A structural SVM approach for reference parsing.

Area of Science:

• Bibliometrics
• Computational Linguistics
• Information Science

Background:

• Automated extraction of bibliographic data is crucial for creating large citation databases.
• Parsing individual references is a necessary preprocessing step for citation-indexing systems.
• The structured nature of references makes reference parsing a suitable sequence learning problem.

Purpose of the Study:

• To study the effectiveness of structural Support Vector Machines (SVM) for reference parsing.
• To compare the performance of structural SVM with conventional SVM and Conditional Random Field (CRF).

Main Methods:

• Implementation of structural SVM with contextual features.
• Comparison of structural SVM, conventional SVM, and CRF using token and chunk-level accuracies.
• Evaluation using binary features and contextual observation features from neighboring tokens.

Main Results:

• Both structural SVM and conventional SVM achieved over 98% token classification accuracy and 95% chunk-level accuracy.
• Structural SVM and CRF demonstrated similar accuracies at token and chunk levels.
• SVM performance improved significantly with the addition of contextual observation features.

Conclusions:

• Structural SVM outperforms conventional SVM when only basic features are used, due to its utilization of contextual label features.
• When contextual observation features are incorporated, SVM performance approaches that of structural SVM.
• Both structural SVM and CRF exhibit superior sequence learning capabilities for reference parsing compared to conventional SVM.