Imbalanced Text Sentiment Classification Based on Multi-Channel BLTCN-BLSTM Self-Attention.

Area of Science:

• Natural Language Processing
• Machine Learning
• Data Science

Background:

• Imbalanced data distribution is a common challenge in practical natural language processing tasks.
• Existing methods often assume data balance, limiting their effectiveness on skewed datasets.
• Sentiment analysis of short, imbalanced text reviews requires specialized approaches.

Purpose of the Study:

• To address the challenge of imbalanced data in short text review sentiment classification.
• To propose a novel fusion multi-channel BLTCN-BLSTM self-attention model for enhanced sentiment analysis.
• To improve the accuracy and reliability of sentiment prediction on imbalanced datasets.

Main Methods:

• Developed a multi-channel BLTCN-BLSTM self-attention network model.
• Utilized word embedding processing for input feature extraction.
• Integrated a self-attention mechanism to enhance sentiment feature extraction.
• Combined focus loss rebalancing and classifier enhancement techniques.

Main Results:

• Achieved an optimal F1 value of 0.893 on the Chnsenticorp-HPL-10,000 corpus.
• Demonstrated significant improvements in accuracy, recall, and F1-measure compared to baseline methods.
• The model effectively integrates the weight of emotional feature words for better sentiment classification.

Conclusions:

• The proposed fusion multi-channel BLTCN-BLSTM self-attention model is effective for imbalanced short-text sentiment classification.
• The method successfully addresses data imbalance issues in sentiment analysis.
• This approach offers a robust solution for extracting and analyzing sentiment from challenging datasets.