Despite the impressive capabilities of conversational AI (ConvAI) systems based on large language models (LLMs) such as OpenAI GPT series, contextual understanding remains siginificant challenge for these models. Utterances convey more information than just their literal meanings; both semantics and context play a significant role. To truly understand the context, a reasoning process is involved by guessing information that is not explicitly stated in the utterances. Semantically correct responses of ConvAI systems cannot be guaranteed as the desired behavior if the context is not well-reflected.

In this thesis, we explore the contextual understanding of ConvAI systems. Firstly, we highlight the important aspects of reasoning for contextual understanding and the challenges it poses due...[ Read more ]

Despite the impressive capabilities of conversational AI (ConvAI) systems based on large language models (LLMs) such as OpenAI GPT series, contextual understanding remains siginificant challenge for these models. Utterances convey more information than just their literal meanings; both semantics and context play a significant role. To truly understand the context, a reasoning process is involved by guessing information that is not explicitly stated in the utterances. Semantically correct responses of ConvAI systems cannot be guaranteed as the desired behavior if the context is not well-reflected.

In this thesis, we explore the contextual understanding of ConvAI systems. Firstly, we highlight the important aspects of reasoning for contextual understanding and the challenges it poses due to the semantic information gap. Here, semantic information refers to the amount of detail provided in a proposition. For example, """"TI will not be in the lab this Wednesday because I am leaving for a conference"""" has more semantic information than """"TI will not be in the lab this Wednesday."""" Bridging the semantic information gap is necessary for the ConvAI systems to correctly interpret the context, as a user may not provide all intentions explicitly. To address this challenge, we examine how ConvAI systems can reason about the user's motivations or prerequisites. We propose a contrastive learning approach to supplement the semantic information gap and improve the reasoning ability during dialogue.

Tracking step-by-step logical processes is a challenge for current LLMs, which can lead to in-coherent responses. To address this, we investigate coreference resolution in dialogue, particularly in conversational question answering. Conversational question answering is more challenging than conventional single-turn question answering because questions may contain implicit expressions with anaphora or ellipsis to avoid redundancy. We propose a reinforcement learning approach to rewrite questions in a self-contained format by resolving coreferences from dialogue history.

Finally, we study the topic drift of a dialogue and how to refer to relevant knowledge for generating on-topic, informative responses. We describe guiding topics in an open-domain chitchat setting, where the system must provide relevant knowledge based on the dialogue context. A sensitive component is necessary to promptly respond to shifts in topics, as the appropriate knowledge to supply changes turn by turn. To address this, we introduce a dialogue system equipped with topic modeling that reacts to topic drifts and guides knowledge stored in the parameters of the models.