Scholarlink

<p>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 over 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. It integrates LSTM, BiLSTM, GRU, and BanglaBERT through a theoretically grounded dynamic weighted ensemble for multi-aspect sentiment classification. We introduce a large-scale dataset of 8755 manually annotated Bangla product reviews across four aspects—Quality, Service, Price, and Decoration—collected from major Bangladeshi e-commerce platforms, with a verified inter-annotator agreement of Fleiss’ <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\kappa = 0.821\)</EquationSource> </InlineEquation>. Our framework incorporates an integrated explainability suite combining SHAP-based feature attribution and attention mechanism visualization to provide transparent, interpretable insights. Experimental results over five random seeds demonstrate that BanglaSentNet achieves <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(85.0\% \pm 0.6\%\)</EquationSource> </InlineEquation> accuracy and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(0.880 \pm 0.007\)</EquationSource> </InlineEquation> F1-score, outperforming standalone deep learning models by 3–7% and traditional machine learning approaches substantially. The explainability components achieve a 9.4/10 interpretability score with 87.6% human evaluation agreement from 18 evaluators. Cross-domain transfer learning experiments reveal robust generalization, with zero-shot performance retaining 67–76% of source domain effectiveness across diverse target domains. Few-shot learning with only 500–1000 samples achieves 90–95% of full fine-tuning performance, significantly reducing annotation costs. This research advances ensemble learning methodologies for low-resource languages and provides actionable, interpretable solutions for Bangla e-commerce platforms.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

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

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Ariful Islam,
  • Md Rifat Hossen,
  • Tanvir Mahmud

摘要

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 over 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. It integrates LSTM, BiLSTM, GRU, and BanglaBERT through a theoretically grounded dynamic weighted ensemble for multi-aspect sentiment classification. We introduce a large-scale dataset of 8755 manually annotated Bangla product reviews across four aspects—Quality, Service, Price, and Decoration—collected from major Bangladeshi e-commerce platforms, with a verified inter-annotator agreement of Fleiss’ \(\kappa = 0.821\) . Our framework incorporates an integrated explainability suite combining SHAP-based feature attribution and attention mechanism visualization to provide transparent, interpretable insights. Experimental results over five random seeds demonstrate that BanglaSentNet achieves \(85.0\% \pm 0.6\%\) accuracy and \(0.880 \pm 0.007\) F1-score, outperforming standalone deep learning models by 3–7% and traditional machine learning approaches substantially. The explainability components achieve a 9.4/10 interpretability score with 87.6% human evaluation agreement from 18 evaluators. Cross-domain transfer learning experiments reveal robust generalization, with zero-shot performance retaining 67–76% of source domain effectiveness across diverse target domains. Few-shot learning with only 500–1000 samples achieves 90–95% of full fine-tuning performance, significantly reducing annotation costs. This research advances ensemble learning methodologies for low-resource languages and provides actionable, interpretable solutions for Bangla e-commerce platforms.