Transfer Learning for Sentiment Analysis Using BERT Based Supervised Fine-Tuning.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning
• Computational Linguistics

Background:

• The proliferation of internet data necessitates advanced sentiment analysis techniques.
• Bangla Natural Language Processing (NLP) faces challenges due to limited standardized labeled data.
• Existing Bangla sentiment analysis often relies on context-independent word embeddings (Word2Vec, GloVe, fastText).

Purpose of the Study:

• To improve Bangla sentiment analysis performance by leveraging transfer learning.
• To integrate BERT's contextual understanding with deep learning models like CNN-BiLSTM.
• To compare the effectiveness of BERT transfer learning against traditional word embeddings and classical machine learning algorithms.

Main Methods:

• Utilized BERT's transfer learning capability within a deep integrated CNN-BiLSTM model.
• Applied transfer learning to classical machine learning algorithms for comparative analysis.
• Evaluated various word embedding techniques including Word2Vec, GloVe, fastText, and BERT transfer learning.

Main Results:

• Achieved state-of-the-art binary classification performance for Bangla sentiment analysis.
• Demonstrated significant outperformance of the BERT transfer learning strategy over all compared embeddings and algorithms.
• The integrated CNN-BiLSTM model with BERT transfer learning showed superior results.

Conclusions:

• BERT transfer learning significantly enhances Bangla sentiment analysis.
• The proposed deep integrated model offers a superior approach compared to existing methods.
• This research provides a robust framework for future Bangla NLP tasks.