Bayesian Text Segmentation for Index Term Identification and Keyphrase Extraction

Proceedings of COLING 2012: Technical Papers, pages 2077–2092, COLING 2012, Mumbai, December 2012. Bayesian Text Segmentation for Index Term Identification and Keyphrase Extraction David Newman,♦ Nagendra Koilada,♦ Jey Han Lau♥♠ and Timothy Baldwin♥♠ ♦ Dept of Computer Science, University of California, Irvine, USA ♥ Dept of Computing and Information Systems, The University of Melbourne, Australia ♠ NICTA Victoria Research Laboratory, Australia newman@uci.edu, nkoilada@uci.edu, jhlau@csse.unimelb.edu.au, tb@ldwin.net ABSTRACT Automatically extracting terminology and index terms from scientific literature is useful for a variety of digital library, indexing and search applications. This task is non-trivial, compli- cated by domain-specific terminology and a steady introduction of new terminology. Correctly identifying nested terminology further adds to the challenge. We present a Dirichlet Process (DP) model of word segmentation where multiword segments are either retrieved from a cache or newly generated. We show how this DP-Segmentation model can be used to successfully extract nested terminology, outperforming previous methods for solving this problem. KEYWORDS: Dirichlet Process, segmentation, Bayesian learning, index term, keyphrase ex- traction. 2077 1 Introduction Index terms and keyphrases are widely used to enhance the browsing/accessibility of docu- ments, especially scientific literature (Gutwin et al., 1999). Index terms are the main terms and concepts in a document, as would be contained in the index of a book or edited collection of papers, for example. They typically number in the hundreds for a single volume, and are conventionally annotated for their key occurrences in the document. Keyphrases are assigned to each document in the form of a handful of terms which capture the key topics or concepts covered in the document. Keyphrases are commonly used as a means of clustering/linking documents in digital libraries. While there is no question of the utility of both index terms and keyphrases, annotating a document with index terms is an arduous and largely unreward- ing task, and annotators are notoriously inconsistent at index term and keyphrase assignment (Leonard, 1977; Kim et al., 2010). As such, there is a real need for computational methods which both speed up and homogenize the assignment of index terms and keyphrases. The focus of this paper is an unsupervised methodology for automatically identifying both index terms in a document collection and keyphrases for a single document. The method- ology is called Dirichlet Process Segmentation (DP-seg) and was originally developed by Goldwater et al. (2009) for the task of word segmentation over phoneme streams, in the con- text of child language acquisition. DP-seg is particularly well suited to the tasks of index term identification and keyphrase extraction because: (1) it is able to identify individual type-level occurrences of a given term; (2) it is context-sensitive and able to robustly deal with “nested” multiword terminology (e.g. are support vector and vector machine also index terms, or just support vector machine?); (3) it is highly sensitive and can be applied to small-scale corpora (important for keyphrase extraction, where the input may be a single document); and (4) it is highly parameterizable. Our contributions in this paper are: (1) the …

Content truncated to 500 words