Exploration of text matching methods in Chinese disease Q&A systems: A method using ensemble based on BERT and...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Text matching is crucial in Natural Language Processing, with significant value in the medical domain due to the complexity of Chinese medical texts.
• A competition released 30,000 Chinese medical Q&A datasets for diabetes, hypertension, hepatitis B, AIDS, and breast cancer, attracting 90 teams for evaluation.
• The unique characteristics of the Chinese language and medical terminology present specific challenges for text matching tasks.

Purpose of the Study:

• To identify the optimal text matching method for Chinese medical Q&A data through participation in an evaluation competition.
• To compare the efficacy of different machine learning models in processing specialized medical text data.
• To explore advanced techniques for improving the accuracy of information retrieval in medical question-answering systems.

Main Methods:

• Utilized Bidirectional Encoder Representations from Transformers (BERT) and a boosted tree model for comparative analysis of Chinese medical Q&A data.
• Conducted feature engineering to extract and analyze the importance of various linguistic and structural features within the Q&A pairs.
• Developed an ensemble model by integrating BERT and boosted tree approaches, validated through correlation analysis to assess synergistic effects.

Main Results:

• The ensemble model achieved the highest F1 score of 0.90825, outperforming all 90 participating teams and single models (BERT: 0.87443, Boosted Tree: 0.86915).
• Feature importance analysis revealed that the out-degree and in-degree of Q&A sentences were critical factors for model performance.
• Correlation analysis showed high similarity (0.9) between models of the same type and moderate discrimination (0.7) between different model types, highlighting the benefit of ensemble diversity.

Conclusions:

• The proposed ensemble method effectively enhances single model performance for Chinese medical Q&A text matching tasks.
• The integrated model demonstrates strong results and good generalization properties, making it a robust solution for real-world applications.
• This study provides valuable insights into optimizing text matching techniques for specialized biomedical datasets.