Synonym Discovery for Structured Entities on Heterogeneous Graphs

With the increasing use of entities in serving people's daily information needs, recognizing synonyms---different ways people refer to the same entity---has become a crucial task for many entity-leveraging applications. Previous works often take a "literal" view of the entity, i.e., its string name. In this work, we propose adopting a "structured" view of each entity by considering not only its string name, but also other important structured attributes. Unlike existing query log-based methods, we delve deeper to explore sub-queries, and exploit tailed synonyms and tailed web pages for harvesting more synonyms. A general, heterogeneous graph-based data model which encodes our problem insights is designed by capturing three key concepts (synonym candidate, web page and keyword) and different types of interactions between them. We cast the synonym discovery problem into a graph-based ranking problem and demonstrate the existence of a closed-form optimal solution for outputting entity synonym scores. Experiments on several real-life domains demonstrate the effectiveness of our proposed method.