Contrasting and Corroborating Citations in Journal Articles

Proceedings of Recent Advances in Natural Language Processing, pages 460–466, Hissar, Bulgaria, 7-13 September 2013. Contrasting and Corroborating Citations in Journal Articles Adam Meyers New York University meyers@cs.nyu.edu Abstract Automatic recognition of CORROBO- RATE and CONTRAST relations between citations may enhance citation analysis. We describe a system that identifies these citation relations using predicate/argument and discourse structures. 1 Introduction The citation of publications has been used to mea- sure the impact of authors, publications, publish- ers, fields of study, etc., as represented in graphs in which nodes represent documents and edges con- nect documents if one refers to the other (citation graphs) or if a third document refers to both (co- citation graphs). NLP may be used to supplement these graphs with information about why docu- ments are cited. While previous work (Teufel et al., 2009; Athar, 2011; Athar and Teufel, 2012) record positive and negative sentiment about cited work, we record information about how cited doc- uments are compared to each other. CONTRAST- ing documents may describe different approaches or different opinions. Documents which COR- ROBORATE each other may follow a single ap- proach. A document that is cited as corrobo- rating with many other documents may be very salient. Some example instances of these CON- TRAST/CORROBORATE relations are provided in Figure 1, the document containing the citations (represented as we or this study) contrasts with [1] and corroborates [2]; [3] and [4] contrast with each other; and [5] and [6] corroborate. We use square brackets and numbers (IEEE style) to represent citations that are the object of this study (Figure 1). We use last name plus date (APA style) for works cited as part of this re- search effort. Examples in this paper (modified for brevity) are from the PubMed Central corpus1. As 1http://www.ncbi.nlm.nih.gov/pmc/ 1. In contrast to [1], we found clear detrimental effects of prophylaxis. 2. This study corroborates a study [2] finding no evidence of cross-hemisphere invasions. 3. FluA and FluB viruses have a common origin [3]. Thus, it is expected that aa residues of PA are conserved between FluA and FluB [4]. 4. Some species shed lots of virus, yet suffer few damaging effects [5]. On the other hand, species of swan, show 100% mortality within days of inoculation with HPAIV (H5N1) [6]. Figure 1: CORROBORATE and CONTRAST in Figure 1, arguments of relations are in bold and signals are underlined. Other important elements (Figure 3) have boxes drawn around them. In this paper, we explore the relations be- tween sentence-internal predicate relations, inter- sentential discourse relations and relations be- tween citations. Then we describe and evalu- ate a system which derives CORROBORATE and CONTRAST relations between citations. 2 How Citation Relations are Encoded 2.1 The phrase/citation connection (Abu Jbara and Radev, 2012) describes a system for identifying the referential scope of each ci- tation, a text fragment that the citation is seman- tically related to–multiple citations can share the same referential scope. We assume that arguments of grammatical and discourse relations are essen- tially …

Content truncated to 500 words