Common sense reasoning with predicate-argument PPMI third-order tensor

Abstract

In this research, we propose a method to acquire common sense data from a large-scale text corpus. The common sense data is calculated as predicate-argument positive pointwise mutual information (PPMI) third order tensor of a word. Specifically, we extract provisional predicates and arguments using the part-of-speech information of words, and a result of evaluating this PPMI model using the common sense reasoning benchmark set JCommonsenseQA, an accuracy of 66.0% was confirmed.