Knowledge-Based Biomedical Word Sense Disambiguation with Neural Concept Embeddings

Area of Science:

• Natural Language Processing
• Bioinformatics
• Computational Linguistics

Background:

• Biomedical word sense disambiguation (WSD) is crucial for NLP tasks like named entity recognition.
• Existing methods often require extensive labeled data, which is impractical for complex domains.
• Advances in neural embeddings offer new possibilities for WSD.

Purpose of the Study:

• To improve state-of-the-art biomedical WSD using knowledge-based approaches and neural embeddings.
• To evaluate methods employing weak supervision, avoiding hand-labeled WSD examples.
• To demonstrate the effectiveness of representation learning in biomedical WSD.

Main Methods:

• Utilized knowledge-based approaches combined with neural word/concept embeddings.
• Employed weak supervision, leveraging MetaMap for concept vector generation.
• Developed a linear-time method and a nearest neighbor framework for WSD.

Main Results:

• Achieved 92.24% accuracy with a linear-time method on the MSH WSD dataset, a 3% improvement over unsupervised methods.
• The nearest neighbor approach reached 94.34% accuracy, halving the error rate.
• Demonstrated that weak supervision with representation learning can match supervised approaches.

Conclusions:

• Neural embeddings and weak supervision significantly advance biomedical WSD.
• This approach is effective for complex, evolving domains like biomedicine.
• External knowledge bases remain vital for achieving high performance in WSD.