Acquiring Selectional Preferences for Knowledge Base Construction

Abstract

Selectional preference, or SP, is an important lexical knowledge that can be applied to many natural language processing tasks, like semantic error detection, metaphor detection, word sense disambiguation, syntactic parsing, semantic role labeling, and machine translation. This paper studies semantic class level SP acquisition for knowledge base construction. Firstly, the noun taxonomy of SKCC, a Semantic Knowledge-base of Contemporary Chinese, is adjusted for SP acquisition. Secondly, a MDL-based tree cut model is implemented. Thirdly, SP in SKCC is introduced as the source of gold standard test set to evaluate SP acquisition performance. Three kinds of predicate-argument relations are investigated in the experiments, including verb-object, verb-subject, and adjective-noun relations. For the verb-object relation, the top1 strict accuracy is 24.74% while the top3 relaxed accuracy reaches 75.26%.