Toward Intelligent Recognition of Ecological Discourse in English-Language Teaching: Symmetric–Asymmetric Multi-Semantic Fusion with Relational-Graph Attention for Aspect-Based Sentiment Classification

Aspect-based sentiment classification associates each target aspect with relevant polarity evidence in sentences containing competing opinion cues and syntactic dependencies. We present two architectures extending aspect-oriented relational-graph attention with part-of-speech, dependency-relation, and dependency-head information. MSR-GAT+BERT combines contextual and syntactic features before graph-conditioned interaction. GMSR-GAT+BERT uses position-aware POS self-attention, separate contextual interaction paths, a BERT [CLS] representation, and shared-parameter gated fusion. Experimental results show that GMSR-GAT+BERT achieves Accuracy/F1 scores of 91.76%/87.57%, 85.80%/72.06%, and 70.00%/71.15% on the Restaurant, Laptop, and Twitter benchmarks, respectively. The comparisons are interpreted descriptively: several tabulated Accuracy percentages require careful treatment of their evaluation denominators, and small differences do not establish statistical significance. The contribution lies in the architectural integration and comparative analysis of complementary graph and auxiliary semantic streams. The approach also suggests a possible computational direction for identifying aspect-specific evaluations in ecological discourse used in English-language teaching; its empirical scope here remains the Restaurant, Laptop, and Twitter benchmarks.