Determination of Power Harassment Expressions in Conversational Speech Using Natural Language Processing

Abstract

In recent years, the number of power harassment consultations is increasing, and power harassment with ambiguous criteria such as mental aggression is rampant. The purpose of this study is to notify the perpetrator when the likelihood of power harassment is judged to be high based on conversational speech. We attempt to use natural language processing to determine whether the target text constitutes power harassment, based on textual data on past precedents that have led to power harassment. The proposed method determines whether the target text constitutes power harassment or not by calculating the similarity (cos-similarity) between the target text and the text of the precedents and comparing it with a threshold value set through the experiment. The resemblance is calculated from a 768-dimensional feature vector obtained from each text’s Bidirectional Encoder Representation from Transformers (BERT). The morphological analyzer is Juman++ and the BERT Japanese Pre-trained Model is used as a pre-trained model. We used two types of surveys to determine thresholds and assess accuracy. In the experiment, we determine the threshold according to the questionnaire results and obtain a high discrimination rate, which shows that our method is effective.

Keywords: Natural Language Processing, Power Harassment, similarity

DOI: 10.54941/ahfe1002738