Ranking-based cited text identification with highway networks

Abstract

In recent years, content-based citation analysis (CCA) has attracted great attention, which focuses on citation texts within full-text scientific articles to analyze the meaning of each citation. However, citation texts often lack the appropriate evidence and context from cited papers and are sometimes even inaccurate. Thus it is necessary to identify the corresponding cited text from a cited paper and examine which part of the content of the paper was cited in a citation. In this study, we proposed a novel ranking-based method to identify cited texts. This method contains two stages: similarity-based unsupervised ranking and deep learning-based supervised ranking. A novel listwise ranking model was developed with the use of 36 similarity features and 11 section position features. Firstly, top-5 sentences were selected for each citation text according to a modified Jaccard similarity metric. Then the selected sentences were ranked using the trained listwise ranking model, and top-2 sentences were selected as cited sentences. The experiments showed that the proposed method outperformed other classification-based and voting-based identification methods on the test set of the CL-SciSumm 2017.