↓ Aller au contenu

Séminaires d’équipe

Neural Natural Language Generation with Limited Annotated Data

Intervenant·e
Raheel Qader (LIG, GETALP)
Date

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