Characterizing In-Text Citations Using N-Gram Distributions

Introduction This article focuses on a Natural Language Processing (NLP) approach for the analysis of citation functions in scientific papers. Bibliometric studies traditionally rely on citation metadata and count the number of times a publication has been cited. However, some recent studies rely also on full text processing on papers, e.g. (Boyack et al., 2013), (Bertin et al., 2013, 2014). The full text content of papers and more specifically the sentences containing citations provide valuable information on the functions of citations that can be exploited through NLP. To study citation acts, we need to consider full text papers and their rhetorical structure. The main question that we want to answer here is whether the most frequent citation patterns are correlated to the rhetorical structure of scientific papers. We investigate the properties of the linguistic patterns that appear in citation contexts. For this, we study the distribution of n-gram classes containing verb forms, and we show the existence of three different types of distributions according to the rhetorical structure.