SentiMedQAer: A Transfer Learning-Based Sentiment-Aware Model for Biomedical Question Answering.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Transfer learning is increasingly applied to Natural Language Processing (NLP) tasks, including question answering (QA).
• Biomedical QA datasets often suffer from limited training data and a prevalence of factoid questions.
• Existing models may not fully capture the nuances required for complex biomedical reasoning.

Purpose of the Study:

• To propose a novel transfer learning-based sentiment-aware model, SentiMedQAer, for enhanced biomedical question answering.
• To address the limitations of insufficient training examples and factoid question dominance in biomedical QA datasets.
• To integrate sentiment analysis into a QA framework for improved performance on reasoning-required questions.

Main Methods:

• Developed a learning pipeline utilizing BioBERT for contextual and domain-specific text encoding.
• Fine-tuned Text-to-Text Transfer Transformer (T5) and RoBERTa models to incorporate sentiment information.
• Employed an XGBoost classifier to output confidence scores for final answer determination.

Main Results:

• SentiMedQAer demonstrated superior performance on the PubMedQA dataset, which features reasoning-required yes/no questions.
• The proposed model achieved a 15.83% improvement over the state-of-the-art (SOTA) methods.
• SentiMedQAer outperformed a single human annotator by 5.91%.

Conclusions:

• SentiMedQAer represents a significant advancement in biomedical question answering.
• The integration of sentiment analysis and transfer learning enhances model accuracy for complex biomedical queries.
• The model shows strong potential for practical applications in biomedical information retrieval and knowledge discovery.