Taxonomy Completion via Triplet Matching Network

Automatically constructing taxonomy finds many applications in e-commerce and web search. One critical challenge is as data business scope grow real applications, new concepts are emerging needed to be added the existing taxonomy. Previous approaches focus on expansion, i.e. finding an appropriate hypernym concept from for a query concept. In this paper, we formulate task, “taxonomy completion”, by discovering both hyponym query. We propose Triplet Matching Network (TMN), find pairs given TMN consists of one primal scorer multiple auxiliary scorers. These scorers capture various fine-grained signals (e.g., or semantics), makes holistic prediction triplet based internal feature representations all Also, innovative channel-wise gating mechanism that retains task-specific information introduced further boost model performance. Experiments four real-world large-scale datasets show achieves best performance completion task previous expansion outperforming methods.