[3Xin4] Poster session 1

[3Xin4-20] Extracting Feature Expressions in Local Assembly Minutes Using SHAP with Hierarchical Classification Structure

〇Hokuto Ototake¹, Keiichi Takamaru², Yuzu Uchida³, Yasutomo Kimura⁴ (1.Fukuoka University, 2.Utsunomiya Kyowa University, 3.Hokkai-gakuen University, 4.Otaru University of Commerce)

Keywords:Local Assembly Minutes, Classification, SHAP

Characteristic expressions such as keywords in the utterances of local assembly minutes are useful for understanding the issues of the region and the speaker's arguments, as well as providing clues for finding dialects.
The authors constructed classifiers that estimate the region and speaker using the utterances in the minutes of local assembly minutes as input, and showed the possibility of extracting expressions that characterize the region and speaker by visualizing the tokens that contribute to the classification using SHapley Additive exPlanations (SHAP).
Utterances in local assembly minutes can be classified hierarchically in the order of prefecture, municipality, and speaker.
In this study, we focus on this hierarchical classification structure and attempt to extract characteristic expressions specific to a narrow region or speaker by visualizing the difference of SHAP values corresponding to each classifier.