Scholarlink

In this study, we propose a method that considers the diversity of query expressions for searching food reviews. For example, if you search for reviews of udon noodles using the query “firmness”. In that case, it is easy to output reviews containing the expression “firmness”. On the other hand, we want to output reviews containing expressions like “elasticity” and “chewiness,” which do not match the exact query but have semantic similarity. Therefore, we propose two approaches using query expansion with ChatGPT, a large language model (LLM). Both approaches use the LLM to obtain several synonymous phrases for a query. The first method trains the BERT model for relevance judgment of reviews based on semantic similarity to the query. To create training data, collect candidate queries in advance and acquire their synonymous phrases. In the second approach, when a search query is input, the synonymous phrases of the query are obtained for on-demand query expansion. We implemented these approaches and compared them to methods such as BM25. The combined use of both proposal approaches showed better performance.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Query Expansion in Food Review Search with Synonymous Phrase Generation by LLM

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Arisa Ashizawa,
  • Ryota Mibayashi,
  • Hiroaki Ohshima

摘要

In this study, we propose a method that considers the diversity of query expressions for searching food reviews. For example, if you search for reviews of udon noodles using the query “firmness”. In that case, it is easy to output reviews containing the expression “firmness”. On the other hand, we want to output reviews containing expressions like “elasticity” and “chewiness,” which do not match the exact query but have semantic similarity. Therefore, we propose two approaches using query expansion with ChatGPT, a large language model (LLM). Both approaches use the LLM to obtain several synonymous phrases for a query. The first method trains the BERT model for relevance judgment of reviews based on semantic similarity to the query. To create training data, collect candidate queries in advance and acquire their synonymous phrases. In the second approach, when a search query is input, the synonymous phrases of the query are obtained for on-demand query expansion. We implemented these approaches and compared them to methods such as BM25. The combined use of both proposal approaches showed better performance.