A Maximum-Entropy approach for accurate document annotation in the biomedical domain

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Information Retrieval

Background:

• The rapid growth of scientific literature necessitates efficient document classification and search tools.
• Ontology-based approaches enable semantic processing, enhancing information retrieval and advancing the Semantic Web.
• Document annotation with ontology concepts is crucial for semantic search and can be framed as a classification task.

Purpose of the Study:

• To develop an automated and robust method for annotating biomedical literature with Medical Subject Headings (MeSH) terms.
• To evaluate the performance of a Maximum Entropy approach for this annotation task.
• To compare the proposed method against other classification approaches like Naive Bayes and Decision Trees.

Main Methods:

• A Maximum Entropy (MaxEnt) classification approach was employed for automated document annotation.
• The method annotates biomedical literature documents using terms from the Medical Subject Headings (MeSH) controlled vocabulary.
• Performance was evaluated using metrics such as precision, recall, and F-measure, including comparisons with Naive Bayes and Decision Trees.

Main Results:

• The Maximum Entropy approach achieved a high average F-measure of 92.4% across 4,078 MeSH terms.
• The method demonstrated robustness to term ambiguity, with an average F-measure of 92.42% on ambiguous terms.
• The MaxEnt approach outperformed Naive Bayes and Decision Trees in terms of F-Measure for both ambiguous and unambiguous MeSH terms.
• The algorithm performed well even with a very small number of training documents.

Conclusions:

• The proposed Maximum Entropy method is a highly accurate and robust solution for annotating biomedical literature with MeSH terms.
• This approach significantly improves the efficiency and quality of searching biomedical documents, contributing to the Semantic Web.
• The method's resilience to term ambiguity and effectiveness with limited training data make it a valuable tool for biomedical information retrieval.