Co-Training for Topic Classification of Scholarly Data

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2357–2366, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Co-Training for Topic Classification of Scholarly Data Cornelia Caragea1, Florin Bulgarov1, Rada Mihalcea2 1Computer Science and Engineering, University of North Texas, TX, USA 2Computer Science and Engineering, University of Michigan, MI, USA ccaragea@unt.edu, FlorinBulgarov@my.unt.edu,mihalcea@umich.edu Abstract With the exponential growth of scholarly data during the past few years, effective methods for topic classification are greatly needed. Current approaches usually re- quire large amounts of expensive labeled data in order to make accurate predictions. In this paper, we posit that, in addition to a research article’s textual content, its ci- tation network also contains valuable in- formation. We describe a co-training ap- proach that uses the text and citation infor- mation of a research article as two differ- ent views to predict the topic of an article. We show that this method improves sig- nificantly over the individual classifiers, while also bringing a substantial reduction in the amount of labeled data required for training accurate classifiers. 1 Introduction As science advances, scientists around the world continue to produce a large number of research ar- ticles, which provide the technological basis for worldwide dissemination of scientific discoveries. Online digital libraries such as Google Scholar, CiteSeerx, and PubMed store and index millions of such research articles and their metadata, and make it easier for researchers to search for scien- tific information. These libraries require effective and efficient methods for topic classification of re- search articles in order to facilitate the retrieval of content that is tailored to the interests of specific individuals or groups. Supervised approaches for topic classification of research articles have been developed, which generally use either the content of the articles (Caragea et al., 2011), or take into account the citation relation between research ar- ticles (Lu and Getoor, 2003). To be successful, these supervised approaches assume the availability of large amounts of labeled data, which require intensive human labeling ef- fort. In this paper, we explore a semi-supervised approach that can exploit large amounts of un- labeled data together with small amounts of la- beled data for accurate topic classification of re- search articles, while minimizing the human ef- fort required for data labeling. In the scholarly do- main, research articles (or papers) are highly inter- connected in giant citation networks, in which pa- pers cite or are cited by other papers. We posit that, in addition to a document’s textual content and its local neighborhood in the citation network, other information exists that has the potential to improve topic classification. For example, in a citation network, information flows from one pa- per to another via the citation relation (Shi et al., 2010). This information flow and the topical influ- ence of one paper on another are specifically cap- tured by means of citation contexts, i.e., short text segments surrounding a citation’s mention. These contexts are not arbitrary, but they of- ten serve as brief summaries …

Content truncated to 500 words