Local assembly minutes record all the statements made in assemblies. Studies have been conducted to analyze local issues and to clarify the efforts and political stance of the assembly members based on the statements in the minutes. Previous studies have used word-based methods such as TF-IDF, which cannot consider multi-word phrases or contexts. In this paper, we construct a BERT-based classifier that predicts the speaker of each statement in the minutes, and calculate the contribution of each token to the prediction by SHAP. We then extract characteristic expressions of political interest from the statements in the minutes on a clause-by-clause basis according to the SHAP values, and analyze the experimental results. By taking into account the dependency relations between the clauses, the context of the extracted expressions can be presented. The analysis results show that our method can extract more characteristic expressions that reveal the political interests of the speaker than TF-IDF. In addition, we confirmed that this method can extract the speaker’s unique phrases and habits of saying, which are difficult to extract with TF-IDF.