PositionRank: An Unsupervised Approach to Keyphrase Extraction from Scholarly Documents

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1105–1115 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1102 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1105–1115 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1102 PositionRank: An Unsupervised Approach to Keyphrase Extraction from Scholarly Documents Corina Florescu and Cornelia Caragea Computer Science and Engineering University of North Texas, USA CorinaFlorescu@my.unt.edu, ccaragea@unt.edu Abstract The large and growing amounts of online scholarly data present both challenges and opportunities to enhance knowledge dis- covery. One such challenge is to auto- matically extract a small set of keyphrases from a document that can accurately de- scribe the document’s content and can fa- cilitate fast information processing. In this paper, we propose PositionRank, an unsupervised model for keyphrase extrac- tion from scholarly documents that incor- porates information from all positions of a word’s occurrences into a biased PageR- ank. Our model obtains remarkable im- provements in performance over PageR- ank models that do not take into account word positions as well as over strong base- lines for this task. Specifically, on several datasets of research papers, PositionRank achieves improvements as high as 29.09%. 1 Introduction The current Scholarly Web contains many millions of scientific documents. For example, Google Scholar is estimated to have more than 100 million documents. On one hand, these rapidly-growing scholarly document collections offer benefits for knowledge discovery, and on the other hand, find- ing useful information has become very challeng- ing. Keyphrases associated with a document typi- cally provide a high-level topic description of the document and can allow for efficient information processing. In addition, keyphrases are shown to be rich sources of information in many natu- ral language processing and information retrieval tasks such as scientific paper summarization, clas- sification, recommendation, clustering, and search (Abu-Jbara and Radev, 2011; Qazvinian et al., 2010; Jones and Staveley, 1999; Zha, 2002; Zhang et al., 2004; Hammouda et al., 2005). Due to their importance, many approaches to keyphrase extrac- tion have been proposed in the literature along two lines of research: supervised and unsupervised (Hasan and Ng, 2014, 2010). In the supervised line of research, keyphrase extraction is formulated as a binary classification problem, where candidate phrases are classified as either positive (i.e., keyphrases) or negative (i.e., non-keyphrases) (Frank et al., 1999; Hulth, 2003). Various feature sets and classification algorithms yield different extraction systems. For example, Frank et al. (1999) developed a system that ex- tracts two features for each candidate phrase, i.e., the tf-idf of the phrase and its distance from the be- ginning of the target document, and uses them as input to Na¨ıve Bayes classifiers. Although super- vised approaches typically perform better than un- supervised approaches (Kim et al., 2013), the re- quirement for large human-annotated corpora for each field of study has led to significant attention towards the design of unsupervised approaches. In the unsupervised line of research, keyphrase …

Content truncated to 500 words