Title-based product search - Exemplified in a Chinese E-commerce portal

Abstract

There are too many products in an on-line shopping website. We need to help buyers to find products they want in an efficient way. A keyword-based IR system seems suitable for searching products. Unfortunately, we observe from real world query logs and find that queries for product search are usually very short. What is worse, a document described a product may have lots of words of related products. It is hard for an IR system to distinguish representative terms from other noisy terms. Hence, we propose a supervised learning method to realize semantic types of each term in product document titles. Then we modify Language Model to improve the relevance of search results. Our methods have significant improvement in search result precision in real world document collection and query collections. © 2010 Springer-Verlag.