A Hybrid Generative/Discriminative Approach To Citation Prediction

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 75–83, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics A Hybrid Generative/Discriminative Approach To Citation Prediction Chris Tanner Brown University Providence, RI christanner@cs.brown.edu Eugene Charniak Brown University Providence, RI ec@cs.brown.edu Abstract Text documents of varying nature (e.g., sum- mary documents written by analysts or pub- lished, scientific papers) often cite others as a means of providing evidence to support a claim, attributing credit, or referring the reader to related work. We address the problem of predicting a document’s cited sources by introducing a novel, discriminative approach which combines a content-based generative model (LDA) with author-based features. Fur- ther, our classifier is able to learn the im- portance and quality of each topic within our corpus – which can be useful beyond this task – and preliminary results suggest its met- ric is competitive with other standard met- rics (Topic Coherence). Our flagship system, Logit-Expanded, provides state-of-the-art per- formance on the largest corpus ever used for this task. 1 Introduction The amount of digital documents (both online and offline) continues to grow greatly for several rea- sons, including the eagerness of users to gener- ate content (e.g., social media, Web 2.0) and the decrease in digital storage costs. Many different types of documents link to or cite other documents (e.g., websites, analyst summary reports, academic research papers), and they do so for various rea- sons: to provide evidence, attribute credit, refer the reader to related work, etc. Given the plethora of documents, it can be highly useful to have a sys- tem which can automatically predict relevant cita- tions, for this could (1) aid authors in citing rele- vant, useful sources which they may otherwise not know about; and (2) aid readers in finding useful documents which otherwise might not have been discovered, due to the documents’ being unpopu- lar or poorly cited by many authors. Specifically, we are interested in citation prediction – that is, we aim to predict which sources each report document cites. We define a report as any document that cites another document in our corpus, and a source as a document that is cited by at least one report. Natu- rally, many documents within a corpus can be both a report and a source. Note, we occasionally refer to linking a report and source, which is synonymous with saying the report cites the source. Citation prediction can be viewed as a special case of the more general, heavily-researched area of link prediction. In fact, past research mentioned in Section 2 refers to this exact task as both cita- tion prediction and link prediction. However, link prediction is a commonly used phrase which may be used to describe other problems not concerning documents and citation prediction. In these general cases, a link may be relatively abstract and repre- sent any particular relationship between other ob- jects (such as users’ interests or interactions). Tra- ditionally, popular techniques …

Content truncated to 500 words