Modeling Semantic and Behavioral Relations for Query Suggestion

Abstract

Query suggestion helps users to precisely express their search intents. The state-of-the-art approaches make great progress on high-frequency queries via click graphs. However, due to query ambiguity and click-through data sparseness, these approaches are limited to reality applications. To address the issue, this paper models both semantic and behavioral relations on hybrid bipartite graphs from click logs. Firstly, to overcome the sparseness, a semantic relation graph is established by multiple morphemes (queries, keywords, phrases and entities), independently of clicks. And then semantic relations between queries and other morphemes on the graph are used to find similar queries for query description. Secondly, in order to find related queries for multiple user intents, a behavioral relation graph is constructed by three kinds of user behaviors. Global clicks between queries and URLs display multiple intents by all users; local clicks imply personal preference by a single user; and query formulations in a session represent relations between queries. Finally, two hybrid methods are proposed to combine both semantic and behavioral relations to suggest related queries. To illustrate our methods, we employ the AOL query log data for query suggestion tasks. Experimental results demonstrate that more than 46.5% of queries get improved suggestions compared with the baseline click models.