tRF-BERT: A transformative approach to aspect-based sentiment analysis in the bengali language

Area of Science:

• Natural Language Processing
• Computational Linguistics

Background:

• Sentiment analysis is increasingly important, especially for Bengali text.
• Aspect-Based Sentiment Analysis (ABSA) in Bengali is challenging due to data scarcity and linguistic complexity.

Purpose of the Study:

• To improve Aspect-Based Sentiment Analysis (ABSA) performance for the Bengali language.
• To evaluate the effectiveness of transformer-based models and a novel hybrid model for Bengali ABSA.

Main Methods:

• Utilized two Bengali benchmark datasets: Cricket and Restaurant.
• Implemented and compared Bidirectional Encoder Representations from Transformers (BERT), Robustly Optimized BERT Approach (RoBERTa), and a proposed transformative Random Forest and BERT (tRF-BERT) hybrid model.
• Benchmarked against existing Random Forest, Support Vector Machine, K-Nearest Neighbors, and Convolutional Neural Network models.

Main Results:

• All implemented models outperformed previous works on the benchmark datasets.
• The proposed tRF-BERT model achieved the highest accuracy and F1 scores.
• Achieved 0.89 accuracy and 0.85 F1 score for aspect detection on the Cricket dataset.
• Achieved 0.92 accuracy and 0.89 F1 score for aspect detection on the Restaurant dataset.

Conclusions:

• The tRF-BERT hybrid model offers a superior approach for Bengali Aspect-Based Sentiment Analysis.
• Transformer-based models, particularly the proposed hybrid, demonstrate significant advancements in handling Bengali NLP tasks.