Enhancing Dialogue Systems with Discourse-Level Understanding Using Deep Canonical Correlation Analysis

Dialogue systems, such as chatbots or virtual assistants, have m ade substantial progress in generating contextually appropriate responses. How ever, these systems face a persistent challenge in maintaining coherence and releva nce across multiple turns in longer conversations. This is especially difficult when th e context becomes complex, with numerous topics, nuanced reference s, or shifting conversational goals. With the objective of enhanced language mo deling, such models often struggle to effectively utilize the entire discourse histo ry, leading to responses that may be locally appropriate but globally inconsistent o r irrelevant [8] The core issue is how dialogue systems manage and interpret discour se history. Current models typically rely on the immediate context (e.g., th e last few utterances) to generate responses, which can lead to a loss of imp ortant information from earlier in the conversation. This limitation becomes more pro nounced 1 in longer dialogues, where the context is spread across many turns and may involve intricate dependencies between utterances.