Citation Segmentation from Sparse & Noisy Data: An Unsupervised Joint Inference Approach with Markov Logic Networks

Citation Segmentation in a Digital Humanities Context. Bibliographies are an important resource for scientific research. Their storage in (online) bibliographic databases offers efficient search functionalities for wide-spread and timely use in international research communities. For this purpose it is crucial to automatically detect the inherent structure of bibliographic references, by isolating and extracting citation subfields (e.g., author, title, venue). Previous approaches in citation segmentation strongly rely on language-specific lexical data and multiple occurrences of the same citation entry in online publication repositories. However, when dealing with multilingual data, the use of language-specific knowledge becomes difficult. Moreover, self-contained data sources like printed bibliographies are naturally short of recurring citation entries, and thus cannot rely on data redundancy. In this work, we present an approach to citation segmentation that operates on sparse and noisy OCR input originating from a single, multilingual bibliography, the Turkology Annual (Turkologischer Anzeiger).1 The Turkology Annual is a bibliography for Turkology and Ottoman studies, comprising 28 volumes which are only available in print. Citation entries containing multiple languages and scripts, the shortage of citation redundancy, frequent OCR errors and inconsistencies in citation structure impede the use of state-of-the-art statistical approaches for citation segmentation.