HCASO: Hybrid Cross-Lingual Attention with Contrastive Alignment and Optimized Lexical Fusion for Bilingual E-Commerce Sentiment Analysis

Shared-encoder multilingual models that confound typologically different language representations scarcely support cross-lingual sentiment analysis. This article introduces HCASO (Hybrid Cross-Lingual Attention with Contrastive Alignment and Optimized Lexical Fusion for Sentiment Analysis) as a unified bilingual model, which consists of four elements: (i) language-specific English and Chinese ELECTRA encoders, (ii) cross-lingual attention that allows structured inter-language communication, (iii) a symmetric InfoNCE contrastive loss in a 256-dimensional space that allows the languages to achieve clearer alignment, and (iv) a two-stage gated lexical fusion module that adaptively combines TF-IDF and sentiment lexicon features with contextual encoder representations via a learned per-dimension mixing coefficient. The model is trained using cross-entropy, MSE intensity regression, and the contrastive objectives. Post-hoc temperature scaling was used to calibrate it. Compared to standard baselines, HCASO achieves the strongest overall balance of performance, with ROC-AUC of 0.9361, 0.8623, and 0.8474 on binary, ternary, and quinary MARC English-Chinese sentiment classification, respectively. HCASO performs best on ordinal consistency (MSE = 0.0455, R² = 0.6356, Top-2 = 0.8211) across settings, with XLM-RoBERTa-Base being slightly more accurate in terms of raw accuracy in quinary classification. Ablation experiments with McNemar tests confirmed that all the components - dual encoders (p < 10⁻²⁶ across all settings), lexical fusion (p < 10⁻¹¹ in the ternary setting), and contrastive alignment (p = 0.028 in the quinary setting) each contribute significantly. HCASO is a statistically validated alternative to the shared-encoder paradigm, which scales at 214.4M parameters and 10.89 GMACs.