Go to ACL ARR 2025 May homepageCRyCHIc: Chat Reconstruction Using Tree in the Context of Human-AI Interaction
Abstract: The conversation list function is widely built into most of the popular Large-Language-Model-based (LLM-based) chat applications. However, it can be hard for the users of these applications to find the chat history they want in the conversation list. One crucial reason for this problem is that sometimes the users tend to talk about multiple topics within one conversation. From this insight, we discussed the benefits of performing a chat tree construction on the chat history and filtering the history according to the tree before sending the user prompt with the history to the LLMs. We believe both LLM performance and user experience can be improved by doing so. A tree-constructing framework named CRyCHIc is then developed to construct the conversation tree efficiently. To test the performance of our framework, we also provide a test dataset called WildChatTree. Our model reaches 68.4\% accuracy and 84.8\% recall with only around 0.7B parameters on this dataset, reaching a performance similar to DeepSeek-V3. Our study offers direction for the future advancement of efficient chat tree construction. We will publicly release our code and models upon acceptance.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: applications, grounded dialog, conversational modeling
Contribution Types: NLP engineering experiment, Approaches low compute settings-efficiency, Data resources
Languages Studied: English
Submission Number: 4303
Loading