User Intent Guided Query Reasoning Over Knowledge Graph

The task of query reasoning over knowledge graph (KG) aims to address user’s requests that involve constrained conditions and necessitate reasoning over paths from KG. However, queries are often ambiguous while incorrect query paths could be reasoned due to misinterpretation of user intent. Existing methods leverage KG structure to disambiguate queries but user often cannot describe their query conditions clearly. To address these challenges, we propose user intent-aware query reasoning (UiQR) over KG, by incorporating user generated data (UGD) to identify user intent. Specifically, we first propose the weighting function to evaluate entity relatedness implied in UGD and prove its submodularity theoretically. Accordingly, we propose a greedy algorithm to efficiently filter optimal candidates and present a variational Bayesian method to reason query paths. Finally, we rank answer entities upon the reasoned paths that are most similar to the query embedding. Extensive experiments on three common datasets show that our UiQR significantly improves the accuracy of query reasoning over KG.