Review Search Interface Based on Search Result Summarization Using Large Language Model

Abstract

For users of e–commerce sites, it is important but difficult to find the desired information from large numbers of reviews efficiently. In this paper, we propose a review search interface that provides users with summaries of reviews as search results. This interface uses large language models (LLMs) to vectorize queries and review sentences and to generate summaries. In this interface, when a user selects a phrase in a review as a query, the query is vectorized using a LLM. Then, the interface searches for reviews based on their similarity to the query vector. Finally, it summarizes the searched reviews using a LLM, and presents the summary as the search results to the user. By using a LLM, the proposed interface can be used for reviews in any domain. We conducted user experiments to quantitatively compare the proposed method with string matching methods. We evaluated the proposed method by taking a questionnaire with seven items, including comprehensiveness of viewpoints and opinions, and readability of search results. As a result, the proposed method was superior in terms of the readability of search results.

This work was supported by JSPS KAKENHI Grant Numbers JP24K03228, JP21H03775, JP22H03905.