Extracting Feature Expressions in Local Assembly Minutes Using SHAP with BERT-Based Classifier

Abstract

Characteristic expressions such as keywords in the utterances of local assembly minutes are not only useful for understanding the issues of the region and the speaker's arguments, but also provide clues for finding dialects. In a classifier that estimates regions and speakers from utterances, tokens that contribute to classification may become expressions that characterize regions and speakers. In this study, we constructed a BERT-based classifier for local assembly minutes from all over Japan, and extracted tokens that contribute to classification based on SHapley Additive exPlanations (SHAP) as feature expressions. As a result of the experiment, the accuracy of the classification was about 50%. From the successfully classified utterances, place names, dialects, and political issues were extracted as region-specific expressions. In addition, we confirmed that it is possible to extract feature expressions consisting of multiple tokens with consideration of the context.