Is long context helpful for dialog response generation?
Abstract: Personalization has been a key challenge in building engaging conversational agents, necessitating models to effectively utilize long-range context to maintain coherence and consistency over extended interactions. In this work, we investigate the potential of large language models (LLMs) to generate coherent and personalized responses in long-term human-human conversations. We experiment with $\textit{fixed context}$ and $\textit{retrieval-based}$ approaches to use the dialogue history between two speakers. We evaluate our methods and perform analysis on four long-term conversational datasets. Our results indicate that including only a few preceding utterances is generally sufficient for response generation. Retrieval or more extended contexts from past dialogues provide minimal benefits for personalizing model responses. Further analysis of instances that benefited most from retrieval reveals that these cases typically involve either explicit references to previously shared information or scenarios requiring stylistic consistency, such as farewell messages.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: spoken dialogue systems, retrieval, conversational modeling, evaluation and metrics
Contribution Types: Model analysis & interpretability, NLP engineering experiment, Data analysis
Languages Studied: English
Submission Number: 4478
Loading