Towards Fine-grained Citation Function Classification

Proceedings of Recent Advances in Natural Language Processing, pages 402–407, Hissar, Bulgaria, 7-13 September 2013. Towards Fine-grained Citation Function Classification Xiang Li Yifan He Adam Meyers Computer Science DepartmentRalph Grishman New York University {xiangli, yhe, meyers, grishman}@cs.nyu.edu Abstract We look into the problem of recogniz- ing citation functions in scientific liter- ature, trying to reveal authors’ rationale for citing a particular article. We intro- duce an annotation scheme to annotate ci- tation functions in scientific papers with coarse-to-fine-grained categories, where the coarse-grained annotation roughly cor- responds to citation sentiment and the fine- grained annotation reveals more about ci- tation functions. We implement a Maxi- mum Entropy-based system trained on an- notated data under this scheme to auto- matically classify citation functions in sci- entific literature. Using combined lex- ical and syntactic features, our system achieves the F-measure of 67%. 1 Introduction Citations in scientific papers serve different pur- poses, from comparing one work to another to acknowledging the inventor of certain concepts. Recognizing citation functions is important for un- derstanding the structure of a single scientific doc- ument as well as mining citation graphs within a document collection. Therefore, this task has attracted researchers from the fields of discourse analysis, sociology of science, and information sciences for decades (Teufel et al., 2006a). Most of the existing research in this area fo- cused on the analysis of citation sentiment, which has achieved good accuracy (see, e.g., (Teufel et al., 2006a)). Citation sentiment analysis systems are usually able to identify positive, neutral, or negative opinions, but if we want to better under- stand the exact function of a citation, we need to know not only whether the authors like the cita- tion, but also how the citation is used in a given context (Section 2). In this paper, we try to reveal citation func- tions more accurately than simply classifying ci- tation sentiment. We first create a two level coarse-to-fine grained annotation scheme (Sec- tion 3). The coarse-level annotation corre- sponds roughly to sentiment categories, includ- ing POSITIVE, NEGATIVE, and NEUTRAL. The fine-grained annotation scheme provides a more detailed description of citation functions, such as Significant, which asserts the importance of an article or a work, and Discover, which ac- knowledges the original discoverer/inventor of a method or material. Using data annotated under this scheme, we train classifiers to determine citation functions, and experiment with features from lexical to syn- tactic levels (Section 4). We predict the fine- grained citation function at 67% in F-measure in our experiments, which is at the same level as the coarse-grained citation sentiment classification (Section 5). 2 Related Work The background for our work is in citation analy- sis. Applications of citation analysis include eval- uating the impact of a published literature through a measurable bibliometric (Garfield, 1972; Luukkonen, 1992; Borgman and Furner, 2002), analyzing bibliometric networks (Radev et al., 2009), summarizing scientific papers (Qazvinian and Radev, 2008; Abu-Jbara and Radev, 2011), generating surveys of scientific paradigms (Mo- hammad et al., 2009), among others. Correctly and …

Content truncated to 500 words