Computational Models for Chinese Selectional Preferences Induction

Selectional preference (SP) is an important kind of semantic knowledge. It can be used in various natural language processing tasks, including metaphor computing, lexicon building, syntactic structure disambiguation, word sense disambiguation, semantic role labeling, anaphora resolution, etc. This paper presents and compares two computational models for Chinese SP induction, a HowNet-based Selectional Association model (SA) and a Latent Dirichlet Allocation model (LDA). Pseudo-disambiguation experiments show that LDA model outperforms SA model in both coverage and accuracy. However, LDA model needs to improve the interpretability of its induced SPs compared with SA model.