LAUSR

The semantic gap between keyword queries and search intents behind them motivates intensive studies on keyword query interpretation, which aims to interpret a keyword query to structured queries (a.k.a. patterns) representing most possibly relevant search intents. However, there still lacks of study on an important issue: how to guarantee the patterns are “reliable”, which means the structured queries can be evaluated as really existing results. In this paper, we regard the reliability as a new metric for ranking patterns, and present a keyword query interpretation approach to find both reliable and relevant pattern trees on an arbitrary summary graph of underlying data. Specifically, we first propose a reliability estimation model to measure how possibly a pattern tree can be evaluated as a nonempty result set by statistics under reasonable assumptions. Second, we develop constrained top-$k$k search algorithms that guarantee to return the optimal pattern trees for a specific keyword query. Moreover, to improve the efficiency of online search, we also design elaborate indexes, search heuristics and pruning strategies. Lastly, we perform comprehensive experiments on two real-world datasets, DBpedia and Yago, with both QALD-9 queries and random queries. The observations indicate our approach improves the accuracy and overall quality of top-$k$k results significantly.