Knowledge based word-concept model estimation and refinement for biomedical text mining.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Computational Biology

Background:

• Large public biomedical databases rely on text mining.
• Knowledge bases (KB) aid biomedical text mining but are suboptimal for machine learning.
• Supervised methods require labeled data, limiting large-scale application.

Purpose of the Study:

• To develop a novel KB-based method for generating word-concept probabilities.
• To leverage unlabeled corpora like MEDLINE for improved text mining.
• To create a method that does not require labeled training data.

Main Methods:

• Generated word-concept probabilities from KB descriptions and MEDLINE texts.
• Utilized MetaMap for entity recognition and co-occurrence analysis.
• Estimated model parameters without labeled training data.

Main Results:

• Achieved higher accuracy in word sense disambiguation (WSD) on the MSH WSD dataset compared to state-of-the-art methods.
• Demonstrated improved performance in document ranking using MEDLINE citations over baseline approaches.

Conclusions:

• The novel method effectively generates word-concept probabilities from KBs and unlabeled text.
• This approach enhances performance in biomedical text mining tasks without requiring labeled data.
• The method offers a scalable and accurate solution for large-scale biomedical text analysis.