A product retrieval system robust to subjective queries

In recent years, electronic markets are increasing rapidly and attracting the attention of customers. In these sites, people search for products using retrieval systems. They, however, often cannot translate their subjective needs into keyword-based queries or adapt to the interfaces. In this paper, we describe a product retrieval system robust to subjective queries. Using a large amount of consumer reviews, the system allows users to input natural language queries and retrieves appropriate products even if the queries are highly subjective. To estimate the correspondence between a query and a review text, the system extracts 3-tuples consisting of a product name/category, its features, and the value from each text using rules based on syntactic patterns. It calculates each product scores based on correspondence rate of 3-tuples and presents ranked relevant products. In experimental results for a accommodation domain, it obtained higher average and total precision for 10 queries compared with a baseline that uses keyword based tf-idf method. Thus, we confirmed the effectiveness for subjective queries.