Séminaire Image : « Coreference Resolution in French Media », Kirill Milintsevich

Nous aurons le plaisir d’écouter Kirill Milinstsevich, post-doctorant à l’Institut National de l’Audiovisuel (INA), Bry-sur-Marne.
Il donnera un séminaire IMAGE le jeudi 11 septembre 2025 à 14h en salle de séminaire F-200.

Titre : « Coreference Resolution in French Media »

Résumé : 
The Institut national de l’audiovisuel (INA) continuously records French TV and radio broadcasts, creating a vast and invaluable archive for research in the humanities and social sciences. This dataset is particularly useful for studying real-world, noisy conditions, such as Automatic Speech Recognition (ASR) transcriptions.
In my postdoctoral project, I evaluate both classic pretrained language models (PLMs) and large language models (LLMs) for automatic coreference resolution in ASR transcriptions of French broadcasts. Coreference resolution is a discourse processing task that identifies links between words in speech or text. For example, in the phrase « John told Peter he was wrong, » the pronoun « he » could refer to either « John » or « Peter, » depending on the context. Automatic systems must resolve such ambiguities, which is especially challenging in spoken language due to disfluencies and ASR errors.
In this talk, I will provide an overview of automatic coreference resolution, its unique challenges in French, and the impact of ASR on performance. I will also present empirical results evaluating how well modern NLP models can address this task.