Locating and parsing bibliographic references in HTML medical articles

Area of Science:

• Bibliometrics
• Computational Linguistics
• Medical Informatics

Background:

• Bibliographic databases rely on accurate citation data, but references are not always a distinct field.
• Automated extraction and indexing of article references are essential for efficient database creation and maintenance.
• Extracting components like author, title, and journal from references is a valuable task for bibliographic data.

Purpose of the Study:

• To develop and evaluate a two-step automated process for locating and parsing references in HTML medical articles.
• To improve the efficiency and accuracy of extracting bibliographic data from scientific literature.
• To minimize human labor in the creation and maintenance of large bibliographic databases.

Main Methods:

• A two-step statistical machine learning process was employed.
• Step 1 (Reference Locating): A two-class classification problem using text and geometric features to identify and decompose the reference section.
• Step 2 (Reference Parsing): Implementation and comparison of two algorithms—Conditional Random Field (CRF) based on sequence statistics and Support Vector Machine (SVM) based on local feature statistics with rule-based correction.

Main Results:

• Near-perfect precision and recall rates (over 99%) were achieved for locating references in 500 articles from 100 medical journals.
• Both reference-parsing algorithms demonstrated high performance: over 99% accuracy at the word level and over 97% accuracy at the chunk level.
• The study successfully automated the extraction of individual reference components.

Conclusions:

• The proposed two-step machine learning approach effectively locates and parses references in medical articles.
• This method significantly enhances the automation of bibliographic data extraction, offering a valuable tool for database management.
• The high accuracy rates demonstrate the feasibility and effectiveness of using statistical machine learning for processing scientific literature references.