Toward Action Comprehension for Searching: Mining Actionable Intents in Query Entities

摘要

Understanding users' potential intents in very short queries is an important study in information retrieval. Most of the current studies focus on organizing search queries into a few predefined intent categories, leaving out many potentially informative actions. In this article we present a novel research for mining a ranked list of such informative actions for a query entity and explore a variety of search strategies for generating a pool of candidate actions. We propose three criteria, that is, significance, representativeness, and diverseness, for evaluating the informativeness of candidate actions and propose two action-mining algorithms for iteratively generating the ranked list of action verbs and verb modifiers respectively based on these criteria. A thorough experiment was performed based on the AM query entity data set, with the quantitative evaluations of generated actions according to the nERR metric and the nDCG metric. The experiment results suggest that our approach could generate both early-satisfying actionable intents to match the search users' aims, that is, the best nERR scores, and informative actions according to their informativeness, that is, the promising nDCG scores. In addition, our approach proved to be very stable even for the essentially most difficult entities. An analysis of the generated actions further indicates that emphasizing the diverseness priority in our action-mining algorithm could steadily improve the nERR and nDCG results. An interesting finding in our study is that the (S) search strategy that queries the Twitter Streaming API does not render as proper results as the other search strategies. This is probably because the (S) strategy has retrieved too much noise from the real-time Twitter stream, which increased the difficulty of selecting proper actions for our action-mining algorithm. As such sensitivity has inevitably limited the diverseness in our action-mining results, our future work will focus on the reasoning of more accurate relationships between entities and actions from the general action pools.