Improving Biomedical Question Answering by Data Augmentation and Model Weighting.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Biomedical question answering (QA) systems face challenges due to specialized domain language and limited large-scale datasets.
• Existing methods struggle with data scarcity and performance degradation, particularly with adversarial samples.

Purpose of the Study:

• To improve the performance and robustness of biomedical question answering systems.
• To address limitations caused by insufficient training data and adversarial examples in specialized domains.

Main Methods:

• Implemented data augmentation techniques: sliding window, summarization, and round-trip translation.
• Proposed a model weighting strategy combining an open-domain model (QANet) and a biomedical-specific model (BioBERT).
• Utilized adversarial training to enhance model robustness.

Main Results:

• Achieved significant performance improvements compared to single models and other BioASQ challenge participants.
• Demonstrated enhanced semantic understanding through data augmentation and adversarial training.
• Showcased improved ability to handle complex biomedical question answering tasks.

Conclusions:

• The proposed approach effectively overcomes data limitations in biomedical QA.
• Data augmentation, model weighting, and adversarial training are crucial for robust biomedical QA.
• The developed methods offer a promising solution for complex biomedical information retrieval.