Topic Shifts: Preserving Comprehension in Conversation

International audienceTopics play an important role in coherence in dialogue, as what is currently discussed constrains the possible contributions of the participants, and initiating a topic while the previous one is still under discussion may be confusing without appropriate signals. However, how to actually define the notion of topic is debated in linguistics and not sufficiently discussed in dialogue modelling.A precise description of topics and topic shifts in conversation would contribute to understanding what it is we perceive when we judge a sequence of utterances to be coherent. In particular, having more understanding of the notion of topic is important to account for language use which does not conform to our intuitions about topical coherence, such as can be found in interactions involving for example patients with schizophrenia. In patients interaction we often see difficulties to produce a coherent speech from the ordinary listener’s point of view. These difficulties can manifest at topic level where patients would use fewer markers to indicate a topic shift, or use unusual topical links such as sound similarity between two different concepts. Formalising topic transitions, and in particular more unusual ones, would thus contribute to identifying cues to support diagnosing schizophrenia.In this thesis we analyse several pieces of conversation containing topic shifts in order to understand the different mechanisms that license topic shifts in dialogue and the way participants acknowledge them. In particular, we investigate topic shifts based on extra-linguistic content and more unconventional topic shifts where speakers take advantage of word similarities to introduce a new topic with no semantic relation with the previous one. We apply two classical discourse modelling theories to our examples, Rhetorical Structure Theory and Segmented Discourse Representation Theory, and show their limits when it comes to modelling topics and topic shifts. In particular, the unique representation for a conversation that they produce does not make it possible to precisely understand the mechanisms that lead to a disruption. Hence, we use a more complex framework, Conversation Oriented Semantics, which is based on Type Theory with Records and enables us to represent each participant's representation of the conversation, as well as different forms of context. We suggest a way to add topics to these representations and account for topic shifts, including unconventional ones