A multi-projection recurrent model for hypernym detection and discovery

Xuefeng Zhang, Junfan Chen, Zheyan Luo, Yuhang Bai, Chunming Hu, Richong Zhang

Abstract: Hypernym detection and discovery are fundamental tasks in natural language processing. The former task aims to identify all possible hypernyms of a given hyponym term, whereas the latter attempts to determine whether the given two terms hold a hypernymy relation or not. Existing research on hypernym detection and discovery tasks projects a term into various semantic spaces with single mapping functions. Despite their success, these methods may not be adequate in capturing complex semantic relevance between hyponym/hypernymy pairs in two aspects. First, they may fall short in modeling the hierarchical structure in the hypernymy relations, which may help them learn better term representations. Second, the polysemy phenomenon that hypernyms may express distinct senses is understudied. In this paper, we propose a Multi-Projection Recurrent model (MPR) to simultaneously capture the hierarchical relationships between terms and deal with diverse senses caused by the polysemy phenomenon. Specifically, we build a multi-projection mapping block to deal with the polysemy phenomenon, which learns various word senses by multiple projections. Besides, we adopt a hierarchy-aware recurrent block with the recurrent operation followed by a multi-hop aggregation module to capture the hierarchical structure of hypernym relations. Experiments on 11 benchmark datasets in various task settings illustrate that our multi-projection recurrent model outperforms the baselines. The experimental analysis and case study demonstrate that our multi-projection module and the recurrent structure are effective for hypernym detection and discovery tasks.