Stance prediction with a relevance attribute to political issues in comparing the opinions of citizens and city councilors

Abstract

This study focuses on a method for differentiating between the stance of citizens and city councilors on political issues (i.e., in favor or against) and attempts to compare the arguments of both sides. We created a dataset by annotating citizen tweets and city council minutes with labels for four attributes: stance, usefulness, regional dependence, and relevance. We then fine-tuned pretrained large language model using this dataset to assign the attribute labels to a large quantity of unlabeled data automatically. We introduced multitask learning to train each attribute jointly with relevance to identify the clues by focusing on those sentences that were relevant to the political issues. Our prediction models are based on T5, a large language model suitable for multitask learning. We compared the results from our system with those that used BERT or RoBERTa. Our experimental results showed that the macro-F1-scores for stance were improved by 1.8% for citizen tweets and 1.7% for city council minutes with multitask learning. Using the fine-tuned model to analyze real opinion gaps, we found that although the vaccination regime was positively evaluated by city councilors in Fukuoka city, it was not rated very highly by citizens.