Citation Role Labeling via Local, Pairwise, and Global Features.

The citation relationship between scientific publications has been successfully used for bibliometrics, information retrieval and data mining tasks, and citation-based recommendation algorithms are well documented. While previous studies investigated citation relationships from various viewpoints, most of them share the same assumption that, if paper1 cites paper2 (or author1 cites author2), they are connected, regardless of citation importance, sentiment, reason, topic, or motivation. However, this assumption is oversimplified. In this study, we propose a novel method to automatically label the massive citations in the scientific repository with different roles, a.k.a. citation role labeling. Unlike earlier studies, we employ pairwise features (similarity between citing and cited paper) and global features (citing and cited paper proximity on the heterogeneous graph), in addition to local features (information extracted solely from the citing paper, e.g. citation textual context). Evaluation result shows pairwise and global features, if properly used, can be very helpful to enhance the citation role labeling performance, especially when full-text data is not readily available. [ABSTRACT FROM AUTHOR]