Recommendation System based on Dialogue using Speaker Summary and Augmented Information

Abstract

Dialogue contains a wealth of information about the preferences and experiences of speakers. This information can be used to personalise and suggest advanced information in various systems, although it is generally underutilised. We propose the SumRec framework for dialogue-based recommendation, using information from speaker summaries and recommendation sentences. In this framework, a large language model (LLM) generates speaker summaries and item recommendation sentences to extract features of both the speaker and the item. The speaker summary focuses on the speaker's preferences and experiences, while the recommendation sentences describe the type of people who would prefer the item. The score estimator then uses this information to predict how much the speaker would like the item. Experimental results showed that SumRec outperformed the baseline on two datasets in different domains.