BanglaSentNet: An Explainable Hybrid Deep Learning Framework for Multi-Aspect Sentiment Analysis with Cross-Domain Transfer Learning

AI summary

Overview

Research area: Natural language processing (sentiment analysis) for a low-resource language, combined with explainable AI (XAI), ensemble deep learning, and cross-domain transfer learning. The application domain is Bangladeshi e-commerce review analytics.

Technical level: Advanced. The paper assumes familiarity with recurrent architectures (LSTM, BiLSTM, GRU), transformer models (BanglaBERT), embedding strategies (GloVe, FastText), ensemble weighting, SHAP/LIME attribution, and domain adaptation protocols.

Scope: The paper presents BanglaSentNet, an explainable hybrid deep learning framework for multi-aspect sentiment classification of Bangla e-commerce reviews, together with a new 8,755-review annotated dataset, an explainability suite, and cross-domain transfer learning experiments across four target domains.

What This Paper Is About

Bangla is spoken by more than 300 million people, yet category-level (multi-aspect) sentiment analysis for the language remains largely unexplored, and most existing Bangla work only predicts general polarity rather than opinions about specific product attributes such as quality, service, price, and decoration. The authors argue that progress is blocked by scarce annotated data, Bangla's morphological complexity, frequent code-mixing with English, informal writing, and performance degradation when a model trained on one product domain is applied to another. Their goal is a single framework that classifies sentiment across four aspects, explains why each prediction was made, and generalizes to new domains with little or no target-domain training data.

Key Contributions

1. A new annotated dataset: 8,755 manually annotated Bangla product reviews collected from major Bangladeshi e-commerce platforms, labeled with Positive, Negative, or Neutral sentiment across four aspects (Quality, Service, Price, Decoration).
2. An explainable hybrid ensemble: A dynamic weighted ensemble of LSTM, BiLSTM, GRU, and BanglaBERT components that combines static embeddings (GloVe, FastText) with contextual BanglaBERT embeddings.
3. An integrated explainability suite: SHAP-based feature attribution, attention mechanism visualization, and LIME, evaluated against human judges for interpretability and agreement.
4. Cross-domain transfer learning analysis: Zero-shot and few-shot protocols tested on four target domains (BanglaBook reviews, social media posts, general e-commerce reviews, news headlines), plus real-world deployment case studies for e-commerce platforms.

Main Findings

• Classification performance: BanglaSentNet achieves 85% accuracy and a 0.88 F1-score, which the authors report outperforms standalone deep learning models by 3-7% and traditional machine learning approaches substantially.
• Explainability quality: The combined explainability suite scores 9.4/10 for interpretability with 87.6% human evaluation agreement. Individually, SHAP scored 8.1 interpretability and 81.2% agreement, LIME 7.8 and 79.5%, and attention 7.3 and 76.3%; the original model without explanations scored 2.1 interpretability with no applicable agreement score.
• Attention patterns are aspect-consistent: For positive Quality sentiment, the highest attention weight (0.250) went to "অসাধারণ" (excellent), followed by "সুপার" (super) at 0.200 and "ভালো" (good) at 0.143. For negative Price sentiment, "বেশি" (expensive) and "দাম" (price/cost) received 0.213 and 0.200. For positive Service sentiment, "বিশ্বস্ত" (trustworthy) and "দর" (rate) received 0.233 and 0.250. For neutral Decoration sentiment, "সাজসজ্জা" (decoration) received 0.231 and "পরিবেশ" (environment) 0.204.
• Dataset composition by platform: Daraz contributed 6,521 reviews, Facebook 1,145, Rokomari 494, Shajgoj 155, and other platforms 440, totaling 8,755.
• Dataset composition by aspect: Quality had 4,000 reviews (2,400 positive, 1,200 negative, 400 neutral), Price 2,500 (1,500/900/100), Decoration 1,255 (750/400/105), and Service 1,000 (600/300/100).
• Aspect vocabulary overlap: Jaccard similarity between aspects ranged from 0.38 (Quality-Service) to 0.58 (Service-Decoration); Quality-Price was 0.52, Quality-Decoration 0.45, Service-Price 0.41, and Price-Decoration 0.50.
• Zero-shot transfer degrades but holds: Applied directly to unseen target domains without fine-tuning, the model retained 67-76% of source domain effectiveness, with a domain gap ranging from 12.7% to 23.2%. The paper's available content states that BanglaBook Reviews achieved the best result, but the remaining per-domain figures are cut off in the provided text.
• Few-shot learning is cost-efficient: With only 500-1000 samples, performance reached 90-95% of full fine-tuning, which the authors present as a significant reduction in annotation cost.
• SHAP reveals aspect-specific feature behavior: Quality, Service, and Decoration features contributed predominantly positively, while the Price aspect showed multiple significant negative contributors, consistent with price-related complaints.

Methodology in Plain English

The authors built a dataset first: they scraped reviews from Bangladeshi e-commerce sites (Daraz, Facebook Marketplace, Rokomari, Shajgoj, and others) using API integration and HTML parsing, then cleaned the text by removing special characters, extra whitespace, and non-Bangla characters, normalizing spelling variants with Bangla tokenizers, and manually verifying spelling against the Bangla Academy Accessible Dictionary. Each review was labeled by three Computer Science undergraduates plus one NLP specialist, with two annotators per review, a third expert resolving disagreements, and majority voting for hard cases. Each review carries sentiment labels for four separate aspects rather than one overall polarity.

For the model, they combined four architectures that learn different things: BanglaBERT (12 transformer layers, 768-dimensional hidden state, 12 attention heads) captures long-range context; BiLSTM (128 hidden units × 2 directions) reads text in both directions to handle Bangla's freer word order; a standard LSTM (256 hidden units, 2 layers) captures forward narrative flow; and a GRU (200 hidden units, 2 layers) provides a lighter, faster sequential model. Words are represented by blending static embeddings (300-dimensional GloVe trained on 2.5 billion Bangla tokens, plus FastText for domain jargon and morphology) with context-dependent BanglaBERT embeddings, using learned weights in the blend. The four models' predictions are combined by adaptive weighted voting whose weights are tuned dynamically based on input characteristics like review length and complexity. BanglaBERT was trained with AdamW at a 2e-5 learning rate with linear scheduling; the recurrent models used Adam at 1e-3 with exponential decay. A calibrated threshold of 0.5 was used for final classification.

For transparency, they applied SHAP to quantify each word's contribution to a prediction, visualized attention weights to show which tokens the model focused on, and used LIME, then asked human evaluators to rate how interpretable and useful the explanations were. For transfer to new domains, they adapted shared features and parameters from the source domain to targets using feature alignment and classifier fine-tuning, with domain-invariant representation learning and instance weighting to avoid negative transfer.

Why This Matters

Impact on research. The work targets a genuinely under-resourced language and combines three threads that are usually treated separately: multi-aspect (not just polarity) classification, interpretability, and cross-domain generalization. It also contributes a reusable annotated dataset and reports a systematic comparison of ensemble components, addressing gaps the authors identify in prior Bangla NLP work, which has mostly focused on document-level sentiment and single architectures.

Real-world applications:

• Pricing optimization, by surfacing how customers talk about cost relative to quality.
• Service improvement, by isolating service-related sentiment from product-related sentiment in the same review.
• Customer experience enhancement and product presentation refinement, using the Decoration aspect labels.
• Market research and brand monitoring at scale, where few-shot transfer lets a team adapt the model to a new product category with only 500-1000 labeled examples.

Industry relevance. The paper explicitly frames deployment case studies for Bangladeshi e-commerce platforms, arguing that explainability matters because businesses need to trust and act on model outputs rather than accept black-box scores. The claimed reduction in annotation cost via few-shot learning is directly relevant to companies that cannot afford large labeling efforts, and the multi-aspect label structure maps onto the specific operational levers (price, service, quality, presentation) that a retailer can actually change.

Future Directions

• Closing the zero-shot domain gap. Performance dropped to 67-76% of source-domain effectiveness with domain gaps of 12.7-23.2%; narrowing this, especially for linguistically distant targets such as social media posts, is an open problem.
• Scaling and diversifying annotation. The current dataset covers 8,755 reviews dominated by one platform (Daraz, 6,521); expanding to more platforms, product categories, and regions would test whether the benchmark generalizes.
• Extending explainability beyond the current suite. The reported methods are SHAP, LIME, and attention visualization; incorporating linguistically grounded explanations such as sentiment lexicons, which the methodology section mentions, invites further evaluation.
• Broadening to other low-resource languages and deployment validation. The framework's design choices are language-agnostic in principle, and the real-world deployment case studies referenced in Section 7.7 are not detailed in the available content, leaving their operational results to be examined in the full paper.

Target Audience

Researchers in low-resource NLP, Bangla language processing, and aspect-based sentiment analysis; practitioners building explainable AI systems for e-commerce or customer feedback analytics; and industry data science teams in South Asian markets who need interpretable, transferable sentiment models without large annotation budgets. Readers without a background in deep learning or transformer architectures will find the paper's technical sections demanding.

Authors’ abstract

Multi-aspect sentiment analysis of Bangla e-commerce reviews remains challenging due to limited annotated datasets, morphological complexity, code-mixing phenomena, and domain shift issues, affecting 300 million Bangla-speaking users. Existing approaches lack explainability and cross-domain generalization capabilities crucial for practical deployment. We present BanglaSentNet, an explainable hybrid deep learning framework integrating LSTM, BiLSTM, GRU, and BanglaBERT through dynamic weighted ensemble learning for multi-aspect sentiment classification. We introduce a dataset of 8,755 manually annotated Bangla product reviews across four aspects (Quality, Service, Price, Decoration) from major Bangladeshi e-commerce platforms. Our framework incorporates SHAP-based feature attribution and attention visualization for transparent insights. BanglaSentNet achieves 85% accuracy and 0.88 F1-score, outperforming standalone deep learning models by 3-7% and traditional approaches substantially. The explainability suite achieves 9.4/10 interpretability score with 87.6% human agreement. Cross-domain transfer learning experiments reveal robust generalization: zero-shot performance retains 67-76% effectiveness across diverse domains (BanglaBook reviews, social media, general e-commerce, news headlines); few-shot learning with 500-1000 samples achieves 90-95% of full fine-tuning performance, significantly reducing annotation costs. Real-world deployment demonstrates practical utility for Bangladeshi e-commerce platforms, enabling data-driven decision-making for pricing optimization, service improvement, and customer experience enhancement. This research establishes a new state-of-the-art benchmark for Bangla sentiment analysis, advances ensemble learning methodologies for low-resource languages, and provides actionable solutions for commercial applications.