DASD: A Dual-Channel Appraisal-Guided Stance Detection Framework

 In the era of big data, social media platforms continuously generate massive volumes of heterogeneous textual data characterized by high velocity, emotional richness, and rapid topic evolution. Effectively analyzing such large-scale unstructured data has become a critical task in big data analytics, particularly for applications such as public opinion monitoring and digital content moderation. As a fundamental natural language processing task, stance detection plays a vital role in extracting opinion-oriented insights from large-scale social media streams. However, it faces significant challenges, including implicit semantic expressions, short text sparsity, and the intertwining of affective and stance signals. These challenges are further exacerbated in long-text contexts, where increased data length, dispersed linguistic cues, and complex contextual dependencies substantially complicate large-scale analytical modeling. To address these limitations, we propose a Dual-channel Appraisal-guided Stance Detection (DASD) Framework, grounded in appraisal theory. Our framework integrates three core components: (1) a dual-channel sentiment encoding module that concurrently captures contextual affective dependencies through the implicit encoding of pre-trained models and computes lexical polarity via lexicon-based explicit encoding, (2) an external knowledge injection module that augments semantic representation through domain-relevant textual resources and structured knowledge, and (3) a feature fusion stance classification module that fuses representations from the three appraisal dimensions—Affect, Judgment, and Appreciation—along with external knowledge to amplify stance-relevant affective signals. Comprehensive evaluation on the VAST (long-text) and SemEval-2016 (short-text) datasets demonstrates that the DASD framework achieves excellent performance across multiple metrics. Notably, on the VAST dataset, it significantly outperforms existing baseline models, highlighting its effectiveness for large-scale, complex text analytics in big data environments.