{"id":8,"date":"2019-02-09T19:46:54","date_gmt":"2019-02-09T18:46:54","guid":{"rendered":"http:\/\/mlia.lip6.fr\/soulier\/?page_id=8"},"modified":"2023-08-01T10:02:51","modified_gmt":"2023-08-01T08:02:51","slug":"research","status":"publish","type":"page","link":"https:\/\/pages.isir.upmc.fr\/soulier\/research\/","title":{"rendered":"Research interests"},"content":{"rendered":"
 <\/div>\n\n\n

Deep learning in information retrieval (IR) and natural language processing (NLP)<\/h1>\n\n\n\n

<\/p>\n\n\n\n

My research is motivated by the proposal of new models based on deep learning for information retrieval and automatic natural language processing.
The common objective of these models is to process and access textual data. (Large) language models are at the core of my various projects, which have applications related to semantic learning, human-machine interaction (in information retrieval and robotics), and continual learning. Since 2018, I have been involved in two major research projects focusing on data-to-text generation and conversational information retrieval.<\/p>\n\n\n\n

Data-to-text generation<\/h2>\n\n\n\n

The objective is to generate textual descriptions of structured inputs (tables\/graphs\/…). This task is particularly interesting in the financial domain, sport journalism, or health since it allows to synthetize and reason over large set of structured data which might be hardly readable for humans.

\"\"

<\/p>\n\n\n\n

Conversational search<\/h2>\n\n\n\n

This work is supported by the ANR JCJC SESAMS<\/a>.
The objective is to support users’ search through interactive and proactive systems, anticipating their needs and guiding users for solving their task. Information need refinement\/understanding, belief tracker, dialog systems, and language generation are examples of tasks that can be addressed in this topic.

\"\"

<\/p>\n\n\n\n

Language models for robotics<\/h2>\n\n\n\n

One emerging hypothesis in robotics is that reinforcement learning algorithms aiming to predict robot actions might be enhanced by the semantics underlying language models. Our objective is to design hybrid models combining RL and LLM to generate instructions in natural language, aiming to guide the action prediction.

\"\"

<\/p>\n\n\n\n

Continual learning and domain adaptation<\/h2>\n\n\n\n

Language models and, more generally neural models, might suffer from catastrophic forgetting while being fine-tuned on additional data. Our objective is to leverage this limitation while maintaining the knowledge learned on previous tasks.

\"\"<\/p>\n\n\n\n