Toxicity evaluation in French audio data with a focus on LGBTphobia
- Speaker
- Diandra Fabre (GETALP)
- Date
- Room
- Room 306, IMAG building
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.
To attend remotely, contact the seminar organiser, who will provide a video-conference link.
