Contextual Parameters’ Impact on Sentiment Polarity Determination

This study proposes a deep learning-based sentiment analysis framework integrated with Halliday’s contextual theory, constructing a context-aware model by deconstructing three-dimensional parameters of field, tenor, and mode. We develop a Contextual Configuration Matrix to quantify domain-specific term density (NI), social power gradient (PG), and rhetorical structure (RST), while designing an attention mechanism for dynamic fusion of contextual features and semantic sentiment. Experiments on dual-domain datasets (2,800 medical consultations and 3,500 e-commerce reviews) demonstrate that introducing contextual parameters improves sentiment classification F1-scores by 14.2% (medical: 0.82 $\rightarrow 0.89$; ecommerce: $0.76 \rightarrow 0.84$) and effectively resolves $23.7 \%$ of domaindependent sentiment ambiguities (e.g., “flexible” shows $61 \%$ negative polarity in medical contexts vs. $89 \%$ positive in ecommerce). This work validates Systemic Functional Linguistics (SFL) as an interpretable pathway for contextual modeling in sentiment analysis.