BanglaSentNet: Hybrid Deep Sentiment Analysis

• The paper demonstrates that a dynamic weighted ensemble of LSTM, BiLSTM, GRU, and BanglaBERT significantly improves multi-aspect sentiment classification accuracy in Bangla reviews.
• It integrates SHAP-based feature attribution and attention visualization to deliver transparent model interpretability for actionable business insights.
• Empirical benchmarks confirm BanglaSentNet's superiority over traditional and deep learning baselines, highlighting its robust cross-domain transfer and few-shot adaptation capabilities.

BanglaSentNet is an explainable hybrid deep learning framework designed for multi-aspect sentiment analysis within Bangla-language e-commerce reviews, addressing significant challenges in low-resource language environments such as morphological complexity, code-mixing, domain shift, and limited annotated data. Integrating four advanced neural architectures—LSTM, BiLSTM, GRU, and BanglaBERT—via a dynamic weighted ensemble, BanglaSentNet sets a state-of-the-art benchmark for multi-aspect and cross-domain sentiment classification, supporting transparent model interpretation through a comprehensive explainability suite (Islam et al., 28 Nov 2025).

1. Architecture and Model Components

BanglaSentNet combines four base models, each engineered to capture complementary linguistic features in Bangla reviews:

• LSTM: Two layers of 256 hidden units, dropout rate 0.3, Adam optimizer (initial learning rate 1e-3, exponential decay), and gradient-norm clipping at 1.0 for stability. Models forward sequential dependencies crucial for sentiment patterns.
• BiLSTM: Stacked architecture (2 layers, 128 units per direction), dropout 0.3, recurrent dropout 0.2, enabling bidirectional context aggregation to suit flexible Bangla word order.
• GRU: Two layers, 200 hidden units, dropout 0.3, offering faster inference and simplified gating, tuned for local sentiment extraction.
• BanglaBERT: Pre-trained transformer with 12 layers (hidden size 768, 12 attention heads), fine-tuned using AdamW (learning rate 2e-5, linear warm-up/decay), layer-wise learning rate decay and gradual unfreezing to maximize transfer of pre-trained knowledge.

The embedding pipeline fuses static word vectors (300-dimensional GloVe from a 2.5B-token Bangla corpus, FastText subword embeddings) and BanglaBERT contextual outputs, optimized via:

$$E_{\mathrm{final}} = \alpha E_{\mathrm{static}} + \beta E_{\mathrm{contextual}}$$

with $\alpha, \beta$ learnable end-to-end parameters.

Final predictions for multi-aspect sentiment are delivered using a dynamic weighted ensemble:

$$\hat{y} = \sum_{i=1}^4 w_i(x)\hat{y}_i, \quad w_i(x) = \frac{\exp(g_i(x))}{\sum_{j=1}^4 \exp(g_j(x))}$$

where $g_i(x)$ are input-dependent attention-based scoring networks. Ensemble training minimizes:

$$\mathcal{L}_{\mathrm{ens}} = \mathcal{L}_{\mathrm{CE}}(y, \hat{y}) + \lambda \sum_{i<j} \mathrm{div}(\hat{y}_i, \hat{y}_j) + \gamma \|w\|_2^2$$

to promote output diversity and prevent overfitting.

2. Dataset Curation and Annotation Protocol

BanglaSentNet introduces a multi-aspect, multi-label dataset comprising 8,755 manually annotated Bangla product reviews, sourced from major platforms:

Source	Review Count
Daraz	6,521
Facebook Marketplace	1,145
Rokomari	494
Shajgoj	155
Other	440

Each review is labeled across four aspects:

Aspect	Total	Positive	Negative	Neutral
Quality	4,000	2,400	1,200	400
Service	1,000	600	300	100
Price	2,500	1,500	900	100
Decoration	1,255	750	400	105

Annotation utilized a two-pass procedure: two CS undergraduates (Group A) conducted initial labeling; adjudication followed by an NLP expert (Group B); disagreements resolved by majority vote involving a third expert. Inter-annotator reliability reached Cohen’s $\kappa \approx 0.84$–$0.88$, demonstrating high consistency.

3. Explainability and Interpretability Mechanisms

BanglaSentNet implements a comprehensive explainability suite to ensure model transparency:

• SHAP-Based Feature Attribution: For each input feature $x_j$, Shapley values are computed:

$$\phi_j = \sum_{S\subseteq F\setminus\{j\}} \frac{|S|!(|F|-|S|-1)!}{|F|!} \left[ f_{S\cup\{j\}}(x_{S\cup\{j\}}) - f_S(x_S) \right]$$

measuring $x_j$'s marginal impact on predictions.

• Attention Visualization: Attention energies $e_t$ are normalized:

$\alpha_t = \frac{\exp(e_t)}{\sum_k \exp(e_k)}$

yielding token-wise importance heatmaps, mapped per aspect.

• Quantitative Interpretability: Evaluation across methods and suite combinations demonstrates measurable gains:

XAI Approach	Interp. Score	Human Agreement (%)
None	2.1	-
Attention	7.3	76.3
SHAP	8.1	81.2
LIME	7.8	79.5
Full Suite	9.4	87.6

A notable outcome is the full suite’s 9.4/10 interpretability and 87.6% human agreement, exceeding standalone XAI approaches.

4. Cross-Domain and Transfer Learning Capabilities

BanglaSentNet’s design explicitly targets robust cross-domain transfer for Bangla sentiment analysis:

• Zero-Shot Performance: Without fine-tuning, model effectiveness is well preserved across diverse targets:

Domain	F1-Score	Drop (%) vs. Source
BanglaBook	0.761	-12.7
Social Media	0.710	-18.3
E-commerce	0.734	—
News Headlines	0.672	-23.2

Aspect-wise transfer (zero-shot) F1 averages: Quality 0.742; Service 0.700; Decoration 0.706.

• Few-Shot Adaptation: Fine-tuning with as few as 50 labeled samples yields +2.5–3.8 F1 improvement; performance plateaus after 500 samples, attaining 90–95% of full fine-tuning effectiveness.
• Domain Adaptation: Techniques include adversarial feature alignment, weighted instance re-sampling, dual adaptation (encoder/classifier fine-tuning), and domain-specific regularization to mitigate distributional shift.

This suggests that BanglaSentNet is well-suited for deployment in domains where annotated data is highly limited, with low annotation overheads and maintained accuracy.

5. Empirical Benchmarks and Comparative Evaluation

BanglaSentNet’s performance is validated against traditional and deep learning baselines on the curated dataset:

Model	Accuracy	F1-Score
LR (TF-IDF)	0.40	0.47
SVM	0.49	0.58
RF	0.43	0.50
CNN+GloVe	0.59	0.75
LSTM+FastText	0.66	0.70
BiLSTM	0.56	0.77
GRU+GloVe	0.64	0.77
LSTM	0.71	0.80
BiLSTM	0.73	0.82
GRU	0.74	0.76
BanglaBERT	0.78	0.85
BanglaSentNet	0.85	0.88

Relative F1 improvement over best single deep model is +3–7%. Ablation analysis reveals steep performance drops (all significant at $p<0.01$):

Removed Component	F1 after Removal
BanglaBERT	0.75
BiLSTM	0.79
LSTM	0.81
GRU	0.78

A plausible implication is that model diversity in both sequence modeling and transformer architectures is critical for robust, aspect-sensitive sentiment extraction in Bangla.

6. Deployment and Commercial Utility

BanglaSentNet has been integrated into a real-world prototype analytics dashboard for major Bangladeshi platforms (Daraz, Facebook Marketplace, Rokomari, Shajgoj). Capabilities include real-time monitoring of aspect-level sentiment, enabling actionable business intelligence:

• Pricing Optimization: Electronics sellers adjusted discounting tiers in response to adverse “Price” sentiment.
• Service Improvement: Automated detection of negative “Service” reviews drove targeted delivery retraining.
• Product UX Enhancement: Multi-aspect outputs guided packaging redesigns by surfacing quality-decoration trade-offs.
• Sentiment Routing: Disaggregation of reviews with mixed aspect sentiment (e.g., “Quality +” and “Price –”) enabled differentiated business unit interventions.

By jointly optimizing a diverse ensemble, leveraging both static and contextual embeddings, and embedding an explainability suite comprising SHAP and attention visualization, BanglaSentNet demonstrates advanced multi-aspect sentiment analysis and interpretable cross-domain generalization for Bangla e-commerce platforms (Islam et al., 28 Nov 2025). Its framework offers practical solutions applicable to low-resource commercial environments, substantiated by empirically robust benchmarks and domain adaptation protocols.