Towards employing native information in citation function classification

Citations play a fundamental role in supporting authors’ contribution claims throughout a scientific paper. Labelling citation instances with different function labels is indispensable for understanding a scientific text. A single citation is the linkage between two scientific papers in the citation network. These citations encompass rich native information, including context of the citation, citation location, citing and cited paper titles, DOI, and the website’s URL. Nevertheless, previous studies have ignored such rich native information during the process of datasets’ accumulation, thereby resulting in a lack of comprehensive yet significantly valuable features for the citation function classification task. In this paper, we argue that such important information should not be ignored, and accordingly, we extract and integrate all of the native information features into different neural text representation models via trainable embeddings and free text. We first construct a new dataset entitled, NI-Cite, comprising a large number of labelled citations with five key native features (Citation Context, Section Name, Title, DOI, Web URL) against each dataset instance. In addition, we propose to exploit the recently developed text representation models integrated with such information to evaluate the performance of citation function classification task. The experimental results demonstrate that the native information features suggested in this paper enhance the overall classification performance.