Examination Of Effective Features For Crf-Based Bibliography Extraction From Reference Strings

Metadata such as bibliographic information about documents are indispensable in the effective use of digital libraries. In particular, the reference fields of academic papers contain much bibliographic information such as authors' names and document titles. We are therefore developing a method for automatically extracting bibliographic information from reference strings using a conditional random field (CRF). The features used by the CRF determine the accuracy of this method. We examine effective features for accurate extraction by experimentally changing the features used. The experiments showed that lexical features were quite effective in accurate extraction and augmenting lexicons properly could lead to further improvements in accuracy.