en informatique, en apprentissage machine ou en traitement automatique du langage naturel (obtenu avant le début du contrat doctoral). Pour postuler sur un postdoctorat, les candidats doivent être titulaires d\u0026rsquo;une thèse de doctorat en informatique, en apprentissage machine ou en traitement automatique du langage naturel (obtenu avant le début du contrat doctoral, les étudiants dont la soutenance est prévue avant fin septembre 2022 peuvent ainsi postuler). Ils doivent avoir une bonne connaissance des méthodes d’apprentissage automatique et idéalement une expérience en collecte et gestion de corpus. Ils doivent également avoir une bonne connaissance de la langue française. Les candidatures doivent contenir : CV + lettre/message de motivation + notes de master + lettre(s) de recommandations; et être adressées à Benjamin Lecouteux (benjamin.lecouteux@univ-grenoble-alpes.fr), Gilles Sérasset (gilles.serasset@univ-grenoble-alpes.fr) et Didier Schwab (Didier.Schwab@univ-grenoble-alpes.fr)\n","date":"5 September 2022","externalUrl":null,"permalink":"/en/news/2022/2022-09-05-call-thesis-postdoc-information-extraction/","section":"News","summary":"","title":"Appel à candidature : Thèse (3 ans) ou PostDoc (2 ans) en extraction d'informations","type":"news"},{"content":"A consistent theme in recent NLP research has been doing more with less. In the realm of research in the low-resource languages spectrum, it is popular to describe new pipelines to solve a wide variety of tasks (POS tagging, ASR pipeline, translation etc.). Indigenous communities around the world are not homogenous. They have different needs, priorities, and challenges that do not always follow the one-size-fits-all paradigm. In this work, I will present my journey as NLP scientist working at the intersection of speech processing and Indigenous research to enable a community-based language documentation framework.\n","date":"16 June 2022","externalUrl":null,"permalink":"/en/seminars/2022-06-16-ferrand/","section":"Team seminars","summary":"","title":"Towards community-based language documentation","type":"seminars"},{"content":"How can language technology address the diverse situations of the world\u0026rsquo;s languages? In one view, languages exist on a resource continuum and the challenge is to scale existing solutions, bringing under-resourced languages into the high-resource world. In another view, presented here, the world\u0026rsquo;s language ecology includes standardised languages, local languages, and contact languages. These are often subsumed under the label of `under-resourced languages\u0026rsquo; even though they have distinct functions and prospects.\nI explore this position and propose some ecologically-aware language technology agendas.\n","date":"2 June 2022","externalUrl":null,"permalink":"/en/seminars/2022-06-02-bird/","section":"Team seminars","summary":"","title":"Local languages, third spaces, and other high-resource scenarios","type":"seminars"},{"content":"Companies and research groups are increasingly aiming to deploy mobile robots to perform various navigation tasks in environments shared with humans. Various social norms have been implemented into social navigation algorithms, but the evaluation of these methods typically studies whether the robot’s navigation is “natural” and “acceptable”, or “uncomfortable”.\nBeyond respecting social norms, humans also use their movement and positioning as a means of communication and interaction, often in combination with other communication channels such as speech. Our first goal is to study how the manner in which a robot moves impacts how people perceive it and interact with it. We take inspiration from the study of vocal prosody, where the (often subtle) manner in which a word or sentence is spoken is a communication signal. Our second goal is to use this understanding of what we call \u0026ldquo;movement prosody\u0026rdquo; in order to design a social navigation architecture that can alter the manner in which a navigation task is performed according to how the robot should be perceived.\nTo achieve both of these goals, we propose an iterative design process, alternating between Human-Robot Interaction experiments and Social Navigation Algorithm design. As a first step in our work, we designed an adaptive social navigation architecture that offers a flexible formulation to control all of the robot\u0026rsquo;s motion parameters. Secondly we conducted two perception experiments to characterize the impact of motion and appearance parameters on people\u0026rsquo;s perceptions of a robot. Based on the preliminary results of these experiments, we extracted the most significant parameters which will be controlled by our navigation architecture. Two further experiments will be conducted in order to observe how people react to the robot when it uses our algorithm to navigate with various parameter sets.\n","date":"12 May 2022","externalUrl":null,"permalink":"/en/seminars/2022-05-12-scales/","section":"Team seminars","summary":"","title":"Adaptive Social Navigation for Robots operating in Human Environments in real ecological use-cases","type":"seminars"},{"content":"This seminar presents BA-LR, a new approach in automatic voice comparison developed in the framework of the LIAvignon chair and Imen Benamor\u0026rsquo;s thesis work. This approach is dedicated to explicabilty and interpretability aspects, contrarily to classical approaches more focused on \u0026ldquo;accuracy\u0026rdquo;. These aspects take a particular importance in the field of voice comparison because this field raises strong societal and ethical issues, mainly related to forensic applications.\nThe first part of the seminar defines automatic voice comparison and proposes a brief introduction on what means explicability and interpretability in AI. Voice comparison is used here as a a case study. The second part of the seminar is devoted to the BA-LR approach.While classical voice comparison approaches produce a single value, the LR, without further explanation, our approach proposes to see the output of the system as the composition of a set of factors, related to the voice attributes. Moreover, the behavior of each attribute is described by three easily understood parameters. Finally, in addition to building a voice comparison system that is more explicit and understandable than standard solutions, our approach allows us to stay within the main paradigm of voice comparison, the likelihood ratio, which is known to be the \u0026ldquo;correct and logical\u0026rdquo; paradigm for forensic applications. Next the limits of the current implementation and the remaining steps to be taken for a fully transparent approach, where each piece of information and decision is defined precisely and understandable by all, is discussed. Finally, a conclusion is proposed, keeping in mind foransic application of voice comparison.\n","date":"17 March 2022","externalUrl":null,"permalink":"/en/seminars/2022-03-17-bonastre/","section":"Team seminars","summary":"","title":"The Binary-Attribute Likelihood Ratio estimation (BA-LR) approach","type":"seminars"},{"content":"Le projet UNseL, financé par la DGA de 2019 à 2021, vise à fournir des outils pour accompagner la spécification de \u0026ldquo;systèmes de systèmes\u0026rdquo; (par exemple, un système de communications sol-air pour un aéroport ou un système de freinage d’urgence). Dans ce cadre, nos travaux proposent l\u0026rsquo;application d\u0026rsquo;une méthode d\u0026rsquo;extraction de contenu sémantique dans un contexte industriel, avec pour objectif la vérification automatique d\u0026rsquo;exigences systèmes rédigées en langue naturelle non contrainte. L\u0026rsquo;étape d\u0026rsquo;extraction utilise une analyse par transduction sémantique, implémentée en s\u0026rsquo;appuyant sur les standards du Web Sémantique du W3C. Elle part d\u0026rsquo;une représentation sémantique des textes, exprimée sous forme de graphes UNL (Universal Networking Language) à \u0026ldquo;sens garanti\u0026rdquo; (obtenu grâce à une étape de désambiguïsation interactive), qui produit d\u0026rsquo;abord une structure semi-formelle et indépendante de la langue source. Les outils développés construisent ensuite automatiquement une ontologie OWL à partir des spécifications du système, exprimées par des énoncés non contraints. Finalement, une vérification automatique des exigences est réalisée à l\u0026rsquo;aide de règles SPARQL génériques et de raisonneurs logiques. Une mise en pratique a été réalisée sur des exigences extraites d\u0026rsquo;une spécification réelle.\n","date":"20 January 2022","externalUrl":null,"permalink":"/en/seminars/2022-01-20-lamercerie/","section":"Team seminars","summary":"","title":"Extraction de contenus sémantiques pour la vérification d’exigences systèmes","type":"seminars"},{"content":"Over the past decade, the confluence of sophisticated algorithms and tools, computational infrastructure, and data science has fueled a machine learning revolution across multiple fields, including speech and handwriting recognition, natural language processing, computer vision, social network filtering, and machine translation. Ensuing advances are changing the way we interact with technology in our daily lives. This is particularly salient when it comes to user input on mobile devices, be it speech, handwriting, touch, keyboard, or camera input. Increased input intelligence boosts device responsiveness across languages, improving not only basic abilities like tokenization, named entity recognition and part-of-speech tagging, but also more advanced capabilities like statistical language modeling and question answering. In this talk, I will give selected examples of what we are doing at Apple to impart input intelligence to mobile devices, with two overarching themes as sub-text: (i) enhancing interaction experience through machine learning, and (ii) transforming users\u0026rsquo; digital lives without sacrificing their privacy.\nDr. Jerome R. Bellegarda is Apple Distinguished Scientist in Intelligent System Experience at Apple Inc., Cupertino, California, which he joined in 1994. Prior to that, he was a Research Staff Member at the IBM T.J. Watson Center, Yorktown Heights, New York. He received the Ph.D. degree in Electrical Engineering from the University of Rochester, Rochester, New York, in 1987. Among his diverse contributions to speech and language advances over the years, he pioneered the use of tied mixtures in acoustic modeling and latent semantics in language modeling. In addition, he was instrumental to the due diligence process leading to Apple\u0026rsquo;s acquisition of Siri personal assistant technology and its integration into the Apple ecosystem. His general interests span machine learning applications, statistical modeling algorithms, natural language processing, man-machine communication, multiple input/output modalities, and multimedia knowledge management. In these areas he has written close to 200 publications, and holds over 100 U.S. and foreign patents. He has served on many international scientific committees, review panels, and advisory boards. In particular, he has worked as Expert Advisor on speech and language technologies for both the U.S. National Science Foundation and the European Commission, served on the IEEE Signal Processing Society (SPS) Speech Technical Committee, was Associate Editor for the IEEE Transactions on Audio, Speech and Language Processing, is currently an Editorial Board member for Speech Communication, and will be a 2022 IEEE SPS Distinguished Industry Speaker. He is a Fellow of both IEEE and ISCA (International Speech Communication Association).\n","date":"1 December 2021","externalUrl":null,"permalink":"/en/seminars/2021-12-01-bellegarda/","section":"Team seminars","summary":"","title":"Input Intelligence on Mobile Devices","type":"seminars"},{"content":"Caroline Rossi commencera par définir le type d\u0026rsquo;explication recherché. L’explicabilité des systèmes neuronaux est souvent définie en distinguant trois niveaux : les systèmes opaques (systèmes propriétaires, dont le fonctionnement n’est pas décrit), les systèmes interprétables (également dits transparents, ou pour lesquels le passage de données d’entrée à la sortie du système est bien décrit) et les systèmes explicables (Doran et al. 2017). C’est ce dernier niveau qui est visé dans le présent projet, où l’explicabilité est fonction de notre capacité à expliquer, sans nécessairement supposer que l’on puisse atteindre une certaine transparence quant au fonctionnement des systèmes de TAN. Elle abordera ensuite les différentes étapes du projet, qui vont du repérage d\u0026rsquo;erreurs dans ACCOLE au développement d\u0026rsquo;une solution logicielle liant le repérage des erreurs probables à une base de données d’explications sur ces erreurs.\n","date":"21 October 2021","externalUrl":null,"permalink":"/en/seminars/2021-10-21-rossi/","section":"Team seminars","summary":"","title":"ExTra : expliquer la traduction automatique neuronale (TAN) pour former des utilisateurs avertis","type":"seminars"},{"content":"à Toulouse du 1er au 5 juillet organisé conjointement avec la Plateforme d\u0026rsquo;Intelligence Artificielle, Loïc Vial, Benjamin Lecouteux et Didier Schwab ont obtenu le prix du meilleur article pour Compression de vocabulaire de sens grâce aux relations sémantiques pour la désambiguïsation lexicale.\nCet article présente une méthode originale qui pallie le manque de données annotés de bonne qualité et qui permet d’obtenir des résultats qui surpassent largement l’état de l’art sur toutes les tâches d’évaluation de la désambiguïsation lexicale.\nPour rappel, la Désambiguïsation Lexicale est une tâche qui vise à clarifier un texte en assignant à chacun de ses mots l’étiquette de sens la plus appropriée depuis un inventaire de sens prédéfini. Il s\u0026rsquo;agira, par exemple, de préférer dans la phraseLa souris mange le fromage le sens de rongeur plutôt que le sens de dispositif électronique pour le mot souris. Ces travaux sont exploités par les auteurs dans plusieurs applications du traitement automatique des langues comme la traduction automatique ou pour concevoir des outils destinés à établir une communication alternative par exemple pour pour des personnes maîtrisant pas ou peu la langue ou des personnes en situation de polyhandicap.\n","date":"12 October 2021","externalUrl":null,"permalink":"/en/news/2021/2019-07-07-best-paper-taln-2019/","section":"News","summary":"","title":"Best paper award for GETALP at TALN 2019","type":"news"},{"content":"In this presentation I will talk about our paper “Facebook Ads Monitor: An Independent Auditing System for Political Ads on Facebook” published at The Web Conference 2020 and followup discussions with civil societies on how to regulate political advertising: https://epd.eu/wp-content/uploads/2020/09/joint-call-for-universal-ads-transparency.pdf.\nPaper abstract: The 2016 United States presidential election was marked by the abuse of targeted advertising on Facebook. Concerned with the risk of the same kind of abuse to happen in the 2018 Brazilian elections, we designed and deployed an independent auditing system to monitor political ads on Facebook in Brazil. To do that we first adapted a browser plugin to gather ads from the timeline of volunteers using Facebook. We managed to convince more than 2000 volunteers to help our project and install our tool. Then, we use a Convolution Neural Network (CNN) to detect political Facebook ads using word embeddings. To evaluate our approach, we manually label a data collection of 10k ads as political or non-political and then we provide an in-depth evaluation of proposed approach for identifying political ads by comparing it with classic supervised machine learning methods. Finally, we deployed a real system that shows the ads identified as related to politics. We noticed that not all political ads we detected were present in the Facebook Ad Library for political ads. Our results emphasize the importance of enforcement mechanisms for declaring political ads and the need for independent auditing platforms.\nOana Goga is a tenured CNRS research scientist in the Laboratoire d’Informatique Grenoble (France) since October 2017. Prior to this, she was a postdoc at the Max Plank Institute for Software Systems and obtained a PhD in 2014 from Pierre et Marie Curie University in Paris. She is the recipient of a young researcher award from the French National Research Agency (ANR). Her research interests are in security and privacy issues that arise in online systems that have at their core user provided data. Her recent research received a Honorable Mention Award at The Web Conference in 2020, obtained the 2020 CNIL-Inria Award for Privacy Protection and was runner-up for the 2019 Caspar Bowden PET Award for outstanding research in privacy enhancing technologies and runner-up for the 2019 CNIL-Inria Award for Privacy Protection.\n","date":"29 April 2021","externalUrl":null,"permalink":"/en/seminars/2021-04-29-goga/","section":"Team seminars","summary":"","title":"Investigating paths towards safe online political advertising","type":"seminars"},{"content":"Les modèles de langue pré-entraînés sont désormais indispensables pour obtenir des résultats à l’état-de-l’art dans de nombreuses tâches du TALN. Tirant avantage de l’énorme quantité de textes bruts disponibles, ils permettent d’extraire des représentations continues des mots, contextualisées au niveau de la phrase. L’efficacité de ces représentations pour résoudre plusieurs tâches de TALN a été démontrée récemment pour l’anglais. Dans cette présentation, nous présentons FlauBERT, un ensemble de modèles appris sur un corpus français hétérogène et de taille importante. Des modèles de complexité différente ont été entraînés à l’aide du nouveau supercalculateur Jean Zay du CNRS. Nous évaluons nos modèles de langue sur diverses tâches en français (classification de textes, paraphrase, inférence en langage naturel, analyse syntaxique, désambiguïsation automatique) et montrons qu’ils surpassent souvent les autres approches sur le référentiel d’évaluation FLUE également présenté ici.\n","date":"14 January 2021","externalUrl":null,"permalink":"/en/seminars/2021-01-14-schwab/","section":"Team seminars","summary":"","title":"FlauBERT : des modèles de langue contextualisés pré-entraînés pour le français","type":"seminars"},{"content":"The field of translation is no exception when it comes to the increasing importance of data collection and data-driven technologies, be it from the perspective of professional translation services or the scientific inquiry into translation. However, academia and industry usually have different data needs. Consequently, there is no data repository that suits translation industry and academia alike.\nThe corpus building initiative “TransBank – A Meta-Corpus for Translation Research” (https://www.transbank.info) addresses this gap by providing an infrastructure for the collaborative collection and sharing of open-access translation data. The main aim of TransBank is to build a general-purpose translation repository, making real-world translation data easily and freely accessible to anyone in the global translation community.\nIn this talk, the motivation behind the TransBank initiative as well as main features of the repository architecture will be outlined. In addition, major theoretical and technical challenges with regard to data collection, data processing and data provision will be discussed.\nMichael Ustaszewski is assistant professor in the Department of Translation Studies at the University of Innsbruck, Austria. He is a trained translator and holds a doctoral degree in linguistics, obtained from the University of Innsbruck, as well as a master’s degree in computational linguistics, obtained from the University of the Basque Country. His research interests lie in the intersection of translation, corpus linguistics and language technology, with a focus on corpus-based translation studies. He is principal investigator of the corpus-building project “TransBank – A Meta-Corpus for Translation Research”.\n","date":"20 February 2020","externalUrl":null,"permalink":"/en/seminars/2020-02-20-ustaszewski/","section":"Team seminars","summary":"","title":"TransBank – A general-purpose translation repository","type":"seminars"},{"content":"Gérard Bailly (1), Erika Rassat (1,2), Anne-Laure Piat-Marchand (1) \u0026amp; Marie-Line Bosse (2)\n(1) GIPSA-Lab\n(2) LPNC\nAccording to the ministry of education, one young adult (16-25) over 10 has reading diffulties, 50% of them being illiterate. France was ranked 34/50 in the PIRLS 2016. A degradation of the reading performance of these 4th grade children was also noticed.\nMastering comprehensive reading is yet a prequisite for accessing other educational disciplines. But reading requires the maturationand synchronization of a complex cognitive network, involving vision, phonological, semantic and pragmatic processing together with the sensorimotor activation of phonetic representations.\nWe will present the framework we developped so far in order to characterize and assess the reading fluency of young readers.\nThis work is performed in the context of the e-FRAN Fluence project, where hundreds of primary schoolers are trained via computer-assisted technologies and monitored in a longitudinal study.\n","date":"13 February 2020","externalUrl":null,"permalink":"/en/seminars/2020-02-13-bailly/","section":"Team seminars","summary":"","title":"Characterizing and assessing the reading fluency of young readers","type":"seminars"},{"content":"L\u0026rsquo;informatique et le Web réinventent notre culture. Qu\u0026rsquo;en est-il de l\u0026rsquo;objet dictionnaire ? Est-il possible de s\u0026rsquo;affranchir de la copie électronique du format papier pour proposer de nouvelles structures de dictionnaire ? Peut-on réussir à créer de nouvelles ressources au moyen de projets collaboratifs et coopératifs ? Comment peut-on interagir avec d\u0026rsquo;autres outils en échangeant des données lexicales ? L\u0026rsquo;objectif de ce mémoire est de répondre à ces questions en concevant des environnements de gestion de bases lexicales accessibles au plus grand nombre pour consulter et éditer des données lexicales sur le Web et de les utiliser dans d’autres applications, principalement pour la compréhension de textes et l’apprentissage des langues.\niPoLex est un entrepôt de données lexicales qui s\u0026rsquo;appuie sur un devin de microstructures afin d\u0026rsquo;analyser et de produire une description fine. Celle-ci servira pour l\u0026rsquo;import des données dans Jibiki, une plateforme de gestion de ressources lexicales en ligne munie d\u0026rsquo;interfaces de consultation, d\u0026rsquo;édition et de programmation (API). Ces outils sont utilisés dans plusieurs projets de construction de dictionnaires tels que la base lexicale multilingue Papillon, le dictionnaire estonien français GDEF, le dictionnaire français-khmer MotÀMot, les dictionnaires DiLAF Langue Africaine-Français et le dictionnaire japonais-français jibiki.fr. Certains de ces projets réutilisent des données issues de lecture optique, de traitements de textes ou de conversion de structures XML. L\u0026rsquo;interface de programmation de Jibiki permet d\u0026rsquo;outiller l\u0026rsquo;accès aux textes à travers la lecture active et les environnements d’apprentissage des langues.\n","date":"16 January 2020","externalUrl":null,"permalink":"/en/seminars/2020-01-16-mangeot/","section":"Team seminars","summary":"","title":"Génie lexico-sémantique multilingue contributif","type":"seminars"},{"content":"In this seminar, I will present an overview of my work on discontinuous constituency parsing.\nConstituency trees represent the syntactic structure of a sentence by a recursive grouping of tokens into labelled phrases such as NP, PP, VP, etc.\nHowever, standard constituency trees are not adequate to model a number of syntactic phenomena related to word order variations, e.g. long distance extractions.\nIn contrast, discontinuous constituency trees use crossing branches to model these phenomena.\nDiscontinuous structures correspond to derivations of mildly context-sensitive formal grammars, such as Linear Context-Free Rewriting Systems (LCFRS), and are therefore much more difficult to parse than projective constituency trees (CFG derivations).\nIn this talk, I will present efficient data-driven parsing algorithms for discontinuous constituency trees, as well as empirical parsing results on English and German discontinuous parsing.\nI will also mention the release of wikiparse: wikipedia automatically annotated with constituency trees, dependency trees and morphological analyses (POS tags, morphological features) for 7 languages.\nReferences:\nhttps://www.aclweb.org/anthology/E17-1118.pdf https://www.aclweb.org/anthology/Q19-1005.pdf https://www.aclweb.org/anthology/N19-1018.pdf https://hal.archives-ouvertes.fr/tel-02302563 wikiparse: data download: http://www.llf.cnrs.fr/wikiparse/ scripts: https://github.com/mcoavoux/wiki_parse ","date":"12 December 2019","externalUrl":null,"permalink":"/en/seminars/2019-12-12-coavoux/","section":"Team seminars","summary":"","title":"Algorithms for Discontinuous Constituency Parsing","type":"seminars"},{"content":"In Natural Language Generation (NLG), End-to-End models have recently gained a strong interest. Such models need a large amount of carefully annotated data to reach satisfactory performance. However, acquiring such datasets for every new NLG application is a tedious and time-consuming task. In the first part of my talk I will present our efforts on collecting a newly created and publicly available company dataset that has been collected from Wikipedia. The dataset consists of around 51K company descriptions that can be used for both concept-to-text and text-to-text generation tasks. We study the performance of several End-to-End models applied to generation of short company descriptions and discuss the challenge in evaluating models trained on such data. In the second part of the talk, I will present a semi-supervised deep learning scheme that can learn from non-annotated data and annotated data when available. It uses an NLG and a Natural Language Understanding (NLU) End-to-End models which are learned jointly to compensate for the lack of annotation. Our experiments on two benchmark datasets show that, with limited amount of annotated data, the method can achieve competitive results while not using any preprocessing or re-scoring tricks. I will also briefly talk about a PyTorch based sequence-to-sequence model that we developed for this project: https://gricad-gitlab.univ-grenoble-alpes.fr/getalp/seq2seqpytorch\n","date":"21 November 2019","externalUrl":null,"permalink":"/en/seminars/2019-11-21-qader/","section":"Team seminars","summary":"","title":"Neural Natural Language Generation with Limited Annotated Data","type":"seminars"},{"content":"L’objectif initial qui a guidé le développement d’ACCOLÉ, est l’annotation manuelle des erreurs de traduction selon des critères linguistiques. L’idée sous-jacente est de pouvoir fournir à un utilisateur une aide dans le choix d’un système de TA à utiliser selon le contexte (compétences linguistiques et informatiques de l’utilisateur, connaissance du domaine du document source à traduire et la tâche pour laquelle il a besoin de traduire le document source.) Pour ce faire, ACCOLÉ doit permettre de détecter quels sont les phénomènes linguistiques qui ne sont pas traités correctement par le système de TA étudié. Nous proposons sur la même plateforme une palette de services permettant de répondre aux besoins d’analyse d’erreurs de traduction. Ainsi, les principales fonctionnalités de la plateforme ACCOLÉ sont la gestion simplifiée des corpus, des typologies d’erreurs, des annotateurs, etc. ; l’annotation d’erreurs ; la collaboration et/ou supervision lors de l’annotation ; la recherche de modèles d\u0026rsquo;erreurs (type d\u0026rsquo;erreurs dans un premier temps, patrons morphosyntaxiques ultérieurement) dans les annotations. La tâche d\u0026rsquo;analyse d\u0026rsquo;erreurs étant déjà fastidieuse, il est important que les personnes la réalisant aient un accès simple à l\u0026rsquo;outil ainsi qu\u0026rsquo;au corpus qu\u0026rsquo;ils souhaitent analyser. La plateforme est donc disponible en ligne (http://lig-accole.imag.fr) sur un navigateur et ne nécessite aucune installation spécifique.\n","date":"17 October 2019","externalUrl":null,"permalink":"/en/seminars/2019-10-17-esperanca-rodier/","section":"Team seminars","summary":"","title":"ACCOLÉ : Annotation Collaborative d'erreurs de traduction pour COrpus aLignÉs","type":"seminars"},{"content":"The General Conference of UNESCO (Paris, 1999) approved February 21 as International Mother Language Day. The idea is to promote linguistic and cultural diversity and multilingualism in all aspects of public life, particularly in education. The focus here is on mother-tongue-based education for all children in the world. Subsequently, UNESCO Ad Hoc Expert Group on Endangered Languages (2003) includes “materials for language education and literacy” as one of the nine suggested factors to assess the vitality of language. Several issues are raised here with respect to the notions of ‘language of education’, ‘orality vs. literacy’, ‘literacy and information technology’, among others:\nWhy is mother-tongue-based education important?\nWhat challenges do multilingual communities face?\nWhat changes has the globalized digital age brought to languages in oral tradition?\nHow foreign is it the currently used script for indigenous languages?\nWhat can information technology do to languages that once had their own script?\nIn this talk, I report on my fieldwork in extraordinarily multilingual communities in Mindanao, the Philippines, and respond to the above addressed issues. In particular, I clarify why mother-tongue-based education is important for children’s sustainable development of conceptual thinking, and introduce a challenging enterprise to revive the understudied minority languages with the lost ancient script called Baybayin, which is anticipated as the script of their own. (This research was made possible by my research collaborators in Mindanao since 2014.)\n","date":"10 October 2019","externalUrl":null,"permalink":"/en/seminars/2019-10-10-katada/","section":"Team seminars","summary":"","title":"The Realm of Unknown Oral Languages and Its Interface with Information Technology: A Report from Mindanao","type":"seminars"},{"content":"The complementary nature between Natural Language and Structured Knowledge Bases has inspired a lot of applications that involve interactions in between, such as Information Extraction, Question Answering and Natural Language Generation.\nNatural language processing models responsible for those interactions are usually designed in a static way, with the abundance of continuously published new information on the web they require a tedious process of maintenance and adaptation.\nIn this direction, I’ll present several contributions from My PhD. thesis to tackle this issue. Firstly, T-REx an architecture for automatically generating alignment datasets between natural language text and structured knowledge bases, alongside with the largest available dataset of the same kind.\nSecondly, I’ll present an encoder-decoder neural network architecture for question generation from knowledge base triples in a zero-shot setup; this model occupied with part-of-speech tag copy actions has proven its capacity of generating questions for unseen relation and entity types during test time, and hence has a large potential of being used in data-augmentation techniques for training Question Answering systems in an evolving way. Finally, I’ll present several delexicalization techniques to generate entity descriptions from structured triples, this method has proven efficiency to generate entity summaries in an under-served multilingual setup with limited training datasets.\n","date":"26 September 2019","externalUrl":null,"permalink":"/en/seminars/2019-09-26-elsahar/","section":"Team seminars","summary":"","title":"Enabling Dynamic Interactions between Natural Language and Structured Knowledge Bases","type":"seminars"},{"content":"Spontaneous Speech Recognition. Application to Situated Corpora in French.\nMore details here.\n","date":"23 August 2019","externalUrl":null,"permalink":"/en/news/2019/2019-08-23-postdoc-proposal-2019/","section":"News","summary":"","title":"Postdoc proposal (18 months)","type":"news"},{"content":"According to the H\u0026amp;H theory (Lindblom, 1970), speakers adapt their articulation effort to the level of intelligibility required by the communicative situation.\nIn a first part of my talk, I will present different studies on speech adaptation to a noisy environment (also called the Lombard effect), showing that speech modifications observed in such a perturbed situation do not limit to hyper-articulation, but reflect a set of different communicative strategies aiming, among others, at improving speech audibility (detection, segregation from the background noise, depending on the noise type), improving segment (audiovisual) intelligibility and facilitating discourse structure and segmentation (through the enhancement of prosodic cues). I will discuss how such adaptation strategies vary between individuals – in particular how some of them make use of the visual modality to improve one’s intelligibility in perturbed conditions – and how they can be affected by the experimental paradigm (reading task vs. interactive game; noise played through headphones vs. loudspeakers).\nIn a second part of my talk, I will go deeper into that speaker-listener interaction loop and present some resent work on how exactly a speaker estimates the required level of intelligibility and adapts his speech to meet the listener’s needs. I will present a study of the different audiovisual markers of understanding vs. misunderstanding emitted by the listener, from which the speaker can estimate online his actual intelligibility. I will also present an ongoing study on speech correction following a misunderstanding, exploring how the speaker’s clarification depends (or not) on what the listener actually perceived.\nI will conclude my talk with some perspectives on speech adaptation and clarification in face-to-face interaction.\n","date":"20 June 2019","externalUrl":null,"permalink":"/en/seminars/2019-06-20-garnier/","section":"Team seminars","summary":"","title":"Speech clarification in face-to-face interaction","type":"seminars"},{"content":"Si Traitement Automatique des Langues et Apprentissage des Langues Assisté par Ordinateur sont intimement liés depuis leurs origines (ou presque), cette alliance ne s\u0026rsquo;est pas faite sans heurts et à l\u0026rsquo;heure actuelle les systèmes d\u0026rsquo;Apprentissage des Langues embarquant de véritables technologies de TAL ne sont pas légion. Toutefois, le TAL s\u0026rsquo;immisce de plus en plus largement dans les technologies du quotidien. À l\u0026rsquo;heure du “Big Data”, l\u0026rsquo;intégration des technologies du TAL aux systèmes tournés vers l\u0026rsquo;ALAO pose des questions méthodologiques.\n","date":"16 May 2019","externalUrl":null,"permalink":"/en/seminars/2019-05-16-loiseau/","section":"Team seminars","summary":"","title":"ICALL ou iCALL — Questions posées par l’intégration du TAL dans l’ALAO ?","type":"seminars"},{"content":"Souvent saluée pour ses effets de démocratisation (Prado 2010 ; Boitet et al. 2010 ; Goltz 2017) et sa contribution au maintien de la diversité linguistique (Cronin 2013 : 59), la TA peut également être considérée comme fondée sur des mythes trompeurs (l’universalité du sens ou la transparence des formes linguistiques), et symptôme d\u0026rsquo;une vision purement instrumentale du langage (voir par exemple Raley 2003 ; ou encore Cronin 2013). Paradoxalement peut-être, elle est aussi souvent interprétée comme un soutien à l\u0026rsquo;hégémonie culturelle continue de l\u0026rsquo;anglais (Raley 2003 ; Poibeau 2017 : 168). Enfin, si elle constitue une technologie clé pour les grandes entreprises mondialisées (Poibeau 2017 : 6), elle est également impliquée dans la baisse des coûts de traduction, que subissent actuellement les traducteurs humains indépendants (Moorkens 2017). Il n\u0026rsquo;est donc pas étonnant que de nombreux travaux en traductologie soient porteurs de représentations négatives de la TA. Les points de rencontre sont cependant de plus en plus nombreux, non seulement parce que la façon dont les développeurs de TA abordent leur tâche peut à la fois refléter et aider à construire la compréhension du langage, du sens et de la traduction (Kenny 2012a), mais aussi parce que c\u0026rsquo;est principalement par l\u0026rsquo;interface avec la TA que les études en traduction rencontrent certaines des questions les plus pressantes de notre époque, liées au regain d\u0026rsquo;intérêt pour l\u0026rsquo;intelligence artificielle et l\u0026rsquo;avenir du travail humain.\nDans cet exposé, nous analysons ces convergences récentes en considérant les usages de la notion de contexte en traduction automatique. Après avoir défini trois grands types de contexte (co-texte, contexte élargi, et contexte situationnel), nous décrivons les principales approches historiquement adoptées en TA et leur utilisation du contexte, afin de montrer que les modèles statistiques puis neuronaux ont permis une prise en compte élargie de celui-ci. Dans un second temps, pour dresser un premier bilan des apports et des limites d’un tel élargissement, nous procédons à l’analyse comparative de sorties de TA statistique et neuronale (avec la paire de langues anglais-français) à partir de deux problèmes souvent relevés: le traitement des anaphoriques (Voita et al. 2018 ; Bawden et al. 2018 ; Hardmeier \u0026amp; Guillou 2018), et celui des métaphores (Isabelle 2017 ; Toral \u0026amp; Way 2018).\n","date":"11 April 2019","externalUrl":null,"permalink":"/en/seminars/2019-04-11-rossi/","section":"Team seminars","summary":"","title":"Le contexte en traduction automatique : un concept à géométrie variable","type":"seminars"},{"content":"Le robot social : utile, futile ou toxique ?\nCe qui sépare le robot social d\u0026rsquo;un objet intelligent connecté n\u0026rsquo;est pas tant la complexité des compétences artificielles qu\u0026rsquo;on lui donne que des caractéristiques que nous ne cernons pas encore et qui le font percevoir comme un \u0026ldquo;autre\u0026rdquo;. Le robot augmente ainsi notre espace social, ce qui est tout à fait nouveau puisque nous avions jusqu’ici « désiré », par les outils techniques et technologies, augmenter seulement les capacités de l’individu (mieux voir, traverser l’atlantique en 7h, augmenter sa mémoire par l’écriture ou le net etc). La question que nous devons tous nous poser sur le fond est pourquoi ce désir aujourd’hui, soudain et effervescent, d’un autre que moi artefactuel, alors que la première statue parlante date de 2000 ans avant l’égypte ancienne. Vaucanson avait réussi à convaincre tous de l’autonomie de ses automates, sans qu’aucun des puissants qui l’accueillaient n’aient souhaité mettre dans son salon le son joueur de flûte ou son canard comme automate compagnon…\nDonc, même si cette rupture entre objet et sujet ne peut pas encore être objectivement rapporté à des caractéristiques techniques (le critère d’autonomie n’est par exemple pas totalement discriminant) on doit constater que les objets technologiques, perçus par nous humain, comme des robots déclenchent en nous une illusion empathique, une illusion d’un autre acceptable dans notre espace social. Une illusion d’un autre avec qui nous croyons communiquer. Or la communication est un élément vital pour l’humain au même titre que l’eau et la nourriture : au 13e siècle un empereur se demandant quelle langue parlerait spontanément un enfant si les personnes qui l’entouraient ne lui adressaient jamais la parole ; il a confié 6 bébés à 6 nourrices, tous les bébés sont morts rapidement. Dans les orphelinats en roumanie en 1990 dans lesquels les enfants ont été à peu près nourris et lavés, mais jamais dans le désir de les considérer communicativement, dans un orphelinat récemment au Maroc d’enfants issus de viols, dont le tabou a involontairement rendu les soignantes non communiquantes, les enfants sont décédés ou sont devenus irrémédiablement psychotiques. De manière moins exceptionnelle, les personnes âgées de plus de 75 ans se fragilisent et meurent 7 fois plus vite q’ils sont en sentiment d’isolement, les opérés cardiaques entre 35 et 45 ans se dégradent 5 fois plus s’ils se sentent isolés, et de plus en plus de jeunes gens au Japon, mais surtout en Europe, souffrent du syndrome d’isolement social hikikimori.\nCela signifie que si à travers nos signaux sociaux : nos paroles, nos regards nos gestes, nos rapprochements ou éloignements, nos manières de nous déplacer ensemble, nous n’échangeons pas assez de nourriture socio-relationnelle, ou de la mauvaise nourriture, nous développons des souffrances intenses, pas toujours conscientes, par ce sentiment d’isolement. Ces fibres sociales qui nous relient, nous les appelons la glu socio-affective et nous avons émis l’hypothèse que lorsque nous tirons sur ces fibres, quand notre culture nous y pousse, avec un sentiment individuel de liberté, nous prenons le risque de les casser et de ne plus pouvoir y laisser passer assez de nourriture utile. Vous l’avez compris, l’hypothèse que nous proposons pour comprendre ce désir aujourd’hui de robot est un déficit de nourriture de l’autre.\nLe robot pourrait ainsi par cette illusion empathique soulager dans un premier temps cette souffrance, comme le miroir posé en face d’un bras valide soulage la douleur fantôme d’un bras amputé. Mais bien sûr le grand danger serait que ce soulagement laisse croire que le robot va réellement nourrir l’humain en mal de l’autre, alors qu’il ne s’agit que d’une illusion, d’un miroir sur lequel l’humain se projette.\nAinsi dans les expériences que nous menons, des personnes, en sentiment variable d’isolement, projettent rapidement des attachements qui soulagent leur douleur d\u0026rsquo;isolement sur le robot Emox. Cet attachement, cette glu, utilise des primitives vocales, gestuelles et proxémiques sans contenus lexicaux, mais dont les caractéristiques prosodiques sont parfaitement contrôlées aussi bien dans leurs contenus que dans la dymanique par laquelle nous « manipulons » la nature de cet attachement. Ces relations s\u0026rsquo;inscrivent dans le modèle DANSE (Dynamics of the Affective Network for Social Entities), à travers une méthode écologique de « Making Thinking » (Fractal) de Living Lab mettant l\u0026rsquo;humain au centre du paradigme technologique de l\u0026rsquo;interaction dont le but est bien de ramener l’humain vers l’humain jusqu’à rendre inutile le robot.\nNous avons ainsi montré que ces bruits purement prosodiques peuvent attacher l’humain d’autant plus vite quand il n’est pas isolé, mais d’autant plus fort et longtemps quand il souffre d’isolement, et ceci sur des axes que nous avons choisis pour ne pas placer l’humain dans la dominance ou la soumission, en explorant un espace proposés de l’altruisme, et dans une relation qu’il perçoit comme positive. Les dimensions de fragilité/robustesse associée à une évaluation visée de « soin tendre » sont particulièrement explorée pour toutes les dimensions de l’artefact robot (son design sa locomotion, ses expressions vocales et visuelles). La théorie D.A.N.S.E., outre explorer cet espace, propose une dynamique co-animée de l’interaction, qui, selon ses paramètres, permet de représenter des états différents du « corps social » - i.e. de la mole émergente d’interactions- en analogie aux états de la matière inerte liquide, solide, gazeux. Au delà des interactions entre humains, ce sont les entités intégrées par l’humain dans son espace social que nous souhaitons observer, analyser, modéliser et simuler, la machine perçue comme entité, cad le robot, étant de plus un outil instrumental ; les entités vivantes non humaines, en particulier les « animaux de compagnie » (pets) permettant d’étendre l’observation des primitives collectées sur l’humain à des primitives plus génériques, voire potentiellement universelles du lien social intra et inter-espèces. Ces primitives identifiées dans l’espace vocal sont plus généralement recherchées dans tous les espaces de la communication directe ou indirecte (par exemple la « navigation »).\nNotre but applicatif est de soulager ainsi la souffrance d’isolement pour que cela amène l’humain à retrouver sa capacité d’entretenir et de faire grandir ses relations avec les autres humains de sa sphère sociale, jusqu’à n’avoir plus aucun désir, plus aucun besoin du robot, et le rejeter.\nMais comment encadrer ce mécanisme afin qu’il ne coupe pas encore plus l’humain des ses relations aux humains, justement car il ne souffre plus quand il est installé dans sa communication avec le robot, dont notre hypothèse est résolument qu’elle est factice et à terme mettra cet humain dans un isolement peut-être irréversible ?\nCette capacité de manipulation, par exemple comme nous l’avons fait par des primitives aussi simples de langage, nous interroge très fortement éthiquement sur les conséquences de ces manipulations, d’autant plus si la raison de l’engouement pour ces technologies se révélait être un grandissement de l’isolement par des relations insuffisantes ou surtout de mauvaise qualité pour notre bien grandir ensemble.\nBien au-delà du robot pensé dès le départ comme compagnon, le robot de services quel qu’il soit, du fait même que nous le percevons robot, est perçu comme communicant ; Cela signifie que l’humain va trouver un sens à ces signaux émis sans but communicatif : que se passe-t-il si cette manipulation implicite, non volontairement contrôlée par le constructeur, et pour cause puisqu’il ne voulait pas explicitement faire de robot social, a des effets psychologiques, comportementaux, sociétaux qui seront toxiques à court ou moyen terme ? Comment et qui rendre responsable de ces effets ?\nA ce moment même où, une loi européenne donnant le statut de personne électronique au robot a été gelée certes, mais après de longues discussions, et je dirais même de combats, en se souvenant que son embryon est née au MIT il y a 10 ans au plus près des GAFA qui proposent de décider pour nous et de nous enseigner les bons usages de l’IA (proposition émise en sept 2017), et où cette loi de personne électronique séduit malgré tout par exemple en Pologne, alors qu’un robot a obtenu un titre de nationalité en arabie saoudite, il est urgent de se poser les questions des effets de cette technologie et de les encadrer afin qu’elle ne soit pas rejetée par une inquiétude collective. Car si notre hypothèse d’isolement s’avérait, le robot serait un moyen efficace de nous en faire prendre conscience pour que nous développions des mécanismes de reconstructions humaines.\nLa question qu’il me semble donc essentiel de poser est : cet artefact est-il seulement futile, auquel cas il n’est pas nécessaire d’y réfléchir ? Pourrait-il utilement favoriser la reconstruction d\u0026rsquo;un espace social humain endommagé, jusqu\u0026rsquo;à devenir inutile ? Dans ce cas comment s’assurer que c’est un mécanisme de reconstruction humaine qui sera mise en œuvre et non pas l’enferment dans l’illusion au début agréable de l’illusion de relation avec les artefacts robots ? Ou au contraire la méconnaissance que nous avons aujourd’hui, toutes sciences humaines confondues, des processus profonds de l\u0026rsquo;interaction est-il un risque de toxicité dans la simulation artificielle de l’interaction qui déclenche l\u0026rsquo;illusion d\u0026rsquo;un autre ?\n","date":"28 March 2019","externalUrl":null,"permalink":"/en/seminars/2019-03-28-auberge/","section":"Team seminars","summary":"","title":"La Théorie D.A.N.S.E. et ses applications en interaction personne-robot, située dans les enjeux sociétaux et éthiques inévitables","type":"seminars"},{"content":"Les textes dramatiques français du xviie siècle présentent une structure intrinsèque extrêmement riche : leur division en scènes, répliques et vers fait de ce corpus un recueil d’interactions qui présentent de vastes potentialités pour une analyse automatisée des phénomènes conversationnels.\nEn se plaçant à des échelles variées (des unités les plus étendues aux plus brèves), il s’agira ici de montrer tout le bénéfice que les études littéraires aussi bien que la linguistique peuvent tirer de l’extraction automatique des motifs récurrents que l’on observe dans ces textes, que ceux-ci concernent la gestion du cadre conversationnel (entrées et sorties des personnages) ; l’organisation des tours de parole en situation de trilogue ou de polylogue ; ou, enfin, les choix lexicaux reposant sur l’emploi d’éléments de langage fréquemment répétés.\n","date":"21 March 2019","externalUrl":null,"permalink":"/en/seminars/2019-03-21-douguet/","section":"Team seminars","summary":"","title":"Le traitement automatique de la parole théâtrale","type":"seminars"},{"content":"Tamgu signifie en coréen l’investigation, la recherche. Ce langage regroupe tous les outils de base nécessaire à l’extraction et la détection d’expressions textuelles. Tamgu permet en particulier de combiner des approches en apprentissage automatique avec des approches plus symboliques, par exemple en intégrant des lexiques généraux ou utilisateurs.\n","date":"14 February 2019","externalUrl":null,"permalink":"/en/seminars/2019-02-14-roux/","section":"Team seminars","summary":"","title":"Tamgu, un langage de programmation pour l’extraction d'information","type":"seminars"},{"content":"The neurobiological disability called dyslexia (\u0026lt; Greek dys- ‘impaired’ + lexis ‘word’) is a specific learning disability that affects only literacy skills. It has been generally assumed that congenital form of dyslexia, termed developmental dyslexia, stems from a particular problem in language acquisition affecting phonological awareness. However, the exact nature of phonological awareness has not yet been made clear.\nThis study spotlights the seemingly mysterious discrepancy in the prevalence of dyslexic populations between stress-timed English (as high as 20%) and mora-timed Japanese (as low as 1%). Stress-timed French falls between the two types of languages. On the basis of English dyslexic reading marked by an overproduction of moraic (CV) units in the absence of rhyme (VC) units, the study strengthens the mora-basic hypothesis and shows that the discrepancy is due to differences in prosodic structures between the two languages. For VC-oriented English, the readers must have rhyme awareness depicting the unit rhyme through prosodic restructuring from CV-C to C-VC. A failure to do so manifests as phonological dyslexia. For mora (CV) oriented Japanese, rhyme awareness and prosodic restructuring are irrelevant. Consequently phonological dyslexia is largely undetected.\nFrom the articulatory phonological point of view, it is suggested that onset consonants are coarticulation of the following vowels. Moras (CVs) are thus formed automatically and essentially free. In contrast, coda consonants are not coarticulation with the preceding vowels. Forming rhymes (VCs) requires a cognitive temporal-spatial decision load, which a dyslexic mind is unable to bear. Mora inclination is explained accordingly.\nThe study deepens the above view and come to claim that mora-forming coarticulation is easy because it is a synchronized articulatory behavior, akin to a synchronized human locomotive behavior. This view conforms to a human neurobiological restriction inclined toward synchronized behavior, which is claimed to be acquired in the process of human evolution.\nDevelopmental dyslexia has a serious impact on children’s learning and forms a quite interdisciplinary field of study ranging from clinics, to brain science, to information processing, to linguistics, and to pedagogy, which offers both technical and conceptual research potentials.\nKeywords: mora inclination, developmental dyslexia, phonological awareness, rhyme awareness, coarticulation, synchronized articulatory behavior\n","date":"5 February 2019","externalUrl":null,"permalink":"/en/seminars/2019-02-05-katada/","section":"Team seminars","summary":"","title":"Explaining Mora Inclination in Phonological Dyslexia","type":"seminars"},{"content":"Computational methods offer exciting new possibilities for recording and processing low resource languages. If we are to extend these methods to encompass all languages we run into a problem: most languages are unwritten. Existing attempts encounter the \u0026ldquo;transcription bottleneck\u0026rdquo;, the fact that it is extremely onerous to transcribe audio in a language that has no established orthography. These talks describe two new ways to address the transcription bottleneck, by rethinking the tasks and the end products in the light of the capacities and motives of speakers, and the requirements of the speech and language processing pipeline. These talks will describe work in progress, based in a remote indigenous community in Australia.\n","date":"14 January 2019","externalUrl":null,"permalink":"/en/seminars/2019-01-14-bird/","section":"Team seminars","summary":"","title":"Scalable Methods for Working with Unwritten Languages 2: Talking about Places and Processes","type":"seminars"},{"content":"","date":"10 January 2019","externalUrl":null,"permalink":"/en/seminars/2019-01-10-schwartz/","section":"Team seminars","summary":"","title":"Intersecting machine learning and linguistic fieldwork: Computational models for St. Lawrence Island Yupik","type":"seminars"},{"content":"Computational methods offer exciting new possibilities for recording and processing low resource languages. If we are to extend these methods to encompass all languages we run into a problem: most languages are unwritten. Existing attempts encounter the \u0026ldquo;transcription bottleneck\u0026rdquo;, the fact that it is extremely onerous to transcribe audio in a language that has no established orthography. These talks describe two new ways to address the transcription bottleneck, by rethinking the tasks and the end products in the light of the capacities and motives of speakers, and the requirements of the speech and language processing pipeline. These talks will describe work in progress, based in a remote indigenous community in Australia.\n","date":"9 January 2019","externalUrl":null,"permalink":"/en/seminars/2019-01-09-bird/","section":"Team seminars","summary":"","title":"Scalable Methods for Working with Unwritten Languages 1: Interactive Respeaking","type":"seminars"},{"content":"Wed 9th January at 2pm - room 306 batiment IMAG\nScalable Methods for Working with Unwritten Languages 1: Interactive Respeaking\nMon 14th January at 2pm - room 306 batiment IMAG\nScalable Methods for Working with Unwritten Languages 2: Talking about Places and Processes\nSteven Bird, Charles Darwin University\nComputational methods offer exciting new possibilities for recording and processing low resource languages. If we are to extend these methods to encompass all languages we run into a problem: most languages are unwritten. Existing attempts encounter the \u0026ldquo;transcription bottleneck\u0026rdquo;, the fact that it is extremely onerous to transcribe audio in a language that has no established orthography. These talks describe two new ways to address the transcription bottleneck, by rethinking the tasks and the end products in the light of the capacities and motives of speakers, and the requirements of the speech and language processing pipeline. These talks will describe work in progress, based in a remote indigenous community in Australia.\n","date":"21 December 2018","externalUrl":null,"permalink":"/en/news/2018/2018-12-21-steven-bird-seminars-2019/","section":"News","summary":"","title":"2 seminars from Steven Bird in january 2019","type":"news"},{"content":"In the last few years, neural networks have quickly gained a dominant position in computational linguistics. In application domains where supervised data is abundant, such as Machine Translation between some of the major world languages, the superior learning capabilities of neural networks have produced models with better performance than the previously available techniques. In such abundant data conditions, these models can be trained from raw data, in an end-to-end fashion, without much injection of external knowledge.\nHowever, in less favorable data conditions, prior knowledge continues to play an important role: it allows the neural components to be guided, not only by direct data observations, but also by hypotheses and principles that come from an understanding of the problem at hand.\nIn my talk, I will try to provide some intuitions about the role of prior knowledge in deep learning for NLP and provide some examples from my own experience with applications such as Language Modelling, NLG, and Semantic Parsing.\n","date":"6 December 2018","externalUrl":null,"permalink":"/en/seminars/2018-12-06-dymetman/","section":"Team seminars","summary":"","title":"Prior knowledge and deep learning: some principles and applications to NLP","type":"seminars"},{"content":"Les techniques de Narration Interactive se sont développées depuis les années 2000 principalement dans le but de fournir un contenu narratif plus évolué aux médias interactifs, pour des applications plus ludiques qu’éducatives. Avec le développement de représentations des connaissances plus sophistiquées et d’approches plus cognitives de la narration, il existe de nouvelles opportunités pour utiliser des techniques narratives dans le cadre de la simulation et de la formation. Cette tendance rejoint au niveau narratif le développement du domaine des « Jeux Sérieux ». Nous présentons plusieurs exemples d’utilisation de techniques narratives dans des applications non ludiques, basées sur une utilisation de techniques de planification par opérateurs ou par tâches.\nDans le domaine de la formation ou de l’éducation des patients, la conversion de modèles en connaissances en fragments narratifs scénarisés peut être utilisée pour créer une diversité de situations résultant de l’interaction entre des connaissances génériques et des données personnelles. Nous présenterons également une approche cognitive de la narration qui vise à contrôler le phénomène de compréhension narratif, et qui a pu être utilisée pour explorer la compréhension causale chez l’enfant.\n","date":"4 December 2018","externalUrl":null,"permalink":"/en/seminars/2018-12-04-cavazza/","section":"Team seminars","summary":"","title":"Nouvelles applications des techniques de narration interactive","type":"seminars"},{"content":"I report work in progress on learning simplified interpreted languages by means of recurrent models. The data is constructed to reflect core properties of natural language as modeled in formal syntax and semantics: recursive syntactic structure and compositionality. Preliminary results suggest that LSTM networks do generalise to compositional interpretation, albeit only in the most favorable learning setting, with a well-paced curriculum, extensive training data, and left-to-right (but not right-to-left) composition.\nBio : While I have experience in different subfields of language science such as field linguistics, language typology, and formal semantics, my current work mainly focuses on computational semantic representations for natural language, including word, phrase, and sentence embeddings. My work has included proposing new models such as the Practical Lexical Function model for syntax-driven vector compositionality or the Boolean Distributional Semantic Model for entailment detection, as well as analyzing existing models and evaluating them on new tasks.\nI hold an undergraduate degree in Linguistics from Moscow State University and a PhD from the University of California, Los Angeles. After finishing my thesis, I was a postdoc at Marco Baroni\u0026rsquo;s COMPOSES group at the University of Trento. Since 2016, I am a researcher (CR CNRS) at the Lorraine laboratory of computer science and its applications (Loria).\n","date":"8 November 2018","externalUrl":null,"permalink":"/en/seminars/2018-11-08-paperno/","section":"Team seminars","summary":"","title":"Limitations in learning an interpreted language with recurrent models","type":"seminars"},{"content":"Didier Schwab, Benjamin Lecouteux et Laurent Besacier présentent le 5 Octobre 2018 deux démonstrations aux R\u0026amp;T Days de l\u0026rsquo;Institut Cognition.\nGazePlay : jeux par le regard pour les enfants en situation de polyhandicap Lig-Aikuma : une application mobile pour la collecte de parole sur le terrain ","date":"5 October 2018","externalUrl":null,"permalink":"/en/news/2018/2018-10-05-rt-days-2018/","section":"News","summary":"","title":"LIG présent au R\u0026T Days de l'Institut Cognition !","type":"news"},{"content":"Our usage of language is not solely reliant on cognition but is arguably determined by myriad external factors leading to a global variability of linguistic patterns. This issue, which lies at the core of sociolinguistics and is backed by many small-scale studies on face-to-face communication, is addressed here by constructing a dataset combining the largest French Twitter corpus to date with detailed socioeconomic maps obtained from national census in France. We show how key linguistic variables measured in individual Twitter streams depend on factors like socioeconomic status, location, time, and the social network of individuals. We found that (i) people of higher socioeconomic status, active to a greater degree during the daytime, use a more standard language; (ii) the southern part of the country is more prone to use more standard language than the northern one, while locally the used variety or dialect is determined by the spatial distribution of socioeconomic status; and (iii) individuals connected in the social network are closer linguistically than disconnected ones, even after the effects of status homophily have been removed. Our results inform sociolinguistic theory and may inspire novel learning methods for the inference of socioeconomic status of people from the way they tweet.\n","date":"20 September 2018","externalUrl":null,"permalink":"/en/seminars/2018-09-20-abitbol/","section":"Team seminars","summary":"","title":"Socioeconomic Dependencies of Linguistic Patterns in Twitter: A Multivariate Analysis","type":"seminars"},{"content":"Séminaire consacré à la conférence Coling 2018.\n","date":"13 September 2018","externalUrl":null,"permalink":"/en/seminars/2018-09-13-boitet/","section":"Team seminars","summary":"","title":"Retour sur la conférence Coling 2018","type":"seminars"},{"content":" # Welcome # This is the web page for the Computational Approaches for Documenting and Analyzing Oral Languages Special Session at ICPhS 2019, the International Congress of the Phonetic Sciences, 5-9 August 2019, Melbourne, Australia.\nSummary # The special session Computational Approaches for Documenting and Analyzing Oral Languages welcomes submissions presenting innovative speech data collection methods and/or assistance for linguists and communities of speakers: methods and tools that facilitate collection, transcription and translation of primary language data. Oral languages is understood here as referring to spoken vernacular languages which depend on oral transmission, including endangered languages and (typically low-prestige) regional varieties of major languages.\nThe special session intends to provide up-to-date information to an audience of phoneticians about developments in machine learning that make it increasingly feasible to automate segmentation, alignment or labelling of audio recordings, even in less-documented languages. A methodological goal is to help establish the field of Computational Language Documentation and contribute to its close association with the phonetic sciences. Computational Language Documentation needs to build on the insights gained through phonetic research; conversely, research in phonetics stands to gain much from the availability of abundant and reliable data on a wider range of languages.\nOur special session is mentioned on the ICPhS website here. You can find a poster of this session here.\nMain goals # The special session aims to bring together phoneticians, computer scientists and developers interested in the following goals:\nRethinking documentary processes: recording, transcription and annotation; Responding to the urgent need to document endangered languages and varieties; Elaborating an agenda and establishing a roadmap for computational language documentation; Ensuring that the requirements of phonetics research are duly taken into consideration in the agenda of Computational Language Documentation; Attracting computer scientists to ICPhS and engaging them in discussions with phoneticians (and linguists generally). Main topics # This special session will focus on documenting and analyzing oral languages, including topics such as the following:\nlarge-scale phonetics of oral languages, automatic phonetic transcription (and phonemic transcription), mobile platforms for speech data collection, creating multilingual collections of text, speech and images, machine learning over these collections, open source tools for computational language documentation, position papers on computational language documentation. Session format # Special sessions at ICPhS will normally be 1.5 hours. For our accepted special session, we chose the \u0026ldquo;workshop\u0026rdquo; type with a more open format suitable for discussion of methods/tools. The exact format is still to be determined. More details will be provided on the format later.\nHow does the submission process work? # Papers will be submited directly to the conference by December 4th and will then be evaluated according to the standard ICPhS review process [see here]. Accepted papers will be allocated either to this special session or a general session. When submitting you can specify if you want to be considered for this special session.\nOrganizers # Laurent Besacier – LIG UGA (France)\nAlexis Michaud – LACITO CNRS (France)\nMartine Adda-Decker – LPP CNRS (France)\nGilles Adda – LIMSI CNRS (France)\nSteven Bird – CDU (Australia)\nGraham Neubig – CMU (USA)\nFrançois Pellegrino – DDL CNRS (France)\nSakriani Sakti – NAIST (Japan)\nMark Van de Velde – LLACAN CNRS (France)\nEndorsement # This special session is endorsed by SIGUL (Joint ELRA and ISCA Special Interest Group on Under-resourced Languages)\n","date":"20 July 2018","externalUrl":null,"permalink":"/en/news/2018/2018-07-20-icphs-2019-special-session/","section":"News","summary":"","title":"ICPhS 2019 Special Session","type":"news"},{"content":"","date":"31 May 2018","externalUrl":null,"permalink":"/en/seminars/2018-05-31-pouliquen/","section":"Team seminars","summary":"","title":"From SMT to NMT at WIPO","type":"seminars"},{"content":"Steven Bird is researching new methods for documenting and revitalising the thousands of small languages still spoken in the world today. His career began with a BSc and MSc in computer science at Melbourne University, followed by a PhD in computational linguistics from Edinburgh University, completed in 1990. Since then he has worked at the Universities of Edinburgh, Pennsylvania, Melbourne, and Berkeley, and conducted fieldwork in Australia, West Africa, Melanesia, Amazonia, and Central Asia. He is co-author of a popular textbook in computational linguistics, and recently developed a new computer science curriculum for secondary students which has been adopted in Australian schools. The Aikuma app developed with his students took out the grand prize in the Open Source Software World Challenge.\nSteven Bird’s page\n","date":"4 April 2018","externalUrl":null,"permalink":"/en/seminars/2018-04-04-bird/","section":"Team seminars","summary":"","title":"Sparse Transcription: Rethinking the Processing of Unwritten Languages","type":"seminars"},{"content":"Dans ce séminaire, je présenterai deux projets scientifiques collectifs [1,2] qui m\u0026rsquo;ont occupé pendant l\u0026rsquo;année 2017. Qu\u0026rsquo;ont-ils en commun ? La découverte d\u0026rsquo;unités linguistiques à partir de la parole brute sans aucune autre supervision. Ou presque…\nIn this seminar, I will present two collective scientific projects [1,2] that occupied me during the year 2017. What do they have in common? Discovering linguistic units from raw speech without any other supervision. Or almost…\nhttps://arxiv.org/pdf/1712.04313.pdf https://arxiv.org/pdf/1802.05092.pdf ","date":"29 March 2018","externalUrl":null,"permalink":"/en/seminars/2018-03-29-besacier/","section":"Team seminars","summary":"","title":"Le défi de découvrir des unités linguistiques à partir de la parole brute / The challenge of discovering linguistic units from raw speech","type":"seminars"},{"content":"Dans ce séminaire je vais parler des principaux domaines de recherches sur lesquels j’ai travaillé : compréhension automatique de la parole et resolution de chaînes de coréférences. Je décrirai les systèmes informatiques, surtout à base d’apprentissage artificiel, mis en place pour modéliser ces problèmes.\nCes systèmes s’appuient sur des modèles qui vont des automates probabilistes à états finis (FSA/FST) aux réseaux neuronaux, en passant par les champs conditionnels aléatoires (CRF), et détiennent l’état-de-l’art sur certaines tâches.\n","date":"22 March 2018","externalUrl":null,"permalink":"/en/seminars/2018-03-22-dinarelli/","section":"Team seminars","summary":"","title":"Compréhension automatique de la parole et résolution de chaînes de coréférences","type":"seminars"},{"content":"L’extraction de lexiques bilingues à partir de corpus a initialement été réalisée en s’appuyant sur des textes en correspondance de traduction (c’est-à-dire des corpus parallèles). Cependant, et en dépit des bons résultats obtenus, ces corpus demeurent des ressources rares, notamment pour les domaines spécialisés et pour des couples de langues ne faisant pas intervenir l’anglais. Dans ce contexte, les recherches en extraction de lexiques bilingues se sont penchées sur d’autres corpus composés de textes partageant différentes caractéristiques telles que le domaine, le genre, la période\u0026hellip; sans être en correspondance de traduction (c’est-à-dire des corpus comparables).\nL’extraction de lexiques bilingues à partir de corpus comparables spécialisés est fortement contrainte par la quantité de données mobilisables. Pour contourner cet obstacle, une solution serait d’associer des ressources externes au corpus spécialisés. Cette solution, quoi que intuitive, va à l\u0026rsquo;encontre du courant dominant puisque de nombreuses études soutiennent l\u0026rsquo;idée que l\u0026rsquo;ajout de documents hors-domaine à un corpus spécialisé diminue la qualité des lexiques extraits. Dans cet exposé nous montrerons comment des corpus de langue générale peuvent venir compléter des corpus de langue de spécialité. Nous présenterons différentes manières d’associer ces données entre elles en exploitant des représentations distributionnelles basées sur modèles vectoriels et neuronaux.\n","date":"20 March 2018","externalUrl":null,"permalink":"/en/seminars/2018-03-20-morin/","section":"Team seminars","summary":"","title":"Extraction de lexiques bilingues à partir de corpus comparables spécialisés : la langue générale au secours de la langue de spécialité","type":"seminars"},{"content":"On désigne par \u0026ldquo;motifs\u0026rdquo; des constructions récurrentes susceptibles de jouer un rôle dans l\u0026rsquo;organisation textuelle et la structuration du discours. Les motifs, en tant que constructions préfabriquées, sont par ailleurs caractéristiques de genres textuels très codifiés. L\u0026rsquo;identification de ces constructions peut s\u0026rsquo;avérer utile dans différents types d\u0026rsquo;application en TAL : classification de documents, traduction automatique, aide à la rédaction, recherche de termes, outils pour la linguistique de corpus\u0026hellip; Après avoir précisé la notion au plan linguistique, nous passerons en revue différentes méthodes dédiées à l\u0026rsquo;identification automatique de motifs : segments répétés ou ngrams, motifs d\u0026rsquo;itemsets, arbres lexico-syntaxiques récurrents. Nous détaillerons les pistes recherches actuelles concernant l\u0026rsquo;utilisation de la syntaxe (analyses en dépendances) pour la découverte et la description de certaines classes de motifs.\n","date":"8 March 2018","externalUrl":null,"permalink":"/en/seminars/2018-03-08-kraif/","section":"Team seminars","summary":"","title":"Analyse en dépendances pour l'extraction automatique de motifs récurrents","type":"seminars"},{"content":"Dans les procédures judiciaires, des enregistrements de voix sont de plus en plus fréquemment présentés comme élément de preuve. En général, il est fait appel à un expert scientifique pour établir si l’extrait de voix en question a été prononcé par un suspect donné (prosecution hypothesis) ou non (defence hypothesis). Ce prosessus est connu sous le nom de “Forensic Voice Comparison (FVC)” (comparaison de voix dans le cadre judiciaire). Depuis l’émergence du modèle DNA typing, l’approche Bayesienne est devenue le nouveau “golden standard” en sciences criminalistiques. Dans cette approche, l’expert exprime le résultat de son analyse sous la forme d’un rapport de vraisemblance (LR). Ce rapport ne favorise pas seulement une des hypothèses (“prosecution” ou “defence”) mais il fournit également le poids de cette décision. Bien que le LR soit théoriquement suffisant pour synthétiser le résultat, il est dans la pratique assujetti à certaines limitations en raison de son processus d’estimation. Cela est particulièrement vrai lorsque des systèmes de reconnaissance automatique du locuteur (ASpR) sont utilisés. Ces systèmes produisent un score dans toutes les situations sans prendre en compte les conditions spécifiques au cas étudié. Plusieurs facteurs sont presque toujours ignorés par le processus d’estimation tels que la qualité et la quantité d’information dans les deux enregistrements vocaux, la cohérence de l’information entre les deux enregistrements,\nleurs contenus phonétiques ou encore les caractéristiques intrinsèques des locuteurs. Tous ces facteurs mettent en question la notion de fiabilité de la comparaison de voix dans le cadre judiciaire. Dans cette thèse, nous voulons adresser cette problématique dans le cadre des systèmes automatiques (ASpR) sur deux points principaux.\nLe premier consiste à établir une échelle hiérarchique des catégories phonétiques des sons de parole selon la quantité d’information spécifique au locuteur qu’ils contiennent. Cette étude montre l’importance du contenu phonétique: Elle met en évidence des différences intéressantes entre les phonèmes et la forte influence de la variabilité intra-locuteurs. Ces résultats ont été confirmés par une étude complémentaire sur les voyelles orales basée sur les paramètres formantiques, indépendamment de tout système de reconnaissance du locuteur.\nLe deuxième point consiste à mettre en oeuvre une approche afin de prédire la fiabilité du LR à partir des deux enregistrements d’une comparaison de voix sans recours à un ASpR. À cette fin, nous avons défini une mesure d’homogénéité (NHM) capable d’estimer la quantité d’information et l’homogénéité de cette information entre les deux enregistrements considérés. Notre hypothèse ainsi définie est que l’homogénéité soit directement corrélée avec le degré de fiabilité du LR. Les résultats obtenus ont confirmé cette hypothèse avec une mesure NHM fortement corrélée à la mesure de fiabilité du LR. Nos travaux ont également mis en évidence des différences significatives du comportement de NHM entre les comparaisons cibles et les comparaisons imposteurs.\nNos travaux ont montré que l’approche “force brute” (reposant sur un grand nombre de comparaisons) ne suffit pas à assurer une bonne évaluation de la fiabilité en FVC. En effet, certains facteurs de variabilité peuvent induire des comportements locaux des systèmes, liés à des situations particulières. Pour une meilleure compréhension de l’approche FVC et/ou d’un système ASpR, il est nécessaire d’explorer le comportement du système à une échelle aussi détaillée que possible (le diable se cache dans les détails).\n","date":"8 February 2018","externalUrl":null,"permalink":"/en/seminars/2018-02-08-ajili/","section":"Team seminars","summary":"","title":"Fiabilité de la comparaison de voix dans le cadre judiciaire / Reliability of voice comparison for forensic applications","type":"seminars"},{"content":"Thesis entitled « Modèle profond pour le contrôle vocal adaptatif d’un habitat Intelligent » (\u0026ldquo;Deep model for adaptive voice control of an Intelligent habitat\u0026rdquo;) at 14:15 in the IMAG amphitheatre.\n","date":"1 December 2017","externalUrl":null,"permalink":"/en/news/2017/2017-12-01-thesis-defense-alexis-brenon/","section":"News","summary":"","title":"Thesis defense by Alexis Brenon on December 14, 2017","type":"news"},{"content":"De plus en plus de disciplines ou de recherches en sciences humaines et sociales, lettres et langues s’intéressent à la dimension spatiale des données ou des sources: en histoire pour la représentation de données géo historiques nécessaires à la compréhension de l’évolution des territoires ou des phénomènes les impactants; en littérature pour la cartographie des lieux dans les romans, des récits de vie des auteurs ; en linguistique pour appréhender la diffusion spatiale des langues ou des dialectes; en géographie pour la reconstitution de trajectoires et de déplacements d’individus à partir de récits ou pour la valorisation de documents cartographiques anciens…. Tous ces besoins lancent de nouveaux défis à la cartographie et à la géo visualisation qui doivent traiter des données spatialisées semi-structurées, multidimensionnelles, multi-formes et définies par une diversité d’échelles d’observation tant géographique que temporelle et selon des niveaux de qualité variables.\nL’objectif de la communication est de présenter certaines problématiques cartographiques et de géo visualisation soulevées par le traitement et la représentation spatialisées des données issues du domaine des humanités numériques en s’appuyant sur des projets de recherche menés au sein de l’équipe Steamer\n","date":"23 November 2017","externalUrl":null,"permalink":"/en/seminars/2017-11-23-davoine/","section":"Team seminars","summary":"","title":"Cartographie et géovisualisation pour la représentation et l'analyse de données spatialisées pour les humanités numériques","type":"seminars"},{"content":"Le thème sera abordé à partir des travaux d\u0026rsquo;analyse des publications de la communauté TAL réalisés autour du corpus NLP4NLP\nqui couvre 50 ans de publications des principales conférences et revues dans le domaine de l\u0026rsquo;analyse du texte et de la parole\net de corpus biomédicaux (projet MIROR). Les apports du TAL adressés ici concerneront l\u0026rsquo;analyse des tendances et des réseaux ainsi que la détection de plagiat ou de \u0026ldquo;spin\u0026rdquo; (embellissement) dans les publications scientifiques.\n","date":"26 October 2017","externalUrl":null,"permalink":"/en/seminars/2017-10-26-paroubek/","section":"Team seminars","summary":"","title":"Le traitement automatique des langues pour l'analyse des publications scientifiques","type":"seminars"},{"content":"Séminaire consacré au MT Summit 2017.\n","date":"5 October 2017","externalUrl":null,"permalink":"/en/seminars/2017-10-05-boitet/","section":"Team seminars","summary":"","title":"Retour sur le MT Summit 2017","type":"seminars"},{"content":"Ixiade proposes a CIFRE thesis as part of a collaboration with the GETALP team of the laboratoire d\u0026rsquo;Informatique de Grenoble.\nStart of thesis: As soon as possible.\nLocation: the position is based in Grenoble.\nThe gross annual salary is 28 800€.\nDescription of the subject:\nFor more than 12 years, Ixiade\u0026rsquo;s experts have worked closely with teams of companies wishing to enhance and develop their strategic, technological and human assets. They work across the board as part of a participatory innovation process. The aim is to bring innovation to the teams through the transmission of best practices and \u0026ldquo;tailor-made\u0026rdquo; recommendations, in particular through upstream analysis of the potential meaning of a new concept or a new product or service.\nOn this last aspect, Ixiade\u0026rsquo;s expertises are based on the realization of semi-directive oral interviews, their written transcription and the detailed analysis of these transcriptions. This last step is particularly time-consuming and the objective of this thesis is the creation of tools based on the automatic processing of the natural language allowing the expert to more easily perceive the underlying opinion of the persons interviewed on the concept studied.\nThis thesis will thus consist in studying the extraction of ideas from textual documents, whether these ideas are linked to a particular field (for example in association with existing practices) or, for the most interesting part, to the expression of opinions that are more general and emotionally rich because they are linked to the profile of the interviewee.\nThe extraction can then be done in a:\\\nendogenous, based on information from a segmentation of the interview corpus.\\ The extraction can then be done in a:\nexogenous, from external resources such as, for example, the lexical resource JeuDemots (large lexical graph on the French language built from games/associations of ideas for about ten years) or from resources learned on a general corpus and/or from a related field. The candidate will be encouraged to publish his/her progress at major conferences in the field in TALN (automatic natural language processing) (ACL, Interspeech).\nThe candidate may benefit from a thesis work that will start jointly on the same field of application but rather focus on the linguistic analysis of available corpora.\nProfile of the candidate sought :\nYou have or have completed a Master 2 Research in Computer Science or in TALN (automatic natural language processing) and wish to prepare a CIFRE doctorate in a company in liaison with a research laboratory. You are passionate about information and communication technologies. You have training and experience in the study and/or development of automatic natural language processing. Knowledge in automatic learning would be a plus. You are interested in opinion analysis in oral productions and wish to discover the working methodologies of the innovation industry.\nPlease send a CV + cover letter + letter of recommendation to Isabelle Fournié (isabelle.fournie@ixiade.com), Didier Schwab (didier.schwab@imag.fr) and Jérôme Goulian (jerome.goulian@imag.fr).\nIxiade : http://www.ixiade.com/fr Getalp : http://lig-getalp-new.imag.fr/\n","date":"12 July 2017","externalUrl":null,"permalink":"/en/news/2017/2017-07-12-cifre-thesis-ixiade/","section":"News","summary":"","title":"Proposal for a CIFRE thesis with the company Ixiade","type":"news"},{"content":"The GETALP team is recruiting a PhD student on a thesis topic funded by the ANR : Custom voice control system in context\n3-year contract, beginning September 2017.\n","date":"26 June 2017","externalUrl":null,"permalink":"/en/news/2017/2017-06-26-thesis-offer-anr/","section":"News","summary":"","title":"Thesis offer financed by the ANR","type":"news"},{"content":"","date":"30 May 2017","externalUrl":null,"permalink":"/en/seminars/2017-05-30-duran/","section":"Team seminars","summary":"","title":"Le temps non marqué et suffixation à quatre niveaux en quechua","type":"seminars"},{"content":"In the online adversiment market there are two main actors: the publishers that offer a space for advertisement in their websites and the announcers who compite in an auction to show their advertisements in the available spaces. When a user accesses an internet site an auction starts for each ad space, the profile of the user is given to the announcers and they offer a bid to show an ad to that user. The publisher fixes a reserve price, the minimum value they accept to sell the space.\nIn this talk I will introduce a general setting for this ad market and I will present an engine to optimize the publisher revenue from second-price auctions, which are widely used to sell on-line ad spaces in a mechanism called real-time bidding. The engine is fed with a stream of auctions in a time-varying environment (non-stationary bid distributions, new items to sell, etc.) and it predicts in real time the optimal reserve price for each auction. This problem is crucial for web publishers, because setting an appropriate reserve price on each auction can increase significantly their revenue.\nI consider here a realistic setting where the only available information consists of a user identifier and an ad placement identifier. Once the auction has taken place, we can observe censored outcomes : if the auction has been won (i.e the reserve price is smaller than the first bid), we observe the first bid and the closing price of the auction, otherwise we do not observe any bid value.\nThe proposed approach combines two key components: (i) a non-parametric regression model of auction revenue based on dynamic, time-weighted matrix factorization which implicitly builds adaptive users’ and placements’ profiles; (ii) a non-parametric model to estimate the revenue under censorship based on an on-line extension of the Aalen’s Additive Model.\n","date":"18 May 2017","externalUrl":null,"permalink":"/en/seminars/2017-05-18-chahuara/","section":"Team seminars","summary":"","title":"Online Mining of Web Publisher RTB Auctions for Revenue Optimization","type":"seminars"},{"content":"Online Mining of Web Publisher RTB Auctions for Revenue Optimization\nIn the online adversiment market there are two main actors: the publishers that offer a space for advertisement in their websites and the announcers who compite in an auction to show their advertisements in the available spaces. When a user accesses an internet site an auction starts for each ad space, the profile of the user is given to the announcers and they offer a bid to show an ad to that user. The publisher fixes a reserve price, the minimum value they accept to sell the space.\nIn this talk I will introduce a general setting for this ad market and I will present an engine to optimize the publisher revenue from second-price auctions, which are widely used to sell on-line ad spaces in a mechanism called real-time bidding. The engine is fed with a stream of auctions in a time-varying environment (non-stationary bid distributions, new items to sell, etc.) and it predicts in real time the optimal reserve price for each auction. This problem is crucial for web publishers, because setting an appropriate reserve price on each auction can increase significantly their revenue.\nI consider here a realistic setting where the only available information consists of a user identifier and an ad placement identifier. Once the auction has taken place, we can observe censored outcomes : if the auction has been won (i.e the reserve price is smaller than the first bid), we observe the first bid and the closing price of the auction, otherwise we do not observe any bid value.\nThe proposed approach combines two key components: (i) a non-parametric regression model of auction revenue based on dynamic, time-weighted matrix factorization which implicitly builds adaptive users\u0026rsquo; and placements\u0026rsquo; profiles; (ii) a non-parametric model to estimate the revenue under censorship based on an on-line extension of the Aalen\u0026rsquo;s Additive Model.\n","date":"10 May 2017","externalUrl":null,"permalink":"/en/news/2017/2017-05-10-seminar-pedro-chahuara/","section":"News","summary":"","title":"Team seminar Pedro Chahuara on Thursday May 18 at 2pm","type":"news"},{"content":"Le rapprochement des approches cognitives et des approches sociales est souvent présenté comme un objectif souhaitable pour mieux comprendre le processus d\u0026rsquo;acquisition du langage (Hulstijn et al., 2014). Cependant, la question reste de savoir comment traduire ce programme dans la réalité de la pratique de la recherche.\nBien que les approches cognitives et sociales soient fondées sur des traditions différentes, la tentative de combiner les deux points de vue dans la recherche sur l\u0026rsquo;acquisition du langage peut bénéficier d\u0026rsquo;entreprises similaires dans d\u0026rsquo;autres domaines, tels que la cognition sociale, la sociologie cognitive, la sociolinguistique cognitive, les neurosciences sociales, etc. L\u0026rsquo;examen de ces tentatives interdisciplinaires conduit à l\u0026rsquo;identification de trois façons de combiner le social et le cognitif: l\u0026rsquo;approche sociale de la cognition, l\u0026rsquo;approche cognitive du social et l’approche dite individualisme complexe (Kaufmann et Clément, 2011; Chevrot, Drager \u0026amp; Foulkes, en préparation, Dupuy, 2004).\nParmi ces options, seule la dernière ne favorise ni le niveau social et collectif, ni le niveau cognitif et individuel (Dupuy, 2004). Au contraire, elle met l\u0026rsquo;accent sur l\u0026rsquo;interaction et la causalité bidirectionnelle entre ces eux. Dans cette perspective, des individus ayant des caractéristiques sociales et cognitives spécifiques interagissent les uns avec les autres dans le cadre de contraintes sociales et cognitives générales. Les caractéristiques des individus peuvent évoluer en raison des interactions entre eux et ces changements peuvent à leur tour modifier les contraintes générales (Hruschka et al. 2009). Dans ce cadre, l\u0026rsquo;acquisition du langage et son usage peuvent être considérés comme les résultats d\u0026rsquo;influences réciproques diffusant dans un réseau de relations.\nNous présenterons des projets susceptibles de mettre en œuvre ce cadre, notamment le projet DyLNet - Language Dynamics, Linguistic Learning, and Sociability at Preschool: Benefits of Wireless Proximity Sensors in Collecting Big Data (Nardy, 2017).\nReferences\nChevrot, J.P., Drager, K. \u0026amp; Foulkes, P. (en préparation). Sociolinguistic Variation and Cognitive Science.\nDupuy, J.-P. (2004). Vers l’unité des sciences sociales autour de l’individualisme méthodologique complexe. Revue du MAUSS, 24(2), 310-328.\nHruschka, D. J., Christiansen, M. H., Blythe, R. A., Croft, W., Heggarty, P., Mufwene, S. S., Pierrehumbert, Janet B., Poplack, S. (2009). Building social cognitive models of language change. Trends in Cognitive Sciences, 13(11), 464–469.\nHulstijn, J. H., Young, R. F., Ortega, L., Bigelow, M., DeKeyser, R., Ellis, N. C., Lantolf, J. P., Mackey, A., Talmy, S. (2014). Bridging the Gap. Studies in Second Language Acquisition, 36(03), 361–421.\nKaufmann, L., \u0026amp; Clément, F. (2011). L’esprit des sociétés. Bilan et perspectives en sociologie cognitive. In L. Kaufmann \u0026amp; F. Clément, La sociologie cognitive, Ophrys (pp. 7–40).\nNardy (2017). DyLNet Project - Language Dynamics, Linguistic Learning, and Sociability at Preschool: Benefits of Wireless Proximity Sensors in Collecting Big Data https://hal-univ-orleans.archives-ouvertes.fr/hal-01396652\n","date":"2 March 2017","externalUrl":null,"permalink":"/en/seminars/2017-03-02-chevrot/","section":"Team seminars","summary":"","title":"Acquisition du langage et usages sociolinguistiques : le social, le cognitif et le réseau","type":"seminars"},{"content":"Language acquisition and sociolinguistic uses: social, cognitive and network\n","date":"1 March 2017","externalUrl":null,"permalink":"/en/news/2017/2017-03-01-seminar-jean-pierre-chevrot/","section":"News","summary":"","title":"Jean-Pierre Chevrot team seminar (Thursday 2 March)","type":"news"},{"content":"The VOCADOM project led by GETALP was selected by the ANR during the generic call in 2016.\nDuration of the project: 48 months - Scientific Manager: Michel Vacher\nPartners : CNRS/LIG, Inria Nancy Grand-Est, GRePS, THEORIS\nProject website : VOCADOM\n","date":"17 January 2017","externalUrl":null,"permalink":"/en/news/2017/2017-01-17-vocadom-launch/","section":"News","summary":"","title":"18-19 January 2017 : Meeting to launch the ANR VOCADOM project","type":"news"},{"content":"Tout le monde a déjà rencontré le problème suivant : on cherche un mot (ou le nom d’une personne) que l’on connaît, sans être en mesure d’y accéder à temps. Les travaux des psychologues ont montré que les personnes se trouvant dans cet état cognitif savent énormément de choses concernant le mot recherché (sens, nombre de syllabes, origine, etc.), et que les mots avec lequel ils le confondent lui ressemblent étrangement (lettre ou son initial, catégorie syntaxique, champ sémantique, etc.).\nMon objectif (à long terme) est de réaliser un programme tirant bénéfice de cet état de faits pour assister un locuteur ou rédacteur à (re)trouver le mot qu’il a sur le bout de la langue. À cette fin, je prévois d’ajouter à un dictionnaire électronique existant un index d’association (collocations rencontrées dans un grand corpus). Autrement dit, je propose de construire un dictionnaire analogue à celui des êtres humains, qui, outre les informations conventionnelles (définition, forme écrite, informations grammaticales) contiendrait des liens (associations), permettant de naviguer entre les idées (concepts) et leurs expressions (mots). Un tel dictionnaire permettrait donc l’accès à l’information recherchée soit par la forme (lexicale : analyse), soit par le sens (concepts : production), soit par les deux.\nL’objectif de cet exposé est de montrer comment construire une telle ressource, comment s\u0026rsquo;en servir, quelles sont les difficultés de construction et quelles sont les possibilités qu\u0026rsquo;offre un telle ressource.\n","date":"12 January 2017","externalUrl":null,"permalink":"/en/seminars/2017-01-12-zock/","section":"Team seminars","summary":"","title":"Si tous les chemins mènent à Rome, ils ne se valent pas tous. Le problème d’accès lexical en production","type":"seminars"},{"content":"In this talk, I will introduce my research activities in the field of medical information retrieval, and in particular its evaluation.\nThe use of the Web as source of health-related information is a wide-spread phenomena, and laypeople often have difficulties finding relevant documents. The goal of the CLEF eHealth evaluation challenge is to provide researchers with datasets to improve consumer health search. I will firstly introduce the task and the datasets built. Then I will describe some experiments and results obtained on this dataset.\n","date":"1 December 2016","externalUrl":null,"permalink":"/en/seminars/2016-12-01-goeuriot/","section":"Team seminars","summary":"","title":"Medical Information Retrieval and its evaluation: an overview of CLEF eHealth evaluation task","type":"seminars"},{"content":"Lorraine Goeuriot, lecturer in the MRIM team, will give the following presentation on December 1st at 2pm in room 306.\nTitle : Medical Information Retrieval and its evaluation: an overview of CLEF eHealth evaluation task\nSummary : In this talk, I will introduce my research activities in the field of medical information retrieval, and in particular its evaluation.\nThe use of the Web as source of health-related information is a wide-spread phenomena, and laypeople often have difficulties finding relevant documents. The goal of the CLEF eHealth evaluation challenge is to provide researchers with datasets to improve consumer health search. I will firstly introduce the task and the datasets built. Then I will describe some experiments and results obtained on this dataset.\n","date":"30 November 2016","externalUrl":null,"permalink":"/en/news/2016/2016-11-30-seminar-lorraine-goeuriot/","section":"News","summary":"","title":"Lorraine Goeuriot Seminar December 1, 2016","type":"news"},{"content":"2 team articles presented at COLING 2016 in Osaka from 10 to 16 Dec. 2016 !\nYou can access it by clicking on the links below!\nhttps://arxiv.org/pdf/1610.01291v1.pdf\nhttps://arxiv.org/pdf/1609.09382v1.pdf\n","date":"19 November 2016","externalUrl":null,"permalink":"/en/news/2016/2016-12-19-coling-2016/","section":"News","summary":"","title":"2 team articles presented at COLING 2016","type":"news"},{"content":"Les technologies de reconnaissance automatique de l’émotion ont gagné une attention croissante dans la dernière décennie tant au niveau académique qu\u0026rsquo;industriel, puisqu’elles ont trouvé de nombreuses applications dans des domaines aussi variés que la santé, l’éducation, les jeux-vidéos, la publicité, ou encore la robotique sociale. Bien que de bonnes performances soient reportées dans la littérature pour des émotions actées, la reconnaissance automatique d’émotions spontanées, comme exprimées dans la vie de tous les jours, reste encore un challenge non résolu, puisque ces émotions sont subtiles, et leur expression, comme leur signification, varient fortement selon de nombreux paramètres locuteur, comme par exemple l’âge, et le genre, mais aussi la personnalité, le rôle social, la langue, et la culture. Dans cette présentation, je décrirai les méthodologies actuelles en acquisition et annotation de données affectives, et présenterai les dernières avancées pour la reconnaissance automatique des émotions à partir du signal de parole.\n","date":"20 October 2016","externalUrl":null,"permalink":"/en/seminars/2016-10-20-ringeval/","section":"Team seminars","summary":"","title":"Vers la reconnaissance automatique d’émotions écologiques","type":"seminars"},{"content":"DBnary is a lexical resource in RDF, mainly from data available on wiktionary. Since 2012 the data are extracted and the graph above shows the evolution of the number of translations available on the English wiktionary.\n","date":"30 May 2016","externalUrl":null,"permalink":"/en/news/2016/2016-05-30-focus-dbnary/","section":"News","summary":"","title":"Focus on: DBnary","type":"news"},{"content":"Due to the team\u0026rsquo;s move, all our numbers have changed. Feel free to update your info by going to the directory.\n","date":"30 May 2016","externalUrl":null,"permalink":"/en/news/2016/2016-05-30-new-contact-information/","section":"News","summary":"","title":"New contact information for GETALP","type":"news"},{"content":"A new website is now available for the GETALP team. It happens at the same time that the team moves to the new IMAG building on campus\u0026hellip; So it moves!\nP.S.: The photo is by Didier Schwab\n","date":"30 May 2016","externalUrl":null,"permalink":"/en/news/2016/2016-05-30-new-website-2016/","section":"News","summary":"","title":"New website for the GETALP team","type":"news"},{"content":" In short # GETALP (Study Group for Machine Translation and Automated Processing of Languages and Speech) was created in 2007, at the same time as the LIG laboratory.\nResulting from the virtuous union of researchers in natural language and speech processing, GETALP is a multidisciplinary team (computer scientists, linguists, phoneticians, translators, signal processing specialists, …) whose objective is to address all theoretical, methodological and practical aspects of multilingual communication and multilingual (written or spoken) information processing.\nGETALP’s methodology consists of a continuous interaction between data collection, fundamental research, development of operational systems, applications and experimental evaluations.\nHighlights # Ecological approach. One specificity of GETALP is its willingness to address the diversity of written or spoken language usage: multiplicity of languages, speakers, dialects, cultures, social contexts and applications, with a special interest in “long tails” (under-resourced languages, atypical speakers, …).\nAgnostic approach. The history of the team and the different scientific cultures it brings together allow a synergy between expert and empirical methods, large-scale analysis (big data) and analysis of phenomena requiring fine-grained annotation (beautiful data), models and induction, etc.\nMultidisciplinarity. The strength of GETALP comes partly from the complementarity of its members, enabling a wide approach: from data collection to system evaluation, from understanding communicative phenomena to industrial applications.\nHuman in the loop. To assist humans in communication and interaction situations, we include them in the automated processes (semi-supervised, interactive and collaborative approaches, analysis of system traces and errors).\nDevelopment of tools and resources. GETALP develops and freely distributes tools and resources (a Web platform for the cooperative development of multilingual lexical databases, written and spoken corpora for processing under-resourced languages, multimodal corpora collected in interaction contexts, a collaborative system for post-editing and evaluating machine translation, Web services for language, etc.).\nResearch topics # The research activities of the GETALP team are organized around two main axes, guided by values of ethics, openness and societal responsibility.\nThe first axis focuses on the scientific methods used and developed within the team:\nmethods for the analysis, collection, processing and representation of data and resources, enabling knowledge discovery and the sharing and accessibility of human knowledge; the construction, application and verification of theories and modelling approaches, contributing to fundamental knowledge in machine learning and computational linguistics; the development of NLP tools and platforms, to bring them to the economic world and support the work of linguists; the development of evaluation methods that measure not only the performance of systems but also their societal impact; the development and evaluation of methods for explaining NLP models, to make them more transparent and interpretable. The methods of the first axis contribute to the second axis, which focuses on NLP tasks and applications:\nmachine translation (written, spoken, visual); automatic speech transcription; multilingual and multimodal processing; automatic linguistic analysis (document classification, text or speech understanding, syntactic and semantic analysis…); processing of human and human–machine interactions, such as the processing of affect in interaction; language documentation. Local, national and international ecosystem # GETALP builds upon a rich local ecosystem and collaborates with other LIG teams and with other Grenoble laboratories. At the national and international level, GETALP is a recognized player in the natural language and speech processing research community (many collaborative and industrial projects, organization of JEP-TALN-RECITAL in 2012, etc.). Finally, a remarkable feature is its international network of collaborations and contacts on all continents.\nThe GETALP team comes from the GEOD and GETA teams of the CLIPS laboratory and has a long history (see also Wikipedia, in French).\nJoin us # For internship, PhD or postdoc offers, contact the team’s permanent members.\nPages from the former website # Contacts Job offers Theses in progress Where to follow NLP news ","externalUrl":null,"permalink":"/en/about/","section":"GETALP","summary":"","title":"About","type":"page"},{"content":" CAPES COFECUB (France-Brésil)\nDescription: This is a 4-year international cooperation project that aims at creating, reinforcing and continuing academic exchanges between French and Brazilian researchers in the domain of multilingual lexica and ontologies. It is co-financed by CAPES (Brazil) and by COFECUB (France), as project CAPES-COFECUB number 707/11. The project started in 2011 and is expected to be concluded by the end of 2014. This project is part of the France-Brazil international laboratory LICIA.\nPartners: LIG, UFRGS (Porto-Alegre)\nCAPES COFECUB (France-Brésil)\nDescription: This is a 4-year international cooperation project that aims at creating, reinforcing and continuing academic exchanges between French and Brazilian researchers in the domain of multilingual lexica and ontologies. It is co-financed by CAPES (Brazil) and by COFECUB (France), as project CAPES-COFECUB number 707/11. The project started in 2011 and is expected to be concluded by the end of 2014. This project is part of the France-Brazil international laboratory LICIA.\nPartners: LIG, UFRGS (Porto-Alegre)\n","externalUrl":null,"permalink":"/en/projects/cameleon/","section":"Projects","summary":"","title":"CAMELEON : Collaborative and Automatic Methods for the Multilingualisation of Lexica and Ontologies","type":"projects"},{"content":" ANR TECSAN 2011-2014 (ANR-2010- TECS-012) # Project objectives # CIRDO Industrial Research is a project aimed at developing an \u0026ldquo;Intelligent Finger and Eye Reactive Companion\u0026rdquo; that represents an enhanced and automated social tele-link product through the integration of innovative services (automatic speech recognition, situation analysis (scenes) in a complex uncontrolled environment) aimed at promoting the autonomy and management by caregivers, patients with chronic diseases or Alzheimer\u0026rsquo;s or related disorders.\nMoreover, this project will allow the validation of generic technologies, a psychological and ergonomic evaluation of the uses of the services developed (concerning utility, usability and accessibility, acceptance, ethical aspects, the economic model,\u0026hellip;) but also critical surveys of the knowledge acquired by human services professionals (SAP), which will then be transferred to the entire sector.\nDuring this project, GETALP will record speech with elderly volunteers. A brochure has been published to inform people who would like to participate in this experiment.\nGETALP Participants # Michel Vacher (responsable), Solange Rossato, Francois Portet, Benjamin Lecouteux\nPartners # **Academic partners:**University Lyon1/LIRIS SAARA team (project leader), CNRS/LIG GETALP team, University Lyon2/GRePS Industrial Partners: CATEL, ISARP, Technosens, FSI Official website of the project: # Cirdo Industrial Research\nMain publications # Making emergency calls more accessible to the older adults through a hands-free speech interface in the house. M. Vacher, F. Aman, S. Rossato, F. Portet, B. Lecouteux, TACCESS (ACM), 25 pages, 2019 (to appear). The CIRDO Corpus: Comprehensive Audio/Video Database of Domestic Falls of Elderly People. M. Vacher, S. Bouakaz, M.-E. Bobillier-Chaumon, F. Aman, R. A. Khan, S. Bekkadja, F. Portet, E. Guillou, S. Rossato, B. Lecouteux, 10th International Conference on Language Resources and Evaluation (LREC 2016), May 2016, Portoroz, Slovenia. ELRA, pp.1389-1396, 2016. Influence of expressive speech on ASR performances: application to elderly assistance in smart home.F. Aman, V. Aubergé, M. Vacher, Text, Speech, and Dialogue: 19th International Conference, TSD 2016, Brno Czech Republic, September 12-16, 2016. CirdoX: an On/Off-line Multisource Speech and Sound Analysis Software.F. Aman, M. Vacher, F. Portet, W. Duclot, B. Lecouteux, Language Resources and Evaluation Conference, May 2016, Portoroz, Slovenia. pp.1978-1985, 2016. Development of Automatic Speech Recognition Techniques for Elderly Home Support: Applications and Challenges.M. Vacher, F. Aman, S. Rossato, F. Portet. Human Aspects of IT for the Aged Population. Design for the Everyday Life. First International Conference, ITAP 2015, Held as Part of the 17th International Conference on Human-Computer Interaction (HCII), Aug 2015, Los Angeles, CA, United States. Springer International Publishing Switzerland, Part II, pp.341-353, 2015, LNCS 9194. CIRDO: Smart companion for helping elderly to live at home for longer.S. Bouakaz, M. Vacher, M.-E. Bobillier-Chaumon, F. Aman, S. Bekkadja, F. Portet, E. Guillou, S. Rossato, E. Desserée, P. Traineau, J.-P. Vimon, T. Chevalier, IRBM, Elsevier Masson, 2014, 35 (2), pp.101-108. Automatic speech recognition of elderly people for home care services.Frédéric Aman, Doctoral thesis, Ecole Doctorale Mathématiques, Sciences et Technologies de l\u0026rsquo;Information, Informatique. University of Grenoble Alpes, 2014. ","externalUrl":null,"permalink":"/en/projects/cirdo/","section":"Projects","summary":"","title":"CIRDO : Compagnon Intelligent Réagissant au Doigt et à l'Oeil (Intelligent Finger and Eye Reacting Companion)","type":"projects"},{"content":"Page carried over from the former website, last updated in May 2018.\nCome and see us # Responsible\nLaurent Besacier\nGETALP Team\nLaboratoire d\u0026rsquo;Informatique de Grenoble\nIMAG Building\n700 Centrale Avenue\nDomaine Universitaire - 38401 St Martin d\u0026rsquo;HèresGETALP is located on the university campus - see access here\nWrite us # GETALP-LIG\nLaboratoire d’Informatique de Grenoble\nBâtiment IMAG\n700 avenue Centrale\nDomaine Universitaire – 38401 St Martin d’Hères (France)\nContact us # A directory of laboratory staff is available here.\nInternships / Jobs # If you wish to submit an unsolicited application, please consult the team\u0026rsquo;s areas of interest and contact the person(s) concerned by e-mail. It is preferable to attach a Curriculum Vitae to your letter (in a standard electronic format).\nThe subjects of the year submitted in MOSIG Master can be read from this link (by typing \u0026ldquo;GETALP\u0026rdquo; in the \u0026ldquo;simple search\u0026rdquo; window on the left)\nWeb Master: getalp-web@imag.fr\n","externalUrl":null,"permalink":"/en/contacts/","section":"GETALP","summary":"","title":"Contacts","type":"page"},{"content":" CNRS Bilateral Projects Franco-Russian cooperation: Development of a multilingual lexical database based on the UNL semantic language and integration of language modules and software ","externalUrl":null,"permalink":"/en/projects/france-russia/","section":"Projects","summary":"","title":"Cooperation France Russia","type":"projects"},{"content":"GETALP creates or contributes to resources (corpora, lexical resources, software…) released to the public and to other researchers. See also the Zenodo, GRICAD GitLab and GitHub collections.\n","externalUrl":null,"permalink":"/en/resources/","section":"Demos and resources","summary":"","title":"Demos and resources","type":"resources"},{"content":"Page carried over from the former website, last updated in May 2018: these offers are closed.\nThesis offers # Custom voice control system in context - 3-year contract, starting September 2017 Internships # Title : Using Language Models to Check Documents \u0026ldquo;Language Quality\u0026rdquo; # Proposed topic in: M2R Informatique, Project # Person(s) responsible : # Herve Blanchon (Herve.Blanchon@imag.fr)LIG-GETALP\nDidier Schwab (didier.schwab@imag.fr)LIG-GETALP\nKeywords: Natural Language Processing, Language Model, Written Document Quality Estimation\nproject duration: 6 months\nMaximum number of students: 1\nPlaces available: 1\nInterview on: 07 November 2017, at 12.11 p.m.\nDescription\nPaid internship with a company\nSupervisors # Hervé BLANCHON (GETALP-LIG) Annelise JOST (Altica) Jérôme GOULIAN (GETALP-LIG) Clarisse BAYOL (Altica) Didier SCHWAB (GETALP-LIG) Keywords # Natural Language Processing Language Model Written Document Quality Estimation Profile of the Candidate # M2 in computer science with an interest in natural language processing and machine learning\nContext # Language Models for a given natural language L calculate the probability that a sequence of words of interest is a valid sequence in the language L. The higher the score, the more likely it is that the sequence is correct in language L. Conversely, if the score is low, it is probable that the sequence is incorrect in language L.\nThe Altica society, located in Voiron (20km away from Grenoble), is specialized in providing translations of documents for its clients. Altica would like to be able to check the quality of both the source document provided by the client and the translated document produced by Altica using a fast and efficient workflow in order to take proper actions.\nOriginality of the proposed subject # To our knowledge language models have not yet been used for the task we propose. On one hand, for a given a source document provided by a client, we would like to give the translator an approximate idea of the quality of the source in terms of lexical and grammatical quality and locate as precisely as possible the problematic segments in this source so that the translator may eventually correct them or ask the client to provide a correction. On the other hand, for a given translation produced by Altica, we would like to give the translator an approximate idea of the quality of its transation in terms of lexical and grammatical quality and locate as precisely as possible the problematic segments in his translation so that he may correct them.\nExpected results # Pratical: # become familiar with the notion of language model build language models using different approaches: n-gram LM, log-linear LM, feed-forward neural LM, recurrent neural network LM evaluate the usability of the proposed LMs in measuring the language quality of a document (spelling, syntax, \u0026hellip;) evaluate the possiblilty of using the proposed LMs to identify problematic segments of a document (for the sake building a user friendy interface of a verification tool) Theoretical: # State of the art: language modelling approaches State of the art: language modelling usages State of the art: language modelling and document cheching Language modelling approaches: Evaluation of pros, cons, suitability for the task References # Dan Jurafsky. (2017). CS 124: From Languages to Information, week 2 (language modelling). https://web.stanford.edu/class/cs124/\nJoshua Goodman. (2003). The state of the art in language modeling. In Proceedings of the 2003 Conference of the North American Chapter of the Association for Computational Linguistics on Human Language Technology: Tutorials - Volume 5 (NAACL-Tutorials \u0026lsquo;03), Vol. 5. Association for Computational Linguistics, Stroudsburg, PA, USA, 4-4. DOI: https://doi.org/10.3115/1075168.1075172\nGraham Neubig. (2017). Neural Machine Translation and Sequence-to-sequence Models: A Tutorial. http://arxiv.org/abs/1703.01619. Section 3.\nDengliang Shi. (2017). A Study on Neural Network Language Modeling. https://arxiv.org/abs/1708.07252\nOren Melamud, Ido Dagan, Jacob Goldberger. (2017) A Simple Language Model based on PMI Matrix Approximations. https://arxiv.org/abs/1707.05266\nYoussef Oualil \u0026amp; Dietrich Klakow (2017). A Neural Network Approach for Mixing Language Models. 5710-5714. https://arxiv.org/abs/1708.06989\nRafal Jozefowicz, Oriol Vinyals, Mike Schuster, Noam Shazeer, Yonghui Wu. (2017). Exploring the Limits of Language Modeling. https://arxiv.org/abs/1602.02410\n","externalUrl":null,"permalink":"/en/employment/","section":"GETALP","summary":"","title":"Employment","type":"page"},{"content":" Acronyme: FORMICAE Title: Algorithme à colonies de fourmis pour la désambiguïsation lexicale multilingue (Ant colony algorithm for multilingual lexical disambiguation) Project Type and Funding: Proj. UPMF Participants: J. Goulian, D. Schwab ","externalUrl":null,"permalink":"/en/projects/formicae/","section":"Projects","summary":"","title":"FORMICAE : Ant colony algorithm for multilingual lexical disambiguation","type":"projects"},{"content":"The GETALP team, which comes from the GEOD and GETA teams of the CLIPS laboratory, has a long history, longer for written language processing (GETA history =\u0026gt; Lien EN) than for oral language processing (GEOD history =\u0026gt; Lien EN), and is building a new common history.\nThematic # GETA (until 2006) # GETA (Groupe d\u0026rsquo;Étude pour la Traduction Automatique) is a multidisciplinary team of computer scientists and linguists. GETA\u0026rsquo;s research topics cover all theoretical, methodological and practical aspects of CAT (Computer Assisted Translation), and more generally multilingual computing. GETA was created by CETA (1961- 1971), a pioneering laboratory of MT in France.\nAt present, GETA remains active in the editor\u0026rsquo;s CAT, but since 1988 has been redirecting its research towards individual CAT, which comprises two components, the translator\u0026rsquo;s CAT and the editor\u0026rsquo;s CAT.\nThe translator\u0026rsquo;s CAT consists in offering linguistic office automation tools to translators (professional or occasional). It\u0026rsquo;s the man who translates. In this field, we work in cooperation with other research groups, which provide linguistic data or tools (lexicons, lemmatisers), and we are interested in computer problems related to the integration of these elements in a form usable by occasional translators, eager to use them from their favorite applications. We have recently developed and proposed, in collaboration with SITE-Eurolang, the Montaigne project, which aims to make the Eurolang-Optimizer software available to the scientific community via the Internet and to use it to create large terminology databases that can then also feed automatic systems.\nThe editor\u0026rsquo;s CAT is the main objective of the work in progress, grouped together in the LIDIA project. The basic idea is to offer a unilingual writer the possibility of writing in his or her own language and, at the cost of a dialogue of standardisation and disambiguation (which should be made as simple and user-friendly as possible), to be translated into several languages, without revision or with minimal revision. So it\u0026rsquo;s dialogue-based CAT (DBMT, for Dialogue-Based Machine Translation) and indirect preediting, but it\u0026rsquo;s the machine that translates. A first model, LIDIA-1, from French to German, Russian and English, has been specified and produced over the last three years.\nGEOD (until 2006) # GEOD\u0026rsquo;s research area is speech and dialogue, to design interaction and spoken communication software and provide systems with a reliable and high-performance language component.\nFor more than fifteen years, the means of communication (mobile phones, Internet) and the media for the electronic dissemination of information (digital radio and television broadcasts) have experienced an ever-increasing boom. At the same time, the progress in digital information processing and computer technology has been enormous. This development has opened up promising prospects for many applications in the field of man-machine or man-man mediatized oral communication, but also for specific applications in the medical field such as remote monitoring of patients at home (intelligent housing). At the same time, thanks to the ease of storage due in part to highly efficient compression algorithms, the corpus of audio and video documents continues to grow. Virtually all multimedia information is now available in digital format and its exploitation opens the field to new applications for indexing and searching documents by content.\nIn this context, GEOD\u0026rsquo;s research theme is centered on Oral Interaction, articulated around two main scientific axes: Recognition (speech, audio and speaker) and Dialogue (modelling and understanding). For these two axes, there are still a certain number of locks linked to the genericity of the models: this characteristic remains an essential objective, located at the centre of our research concerns for the long term.For the Recognition axis, GEOD\u0026rsquo;s research efforts during the period 2001-2005 focused on two sub-themes: the development of multilingual recognition systems for continuous speech and the improvement of their robustness, and the use of speech and sound as a component of multimodal interaction in perceptual spaces. For the Dialogue axis, the objective has been the development of multimodal human-machine dialogue systems.\nGEOD maintains privileged relations with the MICA laboratory (Multimedia, Information, Communication and Applications), one of the CLIPS laboratory\u0026rsquo;s foreign branches.\nDomains # The research on these different types of CAT is centred around computer, linguistic and ergonomic themes.\nThemes with dominant computer science focus\ndistributed architecture and distributed CAT systems (whiteboard technique) specialized languages of the future, adapted to noisy inputs and interactive approaches design and implementation of an IT platform for heterogeneous BDLMs multilingualisation of software and encoding of multilingual texts for CAT Themes with a dominant linguistic focus\ndeclarative formalisms of grammatical specifications (\u0026ldquo;static grammars\u0026rdquo;) and representations of statements (m-structures, f-typed structures) linguistic design and construction of multilingual lexical databases design and testing of MT systems Themes with a dominant ergonomic focus\norganisation of standardisation and disambiguation dialogues integration of multimedia, and in particular multimodal interactive disambiguation techniques in CAT of the editor computer-assisted learning aspects of systems based on complex language skills ","externalUrl":null,"permalink":"/en/history/","section":"GETALP, a long history...","summary":"","title":"GETALP, a long history...","type":"history"},{"content":" A French corpus of audio and multimodal interactions in a health smart home.\nImportant notice # These data have been acquired in the context of a PhD thesis, in a laboratory that works on Health Smart Homes. These databases, devoted to activity recognition and distress recognition, are distributed free of charge, for academic and research use only, in order to be able to compare the results obtained. By downloading these anonymous data, you agree to the following limitations:\nCommercial use: this database is not designed for any commercial use and should only be used by academic researchers. Publications: any publication using the activity database should include a citation of the article: Fleury, A.; Vacher, M. and Noury, N.; « SVM-Based Multi-Modal Classification of Activities of Daily Living in Health Smart Homes: Sensors, Algorithms and First Experimental Results ». IEEE Transactions on Information Technology in Biomedicine, Vol. 14(2), March 2010, pp. 274-283 (DOI: 10.1109/TITB.2009.2037317). Redistribution: this database cannot be, totally or partially, further distributed, published, copied or disseminated, in any way or form, for profit or not. In the same way, all the publications that use the distress call dataset must include a reference to the following book chapter: Vacher, M., Fleury, A., Portet, F., Serignat, J.F., Noury, N.: “Complete Sound and Speech Recognition System for Health Smart Homes: Application to the Recognition of Activities of Daily Living”, in New Developments in Biomedical Engineering, pp. 645–673. Intech Book (2010). For all questions or concerns, do not hesitate to contact us by email: anthony.fleury@mines-douai.fr.\nInformation on these datasets, their contents and the way they are indexed can also be found in the article: Fleury, A.; Vacher, M.; Portet, F.; Chahuara, P. and Noury, N.: “A French corpus of audio and multimodal interactions in a health smart home”. Journal on Multimodal User Interfaces, Springer, 2013, 7 (1), pp. 93-109 (DOI: 10.1007/s12193-012-0104-x).\nContext of the acquisitions # The acquisitions presented here were made during a PhD thesis, in the Health Smart Home of the TIMC-IMAG Laboratory (UMR CNRS 5525), team AFIRM (now part of the AGIM Laboratory, FRE 3405, CNRS-UJF-EPHE). The PhD thesis was advised by Norbert Noury (TIMC-IMAG, now member of the INL Laboratory in Lyon) and Michel Vacher (LIG Laboratory, team GETALP).\nThe smart home of the TIMC-IMAG laboratory # The smart home of the TIMC-IMAG lab is a flat containing a kitchen, a living room, a bedroom, a bathroom, a hall and toilets. This flat is completely equipped and contains everything that can be used for the different activities in the home (including books, clothes, TV, radio, microwave oven, fridge, food, etc.).\nIt is also equipped with several sensors such as microphones, presence infra-red sensors and webcams. It also contains a temperature and hygrometry sensor inside the bathroom, and the dweller wears a home-made accelerometer and magnetometer circuit placed under the left armpit.\nFinally, all the data are stored on different computers placed in the technical room, just next to the flat.\nDistress call dataset # Experimental protocol # The protocol was quite simple. Each participant was alone in the flat. The participant had to go to the living room and close the door. Then, they had to move to the bedroom and read aloud the first half of one of the five successions of sentences, out of 10 normal and 20 distress sentences. Afterwards, the second half of the set of sentences had to be uttered in the living room. Finally, each participant was called 3 times and had to answer the phone and read the predefined phone conversation (5 sentences each). The sentences were uttered in the flat, with the participant sitting down or standing up. They had no instructions on how to orient themselves relative to the microphones or on the way they pronounced the different sentences.\nDataset content and organisation # This dataset has been acquired on 10 different subjects, all native French speakers:\nID Age Gender Weight (kg) Height (m) Sentence set 1 61 M 82 1.70 1 2 22 M 77 1.77 2 3 53 M 67 1.68 3 4 47 M 56 1.70 4 5 31 M 62 1.70 5 6 25 M 78 1.84 2 7 51 F 64 1.65 4 8 29 F 53 1.60 3 9 28 F 61 1.69 1 10 25 M 90 1.88 5 The last column gives the number of the sentence set each subject pronounced: there were 5 different sets of sentences (normal, distress and phone conversations). The sets were distributed as a spreadsheet, one sheet per set (no longer available online).\nThe dataset is organised as follows:\nA directory containing the complete dataset for each subject, with ALL the sounds captured during the sessions. For each subject: 8 directories, one per channel; for each channel, the files of each acquisition: mainly the WAV file with the acquired sound and the XML file with its analysis by our system (segmentation as sound or speech; for sounds, the sound class; for speech, the five most probable sentences); a text file (Resultats.txt) with a first analysis: for sounds, the SNR, the ID of the corresponding WAV/XML files and the sound class; for speech, the SNR, the ID of the WAV/XML files and the three most probable sentences. A directory containing, for each speaker, only the files of the sounds with the best SNR, for each sentence uttered during the experiment (all other sounds have been deleted). The order of the sentences follows the spreadsheet. A directory containing all the sound (and associated) files with the best SNR and all sounds with an SNR above 80% of that value. The order of the sentences is kept, and the XML files contain the timestamps and the identification of all simultaneous sounds, so that each sound can be matched to its sentence. More information on the experiment, the datasets and the processing is given in the references [Vacher2008], [Vacher2010] and [Vacher2011] below.\nData # ID Whole set Best SNR SNR ≥ 80% of best 1 zip (19 MB) zip (2.1 MB) zip (5 MB) 2 zip (22 MB) zip (2.2 MB) zip (4.9 MB) 3 zip (21 MB) zip (2.1 MB) zip (5.3 MB) 4 zip (22 MB) zip (2.2 MB) zip (5 MB) 5 zip (20 MB) zip (2.1 MB) zip (4.6 MB) 6 zip (32 MB) zip (2.1 MB) zip (4.9 MB) 7 zip (32 MB) zip (2.2 MB) zip (4.2 MB) 8 zip (27 MB) zip (1.7 MB) zip (3.6 MB) 9 zip (21 MB) zip (2.1 MB) zip (4.4 MB) 10 zip (19 MB) zip (1.8 MB) zip (3.3 MB) All zip (231 MB) zip (21 MB) zip (45 MB) Activities of Daily Living dataset # Experimental protocol # The first step of the experiment was a thorough visit of the flat, to allow the person to act as much as possible as if they were at home. After this visit, the only requirement was to perform a set of 7 activities, in the order and for the duration they wanted, and as many times as they wanted:\nSleeping: lie in the bed and stay “asleep” for a reasonable time (of your choice). Resting: sit on the couch or a chair in the living room and do what you like when relaxing (watching TV, reading, doing nothing…). Feeding: prepare a breakfast with the equipment and ingredients in the kitchen cupboard, then eat at the kitchen table, then wash the dishes in the kitchen sink. Hygiene: wash your hands and face, and pretend to brush your teeth in the bathroom. Toilets: pretend to go (or go) to the toilet (sit on the toilet, flush…). Dressing: put on and remove the clothes in the chest of drawers near the bed (over your own clothes), before and after sleeping or after the hygiene task. Communication: you will be called on the phone (in the living room) several times; each time, read the sentences indicated near the phone (one long phone conversation and three short ones). For more information about the dataset and the experiments, see the references [Fleury2010], [Fleury2011], [Vacher2011] and [Noury2011] below.\nDataset content, organisation and download # The data of 2 of the 15 subjects are available on this page. To obtain the complete dataset with all participants, fill in the agreement form (PDF) and send it to anthony.fleury@mines-douai.fr, committing to use the dataset correctly; you will then receive the download link for the remaining participants.\nID Age Gender Height (m) Weight (kg) Phone conversation Native French speaker 1 25 F 1.77 63 1 yes 2 25 F 1.78 70 2 yes 3 32 M 1.92 80 3 yes 4 57 F 1.48 50 4 no 5 25 M 1.72 81 5 yes 6 26 M 1.82 66 1 no 7 28 M 1.78 68 2 yes 8 31 M 1.70 63 3 yes 9 26 M 1.84 78 4 yes 10 24 F 1.62 68 5 no 11 29 F 1.63 57 1 yes 12 40 F 1.76 60 2 yes 13 33 M 1.75 79 3 no 14 34 M 1.85 76 4 yes 15 43 M 1.72 69 5 yes The phone conversations were distributed as a spreadsheet (no longer available online).\nDownload # All files are gzipped tar archives.\nID Video Annotations Audio Actimetry (raw) Sensors Complete set 2 02_video (719 MB) 02_annotations (6.3 kB) 02_son (46 MB) 02_actim6d (4.1 MB) 02_sensors (36 kB) 02_all (768 MB) 9 09_video (528 MB) 09_annotations (5.8 kB) 09_son (38 MB) 09_actim6d (2.8 MB) 09_sensors (35 kB) 09_all (568 MB) Video # The videos come from 4 cameras, compressed in MPEG-4 (Xvid) and mixed into a single video: the kitchen at the top left, the kitchen/hall/bathroom door at the bottom left, the bedroom and living room at the top right, and the living room only at the bottom right. Above, subject 2 (left) and subject 9 (right).\nFor ethical reasons there is no camera in the bathroom and toilets. When entering this part of the flat, the dweller partially closed the door to go to the bathroom (hygiene activity) and closed it completely to go to the toilets.\nActivity annotation # An XML file annotates the activities performed by the dweller during the session.\nIt first gives the information needed to synchronise the actimetry sensor (ACTIM6D) with the other data: all data are synchronised because the clocks of all acquisition computers were synchronised to the same NTP server. For this, the position corresponds to the date given in the ACTIM6D_Synchro section of the XML file.\nIt then gives the Activities of Daily Living performed during the session. For each activity, two timestamps are given: the beginning (debut) and the end (fin). In Activite, the id identifies the annotated activity:\nID Activity 1 Sleeping 2 Resting (TV, reading, radio…) 3 Dressing/undressing 4 Eating 5 Toilet use 6 Hygiene 7 Communication In addition, an Advene annotation file contains:\nthe ground-truth annotation of activities (as above, with file positions instead of timestamps); the ground-truth annotation of posture (for some of the subjects, see Actimetry below); the ground-truth annotation of the use of furniture (see below). Both files are in the Annotations archive: Y.xml and AdveneAnnotation_Y.xml, where Y is the subject ID.\nAudio # As for the distress call dataset, there is one directory per microphone and a text file summarising the classification of each sound that occurred in the flat. All files are in one archive.\nActimetry # ACTIM6D is a home-made circuit board with a 3-axis accelerometer and a 3-axis magnetometer, fixed under the left armpit of the subject in a pocket of a specific T-shirt. Its data are used to detect postural transitions and walking.\nThe raw data are given so that users can implement their own postural-transition detection (transitions can be annotated using the video). Each line of the file contains:\nthe experiment number (an integer from the card\u0026rsquo;s flash memory); the second of the recording; the accelerometer values on the X, Y and Z axes (hexadecimal); the magnetometer values on the X, Y and Z axes (hexadecimal); the temperature measured by the board (decimal). For more details about this sensor and how to process its data, see [Fleury2009].\nPresence infra-red sensors # Each movement detected in one of the rooms is logged in a text file (HIS_Y.txt, in the Sensors archive) with the date (DD/MM/YYYY), the time (HH:MM:SS) and the room.\nHygrometry and temperature # A device in the bathroom gives the temperature and hygrometry every 5 minutes, in a text file (Tempature.temp_Y.txt, in the Sensors archive) with a timestamp (DD.MM.YYYY-HH:MM), the temperature (°C) and the hygrometry (%).\nChest of drawers, fridge and cupboard # Sensors record when the chest of drawers, the fridge and the cupboard (containing everything needed to prepare meals) are opened or closed. Each text file (Cupboard_Y.txt, Drawers_Y.txt, Fridge_Y.txt, in the Sensors archive) lists the changes of state of the door: timestamp (DD-MM-YYYY-HH-MM-SS) and new state (0 closed, 1 open).\nReferences # [Fleury2013] Fleury, A.; Vacher, M.; Portet, F.; Chahuara, P.; Noury, N. « A French corpus of audio and multimodal interactions in a health smart home ». Journal on Multimodal User Interfaces, Springer, 2013, 7 (1), pp. 93-109. Distress call dataset # [Vacher2008] Vacher, M.; Fleury, A.; Serignat, J.-F.; Noury, N.; Glasson, H. « Preliminary Evaluation of speech/sound recognition for telemedicine application in a real environment ». Interspeech'08, Brisbane, Australia, 2008, pp. 496-499. [Vacher2010] Vacher, M.; Fleury, A.; Portet, F.; Serignat, J.-F.; Noury, N. « Complete Sound and Speech Recognition System for Health Smart Homes: Application to the Recognition of Activities of Daily Living ». In New Developments in Biomedical Engineering, ISBN 978-953-7619-57-2, pp. 645-673, 2010. [Vacher2011] Vacher, M.; Portet, F.; Fleury, A.; Noury, N. « Development of Audio Sensing Technology for Ambient Assisted Living: Applications and Challenges ». International Journal of E-Health and Medical Communications, 2(1):35-54, 2011. [pdf] Activities of Daily Living dataset # [Fleury2009] Fleury, A.; Noury, N.; Vacher, M. « A wavelet-based pattern recognition algorithm to classify postural transitions in humans ». EUSIPCO 2009, 17th European Signal Processing Conference, Glasgow, Scotland, 2009, pp. 2047-2051. [Fleury2010] Fleury, A.; Vacher, M.; Noury, N. « SVM-Based Multi-Modal Classification of Activities of Daily Living in Health Smart Homes: Sensors, Algorithms and First Experimental Results ». IEEE Transactions on Information Technology in Biomedicine, 14(2), 2010, pp. 274-283 (DOI: 10.1109/TITB.2009.2037317). [Fleury2011] Fleury, A.; Noury, N.; Vacher, M. « Improving Supervised Classification of Activities of Daily Living Using Prior Knowledge ». International Journal of E-Health and Medical Communications, 2(1):17-34, 2011. [pdf] [Noury2011] Noury, N.; Poujaud, J.; Fleury, A.; Nocua, R.; Haddidi, T.; Rumeau, P. « Smart Sweet Home… a pervasive environment for sensing our daily activity? ». In Activity Recognition in Pervasive Intelligent Environments, Atlantis Ambient and Pervasive Intelligence, vol. 4, Atlantis Press, 2011, ISBN 978-90-78677-42-0. [Chahuara2012] Chahuara, P.; Fleury, A.; Portet, F.; Vacher, M. « Using Markov Logic Network for On-line Activity Recognition from Non-Visual Home Automation Sensors ». Lecture Notes in Computer Science vol. 7683, pp. 177-192, 2012. ","externalUrl":null,"permalink":"/en/resources/his/","section":"Demos and resources","summary":"A French corpus of audio and multimodal interactions in a health smart home: distress calls and activities of daily living.","title":"Health Smart Home (HIS) datasets","type":"resources"},{"content":"GEOD started its activity in 1997, in the field of speech and dialogue research, to design interaction and spoken communication software and equip systems with a reliable and efficient language component.\nSince the 1990s, the means of communication (mobile phones, Internet) and the media for the electronic dissemination of information (digital radio and television broadcasts) have grown steadily. At the same time, the progress in digital information processing and computer technology has been enormous. This development has opened up promising prospects for many applications in the field of man-machine or man-man mediatized oral communication, but also for specific applications in the medical field such as remote monitoring of patients at home (intelligent housing). At the same time, thanks to the ease of storage due in part to highly efficient compression algorithms, the corpus of audio and video documents continues to grow. Virtually all multimedia information is now available in digital format and its exploitation opens the field to new applications for indexing and searching documents by content.\nSince 1997, GEOD\u0026rsquo;s main research objectives have been (a) the development of large vocabulary recognition systems for continuous speech, integrating robustness for speech acquisition and recognition, (b) the development of multimodal human-machine dialogue systems, (c) multimodal interaction in perceptual environments and spaces. To achieve these objectives, transversal research (d) on corpora and learning tools is still needed.\nRobust multilingual continuous speech recognition # Acoustic pre-treatment, taking into account usability under real conditions and capable of achieving robust systems under severe acoustic conditions, constitutes the major problem of speech recognition. GEOD focused its efforts on locating the speaker using a microphone antenna (multi-sensor acquisition and signal inter-correlation technique), blind source separation method, acoustic reverberation cancellation by room frequency response estimation.\nIn the field of acoustic modelling for continuous speech recognition, GEOD has concentrated its research activities on techniques based on hidden Markov chains (or HMM), and more specifically on the statistical optimisation of phoneme models (context dependent and independent) by learning on large multi-speaker speech corpora (BREF80, BREF120 : two LIMSI bases and BRAF100 base registered at CLIPS with 10000 sentences and 100 speakers - see action \u0026ldquo;Transverse searches\u0026rdquo;).\nThe team is also studying acoustic modelling for voice servers, with the objective of minimizing signal distortion following speech signal compression, resulting in degraded recognition system performance.\nGEOD is also developing a recognition system for Vietnamese, which is a 6-tone language, as part of CLIPS\u0026rsquo; scientific cooperation with the Hanoi Polytechnic Institute (IPH).\nOn language models and theme detection, the approach chosen by GEOD consisted in using the WEB as a source of information to constitute the learning corpus from which the vocabulary is extracted and developed the language model. A minimum block filtering technique (vocabulary functions) was developed (D. Vaufreydaz), involving the use of linguistic lexicons and dictionaries (BDLEX 50000 and ABU: Association des Bibliophiles Universels, were used).\nSince 2001, GEOD\u0026rsquo;s research has evolved from supporting large vocabulary to multilingualism.\nGEOD\u0026rsquo;s research activities in this theme have focused on the development of multi-speaker acoustic models and language models for the laboratory\u0026rsquo;s French language automatic continuous speech recognition system. The originality lies in the approach that consists in \u0026ldquo;aspiring\u0026rdquo; a large number of websites in a given language and filtering the recovered text data in order to make them usable for calculating statistical language models. An adaptation of this methodology to poorly endowed languages marks a trend towards multilingualism that is becoming increasingly important in this research. Applications to the Vietnamese language, the Khmer language and Spanish-Mexican (Castilian) have been considered and have produced very encouraging results. Extensions of this research theme, in the sense of an \u0026ldquo;enriched transcription\u0026rdquo; (segmentation into speakers, detection of areas of interest, detection of audio \u0026ldquo;jingles\u0026rdquo;,\u0026hellip;) for information search applications by the content in the databases, were also conducted as part of various participation in international evaluation campaigns. Finally, work for biometric applications has been carried out taking into account the often multimodal nature of the field.\nPerceptual environments: speech and sound as a component of multimodal interaction # In collaboration with the TIMC laboratory, GEOD is developing the general concept of an \u0026ldquo;intelligent living space\u0026rdquo;. This involves designing rooms equipped with several types of sensors and managed by a computer system that analyses signals in real time so as to intervene automatically according to the needs, demands and expectations of human stakeholders. The use of sound sensors and specific processing on speech signals or everyday noise is an innovative approach.\nThis theme describes GEOD\u0026rsquo;s research work on speech and sound, in the application context of perceptual spaces and more particularly in the framework of cooperation with the TIMC13 laboratory for Intelligent Housing for Health (HIS). In the TIMC premises, an apartment (30m2) has been equipped to become a prototype HIS. Various algorithms for detection and classification of everyday sounds have been developed and validated for the detection of distress situations of a patient under remote medical monitoring. Similarly, a language model for the GEOD speech recognition system has been adapted for the recognition of distress calls in this environment. Some developments for \u0026ldquo;smart room\u0026rdquo; applications are also presented.\nDevelopment of multimodal H-M dialogue systems # The theme of human-machine dialogue encompasses oral interaction and multimodal interaction (speech and gestures). Modelling human-machine dialogue poses theoretical problems because human dialogue cannot be considered as a fully planned activity: at each moment, the interlocutors can make impacts or breaks, they use strategies that they adapt during the interaction according to the goals to be achieved and the opportunities offered by the situation.\nBetween 1997 and 1999, GEOD was mainly interested in 4 of the modules of the schematic representation of a dialogue system, namely the first 4 of: the management of the goals related to the task, the understanding of the speaker\u0026rsquo;s statements, their interpretation in a dialogue situation and relative to the goals to be achieved, the control and management of the dialogue and finally the generation of outputs: text, speech or graph.\nThis theme describes GEOD\u0026rsquo;s research work on human-machine dialogue. The main advances have been in game theory and dialogue representation theory (SDRT = Segmented Discourse Representation Theory). The analysis and exploitation of corpora continued in order to study the speakers\u0026rsquo; expectations, their modes of understanding and their behaviour in the face of expressive conversational agents. For this purpose, various dialogue situations were simulated, notably within the framework of the ACE project (Expressive Conversational Agent). Finally, the DCR16 method was further developed to obtain a validated procedure in automatic evaluation of dialogue systems. All this research has focused on application areas provided by the PVE (Portal Vocal d\u0026rsquo;Entreprise) project, the aim of which is to develop natural speech dialogue services for the company\u0026rsquo;s social life (e.g. organisation of meetings, personal agendas, etc.). An opening towards dialogue with several speakers has been initiated, which places the team on an original ground.\nTransversal researches # Transversal research related to the axes described above concerns linguistic resources (written and oral corpora) and more generally the tools and methodologies necessary for the acquisition, production or management of these linguistic resources, but also very sophisticated phonetic alignment or automatic learning methods.\n","externalUrl":null,"permalink":"/en/history/geod/","section":"GETALP, a long history...","summary":"","title":"History of GEOD","type":"history"},{"content":"GETA (Groupe d\u0026rsquo;Etude pour la Traduction Automatique) was a multidisciplinary team of computer scientists and linguists. GETA\u0026rsquo;s research topics cover all theoretical, methodological and practical aspects of CAT (Computer Assisted Translation), and more generally multilingual computing. GETA was created by CETA (1961- 1971), a pioneering laboratory of MT in France.\nHistory of GETA (until 2006) # Research in MT, then in CAT, has been conducted in Grenoble since 1961 at CETA, which became GETA in 1971. Our research can be divided into three main periods, 1961-70, 1971-85, and 1986-99.\n1961-70 is the pioneer period, when, under the direction of Prof. Dr. Klaus Bühler, B. Vauquois, CETA (Centre d\u0026rsquo;Etudes pour la Traduction Automatique), CNRS\u0026rsquo;s own laboratory, built the first \u0026ldquo;second generation\u0026rdquo; system in the world, based on a hybrid \u0026ldquo;pivot language\u0026rdquo;, and tested it on very large volumes of texts. It was then MT without pre-publication or post-edition, which we call today MT of the watchman. This period is characterized by the enlightened application of the theory of formal languages and their compilation to the problem of MT, as well as by the integration of advanced linguistic theories, with in particular the passage through a \u0026ldquo;pivot language\u0026rdquo;.\nResearch continued from 1971 within the framework of a university laboratory associated with the CNRS, the GETA (Groupe d\u0026rsquo;Etude pour la Traduction Automatique), directed by B. Vauquois until his death on 30/9/85. GETA has been a member of IMAG since its creation in 1986.\nFrom 1971 to 1980, the GETA, turning to the editor\u0026rsquo;s MT, developed and tested a new CAT methodology (multi-level transfer approach) and a CAT systems generator (Ariane-78) for writing multilingual CAT models or prototypes using specialized languages for linguistic programming (symbolic languages of production rules).\nFrom 1981 to 1987-88, GETA actively participated in technology transfers to industry (ESOPE project, then ADI national PN-TAO project in 81-82 and 83-87), while carrying out in-house a Russian-French \u0026ldquo;pre-operational\u0026rdquo; system for DRET, several studies for the Eurotra project, and helping various foreign groups to build models or prototypes of MT systems. The Ariane-78 system was profoundly revised and extended, and gave birth to the current Ariane-G5.\nThe disappearance of B. Vauquois, in the middle of the PN-TAO, for which his participation was very important, was a great shock. From 1986 to 1990, the team tried to carry out the various actions in progress, while restructuring and reducing its activity of development and experimentation in full scale to the profit of a more fundamental research. From 1988-90, considering that the technology transfer on the editor\u0026rsquo;s CAT had been carried out, GETA reoriented its research towards the individual CAT, which comprises two components, the translator\u0026rsquo;s CAT and the editor\u0026rsquo;s CAT.\nThe main works concern the editor\u0026rsquo;s CAT and are grouped together in the LIDIA project. The basic idea is to offer a unilingual writer the possibility of writing in his or her own language and, at the cost of a standardization and clarification dialogue (the least cumbersome and most user-friendly possible), to be translated into several languages, without revision. So it\u0026rsquo;s dialogue-based CAT (DBMT, for Dialogue-Based Machine Translation) and indirect preediting, but it\u0026rsquo;s the machine that translates. In translator CAT, we study the computer and ergonomic aspects related to the construction of linguistic office automation tools intended for translators (professional or occasional).\nOngoing research focuses on computer science, linguistics and ergonomics, in particular studies on interactive disambiguation techniques, on multilingual lexical bases by acceptance, on the integration of computer-assisted learning (CAL) possibilities, on the individual discovery of linguistic knowledge, and on the experimentation of modern software engineering techniques to (re)build computer tools (languages and specialised environments for linguists and lexicographers) adapted to the different types of CAT, and in particular to the individual CAT of the future.\n","externalUrl":null,"permalink":"/en/history/geta/","section":"GETALP, a long history...","summary":"","title":"History of GETA","type":"history"},{"content":" Title : **Intelligent Habitat for Home Assistance # (non-funded team project)\nProject Type and Funding # This is a long-term research theme that aims to help people, usually elderly or in difficulty, living alone at home.\nThe RESIDE-HIS (2000-2002), DESDHIS (2002-2004) and DESDHIS2 (2004-2008) projects were used as pilot projects. ANR funding for the Sweet-Home (2009-2013) and Cirdo (2010-2014) projects has helped to advance and enrich this theme.The ANR VOCADOM project (2017-2021) is part of this theme.\nKeywords # Intelligent Housing; Smart Home; detection of distress keywords; speech recognition in remote conditions; recognition of everyday sounds; context recognition; ADL (activities of daily living); home automation; voice commands; contextual interaction; intelligent home automation control.\nHistory # The original aim of the RESIDE-HIS project was to detect a state of emergency by sound analysis in the apartment with fusion of information from presence sensors and a fall sensor. The DESDHIS project focused on the recognition of calls for help and the classification of the person\u0026rsquo;s Activities of Daily Living, a variation in ADLs that could be a sign of an emerging health problem. The Sweet-Home project aimed to voice control home automation equipment in the intelligent apartment to assist frail people, using a specialized social link system between the isolated person and his family. The objective of the CIRDO project was the detection of falls by video analysis and the recognition of distress calls in this case. The VOCADOM project selected by the ANR in 2016 aims to carry out a voice command of home automation under realistic conditions of use: presence of noise in the apartment, presence of a visitor\u0026hellip;\nSee also more details on the site corresponding to each project:\npilot projects: RESIDE-HIS / DESDHIS the Sweet-Home project (scientific resp. Michel Vacher): SWEET-HOME the CIRDO project (resp. scientific GETALP Michel Vacher): CIRDO the VOCADOM project (scientific project Michel Vacher): VOCADOM GETALP team participants # Michel VACHER François PORTET Benjamin LECOUTEUX Solange ROSSATO Véronique AUBERGE Fabien RINGEVAL Test platform # Currently, this project is based on the LIG\u0026rsquo;s DOMUS apartment (see below) and the Amiqual4Home platform.\nContact : Michel.Vacher # ","externalUrl":null,"permalink":"/en/projects/smart-home-assistance/","section":"Projects","summary":"","title":"Intelligent Housing for People Assistance (non-funded team project)","type":"projects"},{"content":"","externalUrl":null,"permalink":"/en/members/","section":"Members","summary":"","title":"Members","type":"members"},{"content":" ANR Project Keywords : biometrics, platform, open-source Description : This project focuses on multimodal biometrics with partners such as LIA, EURECOM, IRIT and LIUM. The MISTRAL project offers an open-source software platform for biometric authentication. The objective of this project is to facilitate access to biometric technologies to the academic world (research and teaching centre) as well as to the industrial world by providing a complete, high-performance, modular software platform, easy to master and implement, capable of managing different biometric tasks and different environments. One of the expected outcomes of the project is the animation of a strong MISTRAL platform user community, composed of academic and industrial actors. Contact : http://mistral.univ-avignon.fr/ ","externalUrl":null,"permalink":"/en/projects/mistral/","section":"Projects","summary":"","title":"MISTRAL : Open Source Platform for Multimodal Biometrics","type":"projects"},{"content":" Networked Research Action, funded by the Agence Universitaire de la Francophonie Keywords: Motamot, dictionnary, Jibiki, Franch, vietnamese, khmer, contributory platform, collaborative project Description: This project consists in developing a multilingual lexical system by simultaneously building several bilingual dictionaries sharing at least one language. The bilingual dictionaries will be built online on a \u0026ldquo;Papillon\u0026rdquo; type site built on the Jibiki platform according to a collaborative and volunteer work methodology, inspired by the Wikipedia project. The bilingual links created during the writing of the articles are used on the one hand to generate bilingual reverse links, and on the other hand to create new interlingual links. These dictionaries will also be available in the form of a multimedia version, with a user-friendly and ergonomic interface, the results of which will be accessible via two media: language (text) and speech. For certain languages (which we will treat as examples), synthesis modules will be added to enable learners to participate in language learning and to give sound examples of the \u0026ldquo;isolated words\u0026rdquo; or \u0026ldquo;short sentence book\u0026rdquo; type to learners. In addition, we will develop APIs (Application Programming Interfaces) that can easily be integrated into Internet pages, forums, blogs or any other application allowing dictionaries to be integrated into applications such as text editors, automatic translation, language learning tools, etc. in order to participate in the development of the computerization of poorly endowed languages. The main objectives of this project are therefore: the launch of a dynamic contribution around the construction of each bilingual dictionary in presence. Wikipedia\u0026rsquo;s success shows that this is possible, provided you have simple and easy to use tools. large-scale laboratory experiments such as the DiCo base (Mel’čuk et Polguère 2006) or the PARAX system (Blanc 1996). the development of a testing ground for the validation of several hypotheses formulated in previous work: bijectivity of bilingual links and transitivity of interlingual links, massive contribution on the Web, construction of a multilingual lexical system. Participants Mathieu Mangeot (project manager) Hong-Thai Nguyen Cong-Phap Huyn Geneviève Caelen (MICA) Éric Castelli (MICA) Yeng Vanra Parteners: MICA-CNRS laboratory in Vietnam Cambodia Institute of Technology Association pays perdu (Denis Richer) ","externalUrl":null,"permalink":"/en/projects/motamot/","section":"Projects","summary":"","title":"MOTAMOT : Development of a multilingual lexical system through the construction of bilingual dictionaries targeted at low-computer languages in South-East Asia","type":"projects"},{"content":" ANR Project 2008 - 2011 Keywords: Categorization and search by image content, Multimodal learning and classification, Multilingual text data analysis, UNL Description: The OMNIA project aims to filter documents containing text and images, in a context of large amounts of data, such as those found in intranets and on the Internet, and to present them to users in a content processing tool such as DocuShare (Xerox). The originality of the project is to work on 3 dimensions: image, text, emotion and in a multilingual context. The image and the text give rise to 2 categories relating to the informational aspects and to the specific emotional aspects (coming directly from the image, or relating to its perception expressed in the text). These two types of content will be treated independently (annotation then indexing and categorization), with learning techniques, and then merged at the level of the filtering and query tool. Their \u0026ldquo;primitives\u0026rdquo; will be linked to an interlingual representation of meanings based on English (UNL), paving the way for multilingualism in the \u0026ldquo;publication\u0026rdquo; of document categories, and the handling of questions in natural languages with UNL dictionaries. Participants: David Rouquet, Christian Boitet, Valérie Bellynck, Hervé Blanchon, Achille Falaise, Jérôme Goulian, Jean-Philippe Guilbaud, Hong-Thai Nguyen , Didier Schwab Partners: XEROX, Xerox Research Center Europe, Luca Marchesotti LIRIS, Ecole Centrale de Lyon, Liming Chen Publications : 2010, Achille Falaise, David Rouquet, Didier Schwab, Hervé Blanchon, Christian Boitet, Ontology driven content extraction using interlingual annotation of texts in the OMNIA project, workshop CLIA, COLING, Bejin, China 2010, David Rouquet, Achile Fallaise, Didier Schwab, Hervé Blanchon, Vallérie Bellynck, Christian Boitet, Emmanuel Dellandréa, Ningning Liu, Liming Chen, Alexandre Saidi, Sandra Skaff, Luca Marchesotti, and Gabriela Csurka, Classifiation multilingue et multimédia pour la recherche d\u0026rsquo;images dans le projet OMNIA,workshop RISE, Inforsid, Marseille, France 2009, David Rouquet and Hong-Thai Nguyen, Interlingual annotation of texts in the OMNIA project, LTC09, Poznan, Poland 2009, David Rouquet and Hong-Thai Nguyen, Multilinguïsation d’une ontologie par des correspondances avec un lexique pivot, TOTh09, Annecy, France ","externalUrl":null,"permalink":"/en/projects/omnia/","section":"Projects","summary":"","title":"OMNIA","type":"projects"},{"content":" PEA DGA Participants: L. Besacier Partners: Cassidian/EADS - LIUM - Systran - LIG - U. Edimbourg - LNE - ELDA ","externalUrl":null,"permalink":"/en/projects/pea-trad/","section":"Projects","summary":"","title":"PEA-TRAD : Translation for document analysis assistance","type":"projects"},{"content":"Project ANR BLANC 2011-2015\nKeywords: Experimental phonetics, perception, phonology, universals, syllables, corpus, dialectology, phonetic change, metathesis, speech recognition, rare languages, poorly endowed languages, automatic syllabification.\nParticipants: Solange Rossato, Laurent Besacier\nPartners: GIPSA, LIDILEM, LIG, LPP, ULB, MICA\n","externalUrl":null,"permalink":"/en/projects/syllable-asymmetries/","section":"Projects","summary":"","title":"Phonetic and phonological asymmetries of the syllable","type":"projects"},{"content":" ANR BLANC 2009-2012 Description : Despite the factors, that at a first glance might make a language appear less important, and thus unnecessary as target for human language technologies (HLT), good reasons exist for developing speech recognition systems for literally all languages in the world. First, the diversity of languages in the world is the basis of the rich cultural diversity. However in today’s globalized world, languages are frequently disappearing. The ongoing extinction of many languages is in part caused by a switch to more prevalent languages that might give their speakers an economic advantage. The lack of HLT systems for these languages only accelerates their extinction while on the other side HLT could help to stop this trend by making the less prevalent languages more attractive to their original speakers. A second reason why HLT should be available for all languages is that the political impact of a language can be very volatile. In today’s world, language is one of the few remaining barriers that hinder human-to-human interaction. Events such as armed conflicts or natural disasters might make it important to be able to communicate with speakers of a less-prevalent language, e.g. for humanitarian workers in a disaster area. Here, readily available technology such as speech translation systems can be highly beneficial. Such technology might be far from being perfect, but when being faced with the alternative of having no translation system at all for an unknown language in an emergency situation, the imperfect system will be of great use. Therefore, HLT needs to be developed especially for under-resourced languages! Nowadays, almost all of the techniques and methods in spoken language technologies, in particular the automatic speech recognition (ASR) systems, use statistical approaches. However, given the statistical nature of these methods, a large amount of resources (vocabularies, text corpora, transcribed speech corpora, phonetic dictionaries) is crucial and required to train models and to test the performances of the systems. Consequently, a large speech corpus which contains hours of signals recorded by hundreds speakers (for acoustic modeling) and a text corpus with million words (for language modeling) is currently necessary for building an ASR system for a new language. However, these crucial resources are not directly available for under-resourced languages. Thus, a methodology for rapidly building them is necessary, and in the mean time, strategies to exploit a minimum amount of resources are necessary. From a scientific point of view, the interest and originality of this project consists in proposing viable innovative methods that go far beyond the simple retraining or adaptation of acoustic and linguistic models. Consequently, a significant breakdown is needed to develop ASR systems for Π-languages. For instance we plan to question the use of the word as a fundamental unit for language modelling : low complexity language models could be obtained by using sub-word units (morphemes, syllables or even characters) which might be of interest when few training data is available. Concerning acoustic modelling, the originality of this project lies in the proposal of large coverage multilingual acoustic models based on our knowledge of the world speech sounds systems. Another goal is to explore more “language independent” ASR systems that would adapt themselves (without or with very few supervision) to the audio flow they have at their input. From an operational point of view, this project aims at providing a free open source ASR development kit for Π-languages. This goal is realistic since elements of it have been already developed in a preliminary fashion by some partners of this consortium : text data collection and filtering tools (LIG, LIA), ASR training and decoding tools (LIA), phone mapping tools from a source to a target language (LIG). We plan to distribute and evaluate such a development kit by deploying ASR systems for new under-resourced languages with very poor resources (khmer and lao for instance). If successful, this project would lead to a dynamic user group of our development kit, composed by research teams or individuals doing ASR research for their own language. It is also important to note that some under-resourced languages could show, in the future of their development, a very strong economic potential: Bengali, Malay and Vietnamese are for instance in the top-20 of the most spoken languages in the world. Some other languages may be of great interest for governmental or non governmental projects involving global security (see example of Iraqi dialect in US DARPA projects) or humanitarian issues. Finally, in the objective of saving some endangered languages (some mostly spoken and not written), the possibility to rapidly develop ASR systems to transcribe them is an important step for their preservation and would facilitate access to audio contents in these languages (notably languages from Africa). ","externalUrl":null,"permalink":"/en/projects/pi/","section":"Projects","summary":"","title":"PI : Spoken Language Technologies for PI languages","type":"projects"},{"content":"List carried over from the former website, last updated in November 2019.\nADORTALP : Augmenter et Déployer des Outils Reproductibles de Traitement Automatique des Langues et de la Parole - Carnot Cognition - 30k€ BULB : Breaking the Unwritten Language Barrier - ANR/DFG - 70k€ Chaire IA (MIAI) : AI \u0026amp; Language - 290k€ DBnary : Multilingual Lexical Linked Open Data (projet d\u0026rsquo;équipe, non financé) DECORE : DEep COnvolutional and REcurrent networks for image, speech and text (Labex Persyval) 250k€ HIS : Intelligent Housing for People Assistance (non-funded team project) AXiMAG : interactive Multilingual Access Gateway ON-TRAC : End-to-End Neural Approaches for Speech Translation - ANR 2018 - 200 k€ Projet TecknowMetrix - Modèles de langue pour recommendation de documentations scientifiques - 25k€ THESES-CIFRE : 4 PhD students are currently funded by a CIFRE contract with the following industrial partners: ATOS, Calystene, Ixiade, LNE VOCADOM : Robust voice control adapted to the person and the context for autonomy at home - ANR - 300k€ Past projects # AIM-WEST : Analysis and Integration of Polylexical Expressions in Automatic Translation of Writing and Oral - 30k€ ALFFA : African Languages in the Field - Fundamentals and Automation - ANR 136k€ Ariane-Y : Automatic translator development environment (non-funded team project) APPSy: Asymétries Phonétiques et Phonologiques de la Syllabe (Syllable Phonetic and Phonological Asymmetries) - ANR 48k€ Arioca: Ariane, CASH and UNL working group CAMELEON: Collaborative and Automatic Methods for the Multilingualisation of Lexica and Ontologies - CAPES COFECUB (France-Brazil) CAMOMILE : Collaborative annotation of multi-modal, multi-lingual and multi-media documents - ANR 250k€ CASSIE : Platform of interoperable uses of connected objects driven in natural language - Investments for the Future (BPI) 440k€ CIRDO : Compagnon Intelligent Réagissant au Doigt et à l\u0026rsquo;Oeil (Smart Companion Finger and Eye Reaction) - ANR 210k€ COOP-FR-RUSSIE: Franco-Russian cooperation: Development of a multilingual lexical database based on the UNL semantic language and integration of language modules and software - CNRS 8k€ DILAF: Icomputerization of bilingual editorial dictionaries African language - French - 3k€ FABIOLE : Reliability in voice biometrics - ANR 78k€ FORMICAE: Ant colony algorithm for multilingual lexical disambiguation - UPMF 8k€ INNOVA-Langues : Innovation and transformation of language teaching-learning practices in higher education - ANR-IDEFI INTERABOT : Natural interactions with a companion robot - CDC ( Caisse des Dépôts et consignations) 267k€ Jibiki : Platform for collaborative development of multilingual dictionaries (non-funded team project) KEHATH : Advanced Quality Methods for Post-Edition of Machine Translation - ANR 2014 - 132k€ KINEPHONES: PEPS: Joint Call for Interdisciplinary Project COMUE de Grenoble – CNRS 2014 LexALP: Legal Language Harmonisation System for Environment and Spatial Planning within the Multilingual Alps MISTRAL: Plateforme opensource pour la biométrie multimodale MOTAMOT: Development of a multilingual lexical system through the construction of bilingual dictionaries targeting low-computer languages in South-East Asia Motikz: Word game on iphone OMNIA: Papillon: Towards High Quality Multilingual Lexical Database for Human and Machines PEA-TRAD: Translation for literature review assistance - PEA DGA 50k€ PI: Spoken Language Technologies for PI languages - ANR 135k€ QCOMPERE: Quaero Consortium for Multimodal People Recognition - ANR 81k€ RESIDE-HIS et DESDHIS: Reconnaissance de Signaux de Détresse dans l\u0026rsquo;Habitat Intelligent Santé (Recognition of Distress Signals in Healthy Intelligent Housing) (2000-2008) Sectra_W : Web-based translation corpus operating system: import/export, presentations and comparisons, use of MT systems, collaborative evaluation and post-editing (non-funded team project) SWEET-HOME: Home automation system for home assistance - ANR 291k€ (2009-2013) TELMA: Téléphonie à l\u0026rsquo;usage des malentendants (Telephony for the hearing impaired) TRAOUIERO: TRAduction : Outils Unifiés, Intégrables, Embarquables, et Ressources Opérationnelles (Translation: Unified, Integrable, Embeddable, and Operational Resources Tools) - ANR 358k€ TRIDAN : Translation and Recognition of Images of Digitized Arabic Documents - Rapid DGA - 65k€ UNL : Universal Networking Language (non-funded team project) VideoSense: Multimodal recognition of enriched concepts (static, dynamic, emotional) in multilingual videos through pivotal languages. - ANR 146 k€ (2010-2013) ","externalUrl":null,"permalink":"/en/projects/","section":"Projects","summary":"","title":"Projects","type":"projects"},{"content":"Team publications, updated daily from the GETALP HAL collection.\n","externalUrl":null,"permalink":"/en/publications/","section":"Publications","summary":"","title":"Publications","type":"publications"},{"content":" ANR CONTINT 2011-2014 Description: The QCOMPERE project aims to form a consortium to participate in the multimedia challenge REPERE. The QUAERO program, started in May 2008, is composed of French and German research organizations and the objective of the QUAERO CTC project is to improve the state of the art in structuring and automatic indexing of multimedia documents. Several of its tasks are directly linked to the REPERE call for projects, and each of its partners, in its respective field, has shown technological expertise of the highest level. Participation in REPERE is therefore a natural extension of the work planned in QUAERO. Participants: L. Besacier, G. Quénot Partners: LIMSI, LIG, INRIA LEAR, GREYC, YACAST, VECSYS Research ","externalUrl":null,"permalink":"/en/projects/qcompere/","section":"Projects","summary":"","title":"QCOMPERE : Quaero Consortium for Multimodal People Recognition","type":"projects"},{"content":" Research projects on Intelligent Housing for Health carried out in collaboration between the GETALP teams of the LIG laboratory (formerly the GEOD team of the CLIPS laboratory) and AFIRM of the TIMC-IMAG laboratory between 2000 and 2008. The funding that has enabled progress to be made on this theme is as follows:\nIMAG RESIDE-HIS project (IMAG 2000-2002) ; DESDHIS project of the ACI Santé du Ministère de la Recherche (2002-2004). Theses by Dan Istrate and Gilles Virone ; Departmental Award (Anthony Fleury Doctoral Award) (2004-2008). Thesis defended in 2008. These projects having been completed, this page has not been modified since December 2009.\nThe GETALP team contributed its skills for sound and speech recognition, the AFIRM team contributed its skills on medical sensors and end-user evaluation. This is a long-term research theme that aims to help people, usually elderly or struggling alone at home.\nKeywords: Intelligent Habitat, Smart Home, distress keywords detection, speech recognition, everyday sound recognition, ADL (activities of daily living) HIS apartment of the AFIRM team: map and views Audiovisual presentation of the project Your browser cannot play this video. Download video.\nProject presentation (2 min) Your browser cannot play this video. Download video.\nFrench version Participants: Michel VACHER (GEOD then GETALP) Jean-François SERIGNAT (Former GEOD Team Leader) Norbert Noury (AFIRM team leader over the duration of these projects) Vincent RIALLE (AFIRM) People who have made a significant contribution to these projects: Stéphane CHAILLOL (GEOD) Anthony FLEURY (AFIRM) Nicolas GAC (GEOD) Hubert GLASSON (GETALP) Noë GUIRAND (GETALP) Dan ISTRATE (GEOD) Pelayo Menendez-Garcia (GEOD) François PORTET (GETALP) Gilles VIRONE (AFIRM) Sound and multimodal corpus # Corpus HIS :corpus recorded at the end of the RESIDE-HIS and DESDHIS projects (thesis Anthony Fleury) in the HIS apartment of the TIMC-AFIRM team, participants play scenarios involving ADLs (Activity of Daily Living - ADL).\nRealization # Sound analysis software was developed and implemented in the HIS (Intelligent Housing for Health) of the AFIRM team.\nMain features of AUDITHIS software :\nDebian/LINUX operating system Multichannel acquisition card National Instrument pilote NiDAQmx compatible LINUX interruption if 1/2 full buffer 8 possible channels (microphones) 2 applications running in parallel on a single PC sound analysis - several threa automatic speech recognition (ASR Raphael with Janus recognition engine) GTK+ graphics library functions and libraries developed in C or C+ Significant publications # [1] A. Fleury, M. Vacher, N. Noury, SVM-Based Multi-Modal Classification of Activities of Daily Living in Health Smart Homes: Sensors, Algorithms and First Experimental Results, IEEE Transactions on Information Technology in Biomedicine, vol. 14(2), pp. 274 - 283, 2010. [2] M. Vacher, F. Portet, A. Fleury, N. Noury, **Development of Audio Sensing Technology for Ambient Assisted Living: Applications and Challenges,**International Journal of E-Health and Medical Communications (IJEHMC), IGI, 2011, 2 (1), pp.35-54. [3] A. Fleury, N. Noury, M. Vacher, H. Glasson and J.-F. Serignat, Sound and Speech Detection and Classification in a Health Smart Home, 30th IEEE EMBS Annual International Conference, \u0026ldquo;Personalized Healthcare through Technology\u0026rdquo;, Vancouver, British Columbia, Canada, Aug20-24, 2008, pp. 4644-4647.Advances in Spoken Language Technology,Iasi, Romania, 10-12 May, 2007, pp. 135-146, ISBN: 978-973-27-1516-1. [4] Michel Vacher, Anthony Fleury, Jean-François Serignat, Norbert Noury and Hubert Glasson, Preliminary Evaluation of Speech/Sound Recognition for Telemedicine Application in a Real Environment, INTERSPEECH, Brisbane, Australia, Sep 22-26, 2008, pp. 496-499. [5] A. Fleury, M. Vacher, H. Glasson, J.-F. Serignat and N. Noury, Data Fusion in Health Smart Home: Preliminary Individual Evaluation of Two Families of Sensors, ISG 2008, the 6th International Conference of the International Society for Gerontechnology, Pisa, Italy, June 4-7 , 2008, pp. 6p. [6] M. Vacher, J.F.Serignat, S. Chaillol, Sound Classification in a Smart Room Environment: an Approach using GMM and HMM Methods, the 4th Conference on Speech Technology and Human-Computer Dialogue, Advances in Spoken Language Technology, Iasi, Romania, 10-12 May, 2007, pp. 135-146, ISBN: 978-973-27-1516-1. [7] D. Istrate, M. Vacher, E. Castelli, L. Besacier, J.-F. Serignat,Information Extraction From Sound for Medical Telemonitoring, Information Technology in Biomedicine, IEEE Transactions on, April 2006, vol. 10, issue 2, pp. 264-274, ISSN: 1089-7771. [8] M. Vacher, J.-F. Serignat, S. Chaillol, D. Istrate and V. Popescu, Speech and Sound Use in a Remote Monitoring System for Health Care, Lecture Notes in Computer Science, Artificial Intelligence, Text Speech and Dialogue, vol. 4188/2006, 2006, pp. 711-718, ISBN: 978-3-540-39090-9. [9] M. Vacher, D. Istrate, J.F. Serignat and N. Gac, Detection and Speech/Sound Segmentation in a Smart Room Environment, presented at The 3rd Conference on Speech Technology and Human-Computer Dialogue (Sped 2005), Cluj-Napoca, Romania, May 13-14, 2005, pp. 37-48. [10] M. Vacher, D. Istrate and J.F. Serignat, Sound Detection and Classification through Transient Models using Wavelet Coefficient Trees, presented at 12th European Signal Processing Conference (EUSIPCO 2004), Vienna, Austria, September 6-10, pp. 1171-1174. ","externalUrl":null,"permalink":"/en/projects/reside-his/","section":"Projects","summary":"","title":"RESIDE-HIS and DESDHIS : Recognition of Distress Signals in Healthy Intelligent Housing","type":"projects"},{"content":"Publications :\nGazePlay-fr-20180408\n","externalUrl":null,"permalink":"/en/projects/disability-study-group/","section":"Projects","summary":"","title":"Study group for automatic language and speech processing for people with disabilities","type":"projects"},{"content":" ANR Project Description: This project, stemming from an INPG BQR, aims at the study and algorithmic development of original audiovisual functionalities for the use of hearing impaired people, and at the feasibility study of their integration in an autonomous telephone telecommunication terminal. The specific technical objective of the project is to exploit the visual modality of speech, on the one hand to improve speech sound denoising techniques (minimisation of environmental noise allowing a better exploitation of the auditory remains of the hearing impaired), and on the other hand, by implementing analysis/synthesis techniques for lip reading and gestures of the French Spoken Language (LPC). The LIG\u0026rsquo;s contribution to this project concerns the automatic recognition of speech and multimodal gestures. In connection with this project, L. Besacier is co-supervising a thesis student at the GIPSA laboratory, Nourredine Aboutabit. ","externalUrl":null,"permalink":"/en/projects/telma/","section":"Projects","summary":"","title":"TELMA : Telephony for the hearing impaired","type":"projects"},{"content":"List carried over from the former website, last updated in January 2021.\nStudents enrolled in an doctoral school of the doctoral college of U. de Grenoble # Houssein AHMED ASSOWE (Dir. H. Blanchon) Construction de ressources linguistiques pour l\u0026rsquo;informatisation et la traduction automatique statistique d\u0026rsquo;une langue peu dotée : Le cas du somali.\nSina ALISAMIR (Dir. L. Besacier \u0026amp; F. Ringeval) Weakly supervised learning of human affective behaviors from multimodal interactions with a chatbot.\nAmbre DAVAT (Dir. G. Feng \u0026amp; V. Aubergé) Vers un modèle du «toucher vocal» pour la communication ubiquïte\nThierry DESOT (Dir. M. Vacher \u0026amp; F. Portet) Système personnalisé de commande vocale en contexte\nMaha ELBAYAD (Dir. J. Verbeek \u0026amp; L. Besacier) Cross-lingual image captioning using recurrent neural networks and attention models.\nJibril FREJ (Dir. J-P Chevallet \u0026amp; D. Schwab) Combinaison de connaissances discrètes et continues pour l\u0026rsquo;accès à de l\u0026rsquo;information par le dialogue.\nQianqian FU (Dir. R. Ronfard \u0026amp; B. Lecouteux) Montage automatique de pièces de théâtre, basé sur des modèles informatiques de montage d\u0026rsquo;”idiomes” et l’analyse automatique du discours et du mouvement des acteurs.\nMahault GARNERIN (Dir. L. Besacier \u0026amp; S. Rossato) Gender issues and biases in spoken language technologies\nWilliam HAVARD (Dir. L. Besacier et J-P Chevrot) Méthodes non supervisées sur corpus denses pour analyser l’émergence du lexique en contexte\nDamien HANSEN (Dir. H. Blanchon \u0026amp; E. Esperança-Rodier \u0026amp; V. Bada Université de Liège) Utilisation des nouvelles technologies en traduction littéraire\nEmmanuelle KELODJOUE (Dir Jérôme Goulian \u0026amp; Didier Schwab) Extraction d’opinions pour l’aide à l’analyse multicritère de corpus oraux transcrits\nAli Can KOCABIYIKOGLU (Dir. H. Blanchon \u0026amp; F. Portet) Saisie de prescriptions médicales en langage naturel sur terminaux mobiles\nEric LE FERRAND (Dir. L. Besacier \u0026amp; S. Bird) Computational language documentation of aboriginal languages\nKy Trung NGUYEN (Dir. C. Garbay \u0026amp; F. Portet)Story Generation from Smart Phone Data - A script Approach\nManh Ha NGUYEN (Dir. L. Besacier \u0026amp; Y. Estève) End-to-end speech translation\nLoïc VIAL (Dir. B. Lecouteux \u0026amp; D. Schwab) Modèles joints de clarification de texte et de traduction automatique statistique.\nMarcely ZANON-BOITO (Dir. L. Besacier \u0026amp; A. Villavicencio Univ. Essex, UK) Computational Language Documentation\nTheses recently defended (link to online manuscripts on HAL)\nClick here to view and/or download all recent GETALP theses.\n","externalUrl":null,"permalink":"/en/theses-in-progress/","section":"GETALP","summary":"","title":"Theses in progress","type":"page"},{"content":" ANR: \u0026ldquo;Emergence of products, technologies or services with high value potential\u0026rdquo;, 2010 Call Keywords: operationalization, MT systems, language resources, MT services Description: Traouiero is an Emergence project funded by the ANR over 2 years: 2011 and 2012. The Traouiero project aims to enable the operationalization of software tools and linguistic techniques and resources developed so far by GETALP as operational prototypes. The two partners of this project are our research team and the valorisation structure of UJF, Floralis. In the context of Web 2.0, GETALP proposed replacing the traditional paradigm of translation, a product that is expensive once delivered, with that of multilingual access, a low-cost service in the form of a subscription, allowing the collaborative increase in the quality of \u0026ldquo;pre-translations\u0026rdquo; produced automatically. A young company was created around this concept, AXiMAG, BU (Business Unit) of Floralis (UJF-filiale) from September 2009 to September 2010, and in the process of being set up as a SAS. AXiMAG is beginning to add value to GETALP\u0026rsquo;s research products and is in rapid need of \u0026ldquo;operationalized\u0026rdquo; versions and \u0026ldquo;scaled-up\u0026rdquo; resources. On the other hand, GETALP is working on multimodal phrase book MT, on smartphones, for poorly endowed language pairs, and has anticipated the possibility of making MT embedded in the Ariane-Y project. This project has 3 main parts, named Ariane++, iMAG++, and LING++, which each have a page in this site. At the beginning of the project, our research team has products or resources resulting from work carried out for a long time, and others developed in recent theses for each of these parts. There is currently a unique opportunity to operationalize the whole, by calling upon both highly experienced staff and 3 postdoctoral fellows very interested in the operationalization of the software created for their theses. Each of these 3 parts is accompanied by the partner of valorization, Floralis, which deals with the aspects related to the intellectual property, and the valorization. Participants: Christian Boitet, Daniel Baud, Valérie Bellynck, Laurent Besacier, Hervé Blanchon, Francis Brunet-Manquat, Stéphane Donnet, Jean-Claude Durand, Emmanuelle Esperança-Rodier, Georges Fafiotte, Jérôme Goulian, Jean-Philippe Guilbaud, Mathieu Mangeot, Didier Schwab, Gilles Sérasset, David Cattaneo, Sylviane Chappuy, Mohammad Daoud, Achille Falaise, Ruslan Kalitvianski, Carlos Ramisch, Huyng Vo-Trung, Lingxiao Wang, Ying Zhang, Erwan Agaesse, Marwen Azouzi, Etienne Blanc, Abbas Malik, Hong-Thai Nguyen Partners: Floralis ","externalUrl":null,"permalink":"/en/projects/traouiero/","section":"Projects","summary":"","title":"Traouiero : TRAduction : Outils Unifiés, Intégrables, Embarquables, et Ressources Opérationnelles (Translation : Unified, Integrable, Embeddable, and Operational Resources Tools)","type":"projects"},{"content":"Page carried over from the former website, last updated in May 2018.\nThere are lists for research communities, specialized by theme, in particular:\nhttps://wwwsecu.irit.fr/listes/info/bull-i3 GDR I3 Newsletter Subscribers List (all computer researchers)\nhttp://liste.cines.fr//arc/ln List ln (Natural Language), French-speaking\na lot of archives on: http://listserv.linguistlist.org/mailman/listinfo\nhttp://fr.groups.yahoo.com/group/eureka/?v=1\u0026amp;t=search\u0026amp;ch=web\u0026amp;pub=groups\u0026amp;sec=group\u0026amp;slk=3 Eurêka is a mailing list of language resources in all their forms \u0026ndash;\u0026gt; no longer seems active in recent years\n","externalUrl":null,"permalink":"/en/nlp-news/","section":"GETALP","summary":"","title":"Where to follow TAL news?","type":"page"}]