Transfer Learning for Sentiment Classification Using Bidirectional Encoder Representations from Transformers (BERT)...

Area of Science:

• Natural Language Processing
• Machine Learning
• Computational Linguistics

Background:

• Sentiment analysis is vital for recommender systems and understanding online user opinions.
• Existing methods often rely on manual feature engineering and shallow learning, limiting generalization.
• Conflicting results exist regarding the efficacy of various methodologies for predicting review usefulness.

Purpose of the Study:

• To develop a generalized sentiment analysis approach using transfer learning.
• To apply and evaluate a Bidirectional Encoder Representations from Transformers (BERT)-based model.
• To compare the performance of BERT classification against conventional machine learning techniques.

Main Methods:

• Utilized a BERT-based model for sentiment classification.
• Employed transfer learning to enhance model generalization.
• Conducted comparative experiments on Yelp reviews, evaluating BERT against other machine learning approaches.
• Investigated the impact of batch size and sequence length on BERT classifier performance.

Main Results:

• The proposed BERT model demonstrated superior performance, achieving high accuracy and outstanding prediction.
• Fine-tuned BERT classification outperformed other methods on positive and negative Yelp reviews.
• Batch size and sequence length were identified as significant factors influencing BERT classification performance.

Conclusions:

• BERT-based transfer learning offers a more generalized and effective approach to sentiment analysis.
• The model shows significant improvements over traditional methods for classifying online review sentiment.
• Further research should consider the impact of hyperparameters like batch size and sequence length for optimal BERT performance.