All Fingers are not Equal: Intensity of References in Scientific Articles

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1348–1358, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics All Fingers are not Equal: Intensity of References in Scientific Articles Tanmoy Chakraborty Dept. of Computer Science & UMIACS University of Maryland, College Park, USA tanchak@umiacs.umd.edu Ramasuri Narayanam IBM Research, India ramasurn@in.ibm.com Abstract Research accomplishment is usually measured by considering all citations with equal impor- tance, thus ignoring the wide variety of pur- poses an article is being cited for. Here, we posit that measuring the intensity of a refer- ence is crucial not only to perceive better un- derstanding of research endeavor, but also to improve the quality of citation-based applica- tions. To this end, we collect a rich annotated dataset with references labeled by the inten- sity, and propose a novel graph-based semi- supervised model, GraLap to label the in- tensity of references. Experiments with AAN datasets show a significant improvement com- pared to the baselines to achieve the true labels of the references (46% better correlation). Fi- nally, we provide four applications to demon- strate how the knowledge of reference inten- sity leads to design better real-world applica- tions. 1 Introduction With more than one hundred thousand new schol- arly articles being published each year, there is a rapid growth in the number of citations for the rel- evant scientific articles. In this context, we high- light the following interesting facts about the pro- cess of citing scientific articles: (i) the most com- monly cited paper by Gerard Salton, titled “A Vector Space Model for Information Retrieval” (alleged to have been published in 1975) does not actually ex- ist in reality (Dubin, 2004), (ii) the scientific authors read only 20% of the works they cite (Simkin and Roychowdhury, 2003), (iii) one third of the refer- ences in a paper are redundant and 40% are perfunc- tory (Moravcsik and Murugesan, 1975), (iv) 62.7% of the references could not be attributed a specific function (definition, tool etc.) (Teufel et al., 2006). Despite these facts, the existing bibliographic met- rics consider that all citations are equally significant. In this paper, we would emphasize the fact that all the references of a paper are not equally influ- ential. For instance, we believe that for our current paper, (Wan and Liu, 2014) is more influential refer- ence than (Garfield, 2006), although the former has received lower citations (9) than the latter (1650) so far1. Therefore the influence of a cited paper com- pletely depends upon the context of the citing paper, not the overall citation count of the cited paper. We further took the opinion of the original authors of few selective papers and realized that around 16% of the references in a paper are highly influential, and the rest are trivial (Section 4). This motivates us to design a prediction model, GraLap to automati- cally label the influence of a cited paper with respect to a citing paper. Here, we label paper-reference pairs rather than references alone, …

Content truncated to 500 words