Unsupervised discovery of information structure in biomedical documents

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Text Mining

Background:

• Information structure (IS) analysis classifies biomedical text into research components like objectives, methods, results, and conclusions.
• IS analysis aids biomedical text mining and literature review but typically requires extensive domain-specific annotated data.
• Current supervised methods are limited by the high cost and domain-specificity of manual annotations in biomedicine.

Purpose of the Study:

• To investigate an unsupervised approach for information structure analysis in biomedical literature.
• To evaluate the performance of various unsupervised methods on a large corpus of PubMed abstracts.
• To determine if unsupervised methods can achieve performance comparable to supervised approaches and reduce annotation costs.

Main Methods:

• Evaluation of several unsupervised machine learning algorithms for IS analysis.
• Application of algorithms to a large corpus of biomedical abstracts from PubMed.
• Comparison of unsupervised method performance against established IS schemes and supervised methods.

Main Results:

• The best unsupervised algorithm, a multilevel-weighted graph clustering algorithm, achieved F-scores over 0.70 for most IS categories.
• Unsupervised approach performance closely rivals lightly supervised methods, proving sufficient for practical tasks.
• Unsupervised learning identified novel information categories beyond existing IS schemes.

Conclusions:

• Unsupervised IS analysis is a viable and effective approach for biomedical literature.
• This method significantly reduces the need for expensive, domain-specific manual annotations.
• The unsupervised approach facilitates broader application of IS analysis across diverse biomedical sub-domains and offers new insights.