Description
Date depot: 1 octobre 2020
Titre: Reinforcement learning models for search-oriented conversational systems
Directrice de thèse:
Laure SOULIER (ISIR (EDITE))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé:
In conversational systems, interactions engaged by the conversational system are inherently correlated with the quality of the outcomes. Indeed, while the goal of chit-chat bots is to provide discussions that are in the continuity of the conversation, task-oriented bots generally focus on identifying the right response with respect to a particular question (whether associated to a slot filling step or not) as done in ad-hoc retrieval. However, the multi-turn setting imposed by conversational systems offers the possibility to engage the system in a constructive interaction with the user to help the user accomplishing his/her task. In conversational systems, this feature is mainly addressed using decision-making processes based on reinforcement learning models, such as POMDP or deep reinforcement learning. These approaches are used to anticipate the next action of the user and accordingly moderate the system response. In the context of conversational search, the framework is different since it assumes that system-to-user interactions are specific to the IR purposes.
For instance, the search-oriented conversational system might ask the user to refine the query or to express his/her preferences between queries reformulated by the conversational system or between documents retrieved by the search engine. These bi-directional interactions within search-oriented conversational systems open a new perspective to user-driven system-mediated scenarios, as defined in the context of collaborative search.
Search-oriented conversational systems have led to exciting research work these last years: the definition of the new paradigm and its underlying challenges, interactive IR or query reformulation models to mimic the behavior of conversational systems, and more recently the introduction of reinforcement learning frameworks for ranking or interactive IR.
This thesis is in line with this last research direction and aims at exploiting machine learning techniques to the benefit of IR models.
Doctorant.e: Erbacher Pierre