A Multi-View Fusion Framework with Syntax-Aware, Sentiment-Salient, and Quantitative Information for Financial Sentiment Analysis

Sentiment analysis in the financial domain remains a demanding challenge, shaped by dense, domain-specific language, complex syntactic forms, and the critical influence of numerical information. While pre-trained language models such as FinBERT have achieved strong performance, their sequential processing paradigm often overlooks the explicit structural relationships that shape financial narratives. Existing enhancements, such as syntax-aware graph networks, typically address only one structural dimension, leaving other crucial information streams untapped. In this paper, we challenge this single-view paradigm and propose a novel multi-view fusion framework that deconstructs and interprets financial text from three complementary perspectives simultaneously. First, we build a syntax-aware dependency graph to model the logical relationships between words. Second, we generate a sentiment-salient semantic field that dynamically highlights emotionally charged terms. Third, we introduce a dedicated numerical-awareness and quantitative enhancement module that extracts, standardizes, and encodes diverse numerical entities, allowing the model to ground sentiment in concrete figures. These three structured views — syntactic, semantic, and quantitative — are integrated through a sophisticated, multi-stage fusion mechanism, enabling them to synergistically inform the final sentiment prediction. Extensive experiments on two benchmark datasets, Financial PhraseBank, and FiQA, demonstrate that our framework significantly outperforms a wide range of strong and directly comparable baselines. Ablation studies further validate that each view contributes meaningfully to the final performance, confirming the efficacy of our multi-view approach in achieving a more robust and nuanced understanding of financial sentiment.