Interactive User Intent Modeling

Abstract

In information retrieval systems, users often have difficulties in forming precise queries to express their information need. One approach to express information need is to explore the information space by providing relevance feedback to recommended items. This feedback is then used to model user search intent. Studies have shown how retrieval performance could be improved by allowing users to give feedback to multiple items such as keywords and documents instead of keywords only. In this thesis, I extend an existing user model which uses document-level and keyword-level feedback to include session-level feedback, and study the usefulness of this extension. By conducting simulation studies in various settings, I investigate the effect of session-level feedback. Based on these simulation results, I conclude that additional session-feedback helps in finding relevant documents by improving F1-score. Results show that more the additional session-feedback, more the improvement in F1-score. However, trade-off of session-feedback instead of document and keyword feedback results in drop in document F1-score, therefore indicating that session-feedback is less informative than document and keyword feedback.