An Analysis of Native Japanese Verbs Representation by Word Embeddings

Abstract

This study describes an analysis of the actual situation in which the distance between representations is measured using natural language processing methods. The distance between representations and neighbor words of each representation were measured by calculating the word embeddings based on the co-occurrence frequencies of the examples in the corpus. As a result of the analysis, we clarified that there were words such as “naku (〈泣く〉, 〈鳴く〉, etc.)”, for which the distance between the representations was far apart and their neighbor words feature each representation, and also, there were words such as "wakaru (〈分かる〉,〈判る〉, etc.)", for which the distance between the representations was close and their neighbor words were difficult to explain the differences. In addition, we also measured the distance between the representations in the Heisei period and the early Showa period, demonstrating that the usage of these representations has changed over time.