Transformer-driven automated analysis of social media narrative structure: An exploration based on sentiment framing and thematic agenda

Abstract

With the rapid development of social media, narrative texts in public event scenarios have become important carriers of public opinion, making the automatic analysis of social media narrative structures increasingly crucial. Existing research on this task suffers from insufficient integration of multi-dimensional information such as sentiment, topic and time, and poor adaptability to complex scenarios like cross-events and noisy texts. To address these issues, this study proposes a sentiment-topic-temporal attention fusion model (ST-TAN), which takes RoBERTa as the basic semantic encoding module and integrates three core modules to realize joint modeling of sentiment and topic and capture temporal dependence of narrative units. Experimental results show that the ST-TAN model comprehensively outperforms four types of baseline models in narrative structure recognition, sentiment classification and topic classification tasks, with good cross-event generalization ability and noisy text robustness. This research enriches the theoretical connotation of social media narrative analysis and provides effective technical support for practical fields such as public event governance and public opinion monitoring. The study further incorporates a comprehensive discussion of ethical considerations, addressing user privacy, data anonymization, potential biases, and responsible use, thereby ensuring alignment with responsible innovation principles.