Information Retrieval: 26th China Conference, CCIR 2020, Xi'an, China, August 14-16, 2020, Proceedings

As an essential part in web search, search snippets usually provide result previews for users to either gather useful information or make click-through decisions. In complex search scenarios, users may need to submit multiple queries to search systems until their information needs are satisfied. As user intents tend to be ambiguous, incorporating contextual information for user modeling has been proved effective in many session-level tasks. Therefore, the generation of search snippets may also benefit from the integration of context information. However, to our best knowledge, most existing snippet generation methods ignore user interaction and focus merely on the query content. Whether it is useful of exploiting session contexts to improve search snippets still remains inscrutable. To this end, we propose a snippet generation model which considers session contexts. The proposed method utilizes the query sequence as well as users’ interaction behaviors within a session to model users’ session-level information needs. We also adopt practical log-based search data to evaluate the performance of the proposed method. Experiment results based on both expert annotation and user preference test show the effectiveness of considering contextual information in search snippet generation.