Improving Search Effectiveness with Field-based Relevance Modeling

Binsheng Liu, Xiaolu Lu, Oren Kurland, J. Shane Culpepper

Abstract: Fields are a valuable auxiliary source of information in semi-structured HTML web documents. So, it is no surprise that ranking models have been designed to leverage this information to improve search effectiveness. We present the first (initial) study of utilizing field-based information in the relevance modeling framework. Fields play two different, and integrated, roles in our models: sources of information for inducing relevance models and units on which relevance models are applied for ranking. Our preliminary results suggest that field-based relevance modeling can improve precision at top ranks; specifically, to a greater extent than the commonly used BM25F and SDM-Fields field-based models. Further analysis shows that using field-based relevance models mainly improves the effectiveness of tail queries. Our findings suggest that using field-based information together with relevance modeling is a promising area of future exploration.