Topic associations in document models for ad hoc information retrieval

Abstract

The integration of topic models into ad hoc retrieval has been studied by many researchers in the past decade and has achieved improved effectiveness and efficiency under the language modeling framework. However, the relationships between topics (i.e., topic associations) are rarely explored in information retrieval (IR). Topic associations can potentially improve the performance of IR systems. For example, when searching for "viral marketing", one might also be interested in seeing the results related with "word-of-mouth" or "social media". We build a topic-cluster-based document model to incorporate topic associations into the Latent Dirichlet Allocation (LDA) topic model. To discover the relationships between topics, we propose two clustering approaches based on topic co-occurrence and semantic similarity. To evaluate the performance of our proposed model, we plan to conduct a user study on ISTopic, an intelligent literature search tool for the IS community, and compare it with the query likelihood model and LDA-based model in terms of effectiveness, efficiency, and satisfaction, which are frequently adopted in user studies as measures for usability of IR systems. In addition, we plan to evaluate it on several TREC ad hoc test collections using mean average precision as the quantitative measure following the IR literature.