Visualization of the Gap Between the Stances of Citizens and City Councilors on Political Issues.

In local government, clarifying the gap between the political opinions of citizens and city councilors is important for reflecting the will of the people in politics. In this study, we focus on the difference between the stances (i.e., favor or against) of citizens and city councilors on political issues, and attempt to compare the arguments of both sides. Using a dataset of texts collected from citizen tweets and city council minutes, we performed a detailed analysis of the opinions expressed, based on the fine-tuning of pretrained language models. In particular, the model predicted the labels of four attributes: stance, usefulness, regional dependency, and relevance. In our experiments, we targeted two political issues: "eliminating nursery school waiting lists" and "attracting integrated resorts (IR)." However, the clues enabling the prediction of labels for the four attributes varied depending on the target. We, therefore, introduced a target-attention mechanism, which extracted helpful information from target sentences, to improve the prediction performance. In addition, we improved the prediction of attribute labels that considered relevance-related issues, by adopting multitask learning of relevance and other attributes. Our experimental results showed that the macro F1-scores for stance and regional-dependency attributes were improved by up to 1.0% and 3.7%, respectively, when using the target-attention mechanism, and by up to 7.2% for usefulness with multitask learning. Using the trained model to analyze real opinion gaps, we found that the citizens of Osaka City were relatively more supportive of attracting IR than the citizens of Yokohama City.