Ensemble-style Self-training on Citation Classification

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 623–631, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Ensemble-style Self-training on Citation Classification Cailing Dong Ulrich Sch¨afer German Research Center for Artificial Intelligence (DFKI) GmbH Language Technology Lab, Campus D 3 1, D-66123 Saarbr¨ucken, Germany cherinedong@gmail.com ulrich.schaefer@dfki.de Abstract Classification of citations into categories such as use, refutation, comparison etc. may have several relevant applications for digital libraries such as paper browsing aids, reading recommendations, qualified citation indexing, or fine-grained impact factor calculation. Most citation classifi- cation approaches described so far heavily rely on rule systems and patterns tailored to specific science domains. We focus on a less manual approach by learning domain- insensitive features from textual, physical, and syntactic aspects. Our experiments show the effectiveness of this feature set with various machine learning algorithms on datasets of different sizes. Further- more, we build an ensemble-style self- training classification model and get bet- ter classification performance using only few training data, which largely reduces the manual annotation work in this task. 1 Introduction Still in the current age of Semantic Web, struc- tured repositories, ontologies and huge databases, scientific knowledge is preserved and transported in rather unstructured pieces of information, vulgo scientific papers or other similar, textual represen- tations. Except bibliographical metadata, (manu- ally) assigned keywords or coarse-grained classi- fications, there is no further, general method to un- lock the written treasures without tedious reading, while full-text search is generally considered too imprecise and may be misleading. Citations, spread over scientific text, are impor- tant anchors that help to structure the broad pub- lication space. They are of invaluable importance to beginners in a scientific field as they ultimately point to seminal, original work and knowledge not explicitly available or repeated in every pub- lication. Citations are also the primary discourse links in scientific discussion which typically span over years or even decades. Furthermore, citations are helpful to understand and reproduce findings. Thus, they form a predominant text feature for ev- ery reader. Wan et al. (2009), e.g., present a study on user needs for browsing scientific publications and show that citations play an important role, Sch¨afer and Kasterka (2010) suggest a novel user interface for navigating in typed citation graphs. Garfield (1965) is probably the first to discuss an automatic computation of a citation classifica- tion. Many studies on citation classification are generally derived from the four-dimensional cita- tion schema proposed by Moravcsik and Muruge- san (1975). They distinguish between confirma- tive vs. negational; conceptual (theory) vs. op- erational (method); evolutionary (build on cited work) vs. juxtapositional (alternative to cited work); and organic (necessary to understand, re- produce) vs. perfunctory (citation out of polite- ness, policy, piety). The number of different pro- posed classes (also called citation functions) varies from 3 to 35 (Garfield, 1965; Garzone, 1996; Mer- cer and DiMarco, 2004; Teufel et al., 2006; Har- wood, 2009). Most approaches try to identify more detailed dimensions and mutually exclusive classes by making use of …

Content truncated to 500 words