List-wise learning to rank biomedical question-answer pairs with deep ranking recursive autoencoders.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Machine Learning

Background:

• Biomedical question answering (QA) systems are crucial for accessing vital information.
• Current systems often use information retrieval (IR) but struggle with semantic understanding.
• Existing methods fail to accurately capture syntactic and semantic relatedness for natural language answers.

Purpose of the Study:

• To introduce a novel deep learning architecture, ranking recursive autoencoders (rankingRAE), for improved biomedical QA.
• To enhance the ranking of candidate answers for biomedical questions.
• To address limitations of traditional IR-based approaches in semantic relatedness.

Main Methods:

• Developed a rankingRAE architecture utilizing deep learning for semantic feature representation.
• Converted answer ranking into simultaneous binary classification tasks.
• Employed recursive autoencoders for unsupervised learning of semantic representations.
• Introduced a listwise "ranking error" loss function to penalize incorrect answer rankings.

Main Results:

• The rankingRAE architecture demonstrated robustness and effectiveness.
• Experimental results showed superior performance compared to classical IR models and other deep learning approaches.
• The model achieved strong performance on the BioASQ 2013-2018 benchmarks.

Conclusions:

• The proposed rankingRAE model offers a significant advancement in biomedical question answering.
• Deep learning, specifically rankingRAE, provides a more effective approach to semantic understanding and answer ranking.
• This work contributes to more accurate and reliable retrieval of biomedical information.