Visualizing Feature Expressions in Minutes of the National Diet and Local Assemblies

Abstract

This paper describes the visualization of features within the minutes of the Diet and local assemblies. Initially, we employ BERT to develop a binary classifier that determines whether an input sentence originates from a local assembly or the Diet. Subsequently, we visualize the segments of the sentence that the classifier identifies as cues for its inference using explanatory methods such as SHAP and Integrated Gradients. We then analyze the visualization results to explore the types of expressions present in sentence parts that acted as inference cues. Our findings reveal that vocabulary specific to local and national administration is prominently visualized in the minutes of local assemblies and the Diet, respectively. Moreover, we identify various expressions characteristic of both the Diet and local assemblies.