Computational models of dialogue often rely on fine-tuned variants of Large Language Models (including extra symbols to denote e.g. turn boundaries or non-verbal signals). However, most current models assume the conversation is continuous, with time gaps between turns that typically last less than one second.
However, there are certain types of dialogues in which time gaps can be much larger. One notable instance is of course instant messaging, where responses can be sent hours or even days after the previous message. This is also the case for chat conversations in business settings (think for instance about insurance companies chatting with their customers for case-handling purposes).
Currently, those time gaps are simply ignored when training/fine-tuning an LLM for conversations. But it not unreasonable to expect that those time gaps have an influence on both the type of dialogue contributions (e.g. a question such as "How has your week been?" is more likely after a large time gap) and the linguistic form of those contributions (the use of pronouns and deictics may be more frequent for contributions that directly follow one another).
Can we reduce the model perplexity by explicitly including information about those time gaps between turns, for instance through the use of special symbols to denote time gaps? This thesis will investigate this question in more details.