Sentiment interpretability analysis on Chinese texts employing multi-task and knowledge base.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models, while powerful, often function as "black boxes," limiting their use in critical applications due to low interpretability.
• Interpreting sentiment analysis models is crucial for understanding their decision-making processes and ensuring reliable outcomes.

Purpose of the Study:

• To develop a comprehensive method for enhancing the interpretability of sentiment analysis models.
• To address the challenges of low interpretability and incomplete evidence extraction in sentiment models, particularly for Chinese texts.

Main Methods:

• Implemented an attention-based analysis by training sentiment classification and generation tasks to capture multi-perspective attention scores.
• Integrated an external knowledge base and leveraged character scores for improved and complete sentiment evidence phrase extraction.
• Evaluated the method on a sentiment interpretability dataset.

Main Results:

• Demonstrated significant improvements in model performance.
• Achieved a 1.3% increase in accuracy, a 13% increase in Macro-F1 score, and a 23% increase in Mean Average Precision (MAP).

Conclusions:

• The proposed method effectively enhances sentiment model interpretability by combining attention mechanisms and external knowledge.
• The approach provides a robust solution for more transparent and reliable sentiment analysis, particularly beneficial for Chinese language processing.