Language Model-based Context Augmentation for World Knowledge Bases

Abstract

Lack of background knowledge about the everyday world is an obstacle on the way to simulate usual situations and their changes. In this paper we present a simple idea for extending common sense knowledge bases for Japanese language by using a language model. We investigate several semantic categories for which specific knowledge is collected with mask prediction functionality of BERT and the polarity calculation with both next sentence prediction and masking with lexicons. We describe the experimental results and analyze the discrepancies between human evaluators and utilized sentiment analyzer.