As aspect-based sentiment analysis (ABSA) extends its application across various domains, effectively utilizing this technology in political opinion analysis poses significant challenges, particularly due to limited training data in sensitive areas. To address this, we constructed two datasets: a 7,334-sample three-class sentiment analysis dataset from political news via web scraping, and a high-quality dataset of 100 manually annotated news sentences designed for testing aspect-level sentiment computation. "This paper proposes a sentiment classification method based on the mT5 model, combined with Named Entity Recognition (NER) and Semantic Role Labeling (SRL). Experimental results show that our method performs comparably to existing approaches on the SemEval-2014 (Rest14) dataset, while achieving 84% accuracy in aspect entity recognition and 69% overall recognition accuracy on our custom dataset. These findings suggest that the proposed method can significantly enhance model performance in scenarios with limited and sensitive training data.

Political Opinion Analysis,mT5 ModelSemantic Role Labeling (SRL),mT5 Model,News Sentiment Analysis