Error Detection of CRF-Based Bibliography Extraction from Reference Strings

Abstract

We proposed a parsing method for reference strings usually listed at the end of research papers to extract important bibliographies such as a title from them. The method uses a conditional random field (CRF) to estimate the correct bibliographic label for each token in the token sequence generated from a reference string. Although we achieved reasonable parsing accuracies for a Japanese academic journal, errors are inevitable. Therefore, this paper proposes ways to increase confidence for CRF-based bibliography parsing to detect such parsing errors. This paper also reports an empirical evaluation of the proposed parsing on the basis not only of its accuracies but also of how easy it is to detect errors. The experiments showed that the proposed measures reasonably indicated parsing errors and could be used to improve the quality of extracted bibliographies at a moderate manual post-editing cost.