SSABE-TSCM: drift-aware and interpretable financial sentiment analysis for low-resource Bangla via adaptive semi-supervised and temporal contrastive modeling

Analyzing the tone of Bangla financial news is challenging because labeled data are scarce, the language is morphologically rich, and economic discourse shifts over time. We address these hurdles with a three-part framework. First, SSABE a S emi- S upervised A daptive B oosting E nsemble iteratively refines pseudo-labels, adjusts model weights by recent performance, and applies sector-aware voting to distill reliable labels from limited data. Second, the T emporal S entiment C ontrastive M odule ( TSCM ) aligns yearly embedding prototypes via contrastive loss, keeping the classifier robust against vocabulary drift and shifting economic regimes. Third, Temporal-SHAP yields token-level attributions that reveal how term importance changes across years and industries, thereby making the system transparent to analysts. Evaluated on a 5-year (2018–2023) Bangla financial news corpus spanning eight sectors, our pipeline attains a macro-F 1 of 0.782 and 91.4 % explanation fidelity surpassing fine-tuned transformer and self-training baselines by 6 %–12 % absolute. Performance remains stable when labels are scarce, sectors are imbalanced, or economic shocks such as the inflation and currency decline of 2023 occur. Moreover, yearly sentiment scores and Temporal-SHAP attributions track inflation and exchange-rate trends, confirming real-world relevance. The proposed framework offers a scalable, interpretable solution for monitoring emerging-market news, supporting regulators, policymakers, and investors who rely on trustworthy Bangla-language insights.