Cost Evaluation of CRF-Based Bibliography Extraction from Reference Strings

Abstract

The effective use of digital libraries demands maintenance of bibliographic databases. Especially, the reference fields of academic papers are full of useful bibliographic information such as authors’ names and paper titles. We, therefore, propose a method of automatically extracting bibliographic information from reference strings using a conditional random field (CRF). However, at least a few hundred reference strings are necessary for training the CRF to achieve high extraction accuracies. As described herein, we propose the use of active sampling and pseudo-training data to reduce the amount of training data. Then we evaluate the associated training costs by experimentation.