In this paper we focus of the clustering of citation contexts in scienti c papers. We use two methods, k-means and hierarchical clustering to better understand the phenomenon and types of citations and to explore the multidimensional nature of the elements composing the contexts of citations in di erent sections of the papers. We have analyzed a data set of seven peer-reviewed academic journals published by PLOS. The obtained clusters show that the Methods section is speci c in nature, regardless of the journal. A proximity between some of the journals can be observed.
A lot of research currently takes place around citation contexts in scienti c papers. Although these themes are not recent, there is a renewed interest in this eld with the implementation of di erent technics that come from text-mining. The main challenge of these studies is to propose a method to analyze citation contexts at a large scale taking into account various criteria.
We propose a multidimensional approach to this problem which is based on
clusters. Clustering algorithms allow us to select similar contexts, that should be
considered members of a cluster. This study provides new results around citation
contexts and is related to two previous studies on similar problems [
To perform this study we have analyzed a data set of seven peer-reviewed academic journals published in Open Access by the Public Library of Science (PLOS). Six of the journals are domain-speci c (PLOS Biology, PLOS Computational Biology, PLOS Genetics, PLOS Medicine, PLOS Neglected Tropical Diseases and PLOS Pathogens ) and the 7th is PLOS ONE, which is a general journal that covers all elds of science and social sciences. We have used for our experiments the entire data set of about 80,000 research articles in full text published up to September 2013.
The data set is in the XML JATS format, where the sections and paragraphs
that are identi ed as distinct XML elements, as well as the in-text references that
linked to the corresponding elements in the bibliography of the article. Various
aspects of the processing of this corpus and the distributions of in-text references
and thier contexts with respect to the IMRaD structure have been the object of
previous studies [
We have considered the articles in the corpus that are organised following the IMRaD structure (Introduction, Methods, Results and Discussion). As this is part of the editorial requirements of the journals, the vast majority of the papers share this structure. For each journal and for each of the four section types, we have extracted a random sample of 1000 sentences that contain in-text citations. These sentences will be considered as citation contexts in our experiment.
The pre-processing of the corpus consists in removing all punctuation marks and numerical values so as not to introduce bias as to the overrepresentation of the bibliographic references. We also used a stopword list and the terms obtained were stemmatized. 2.3
We analyse citation contexts using a multidimesional approach. We used two
complementary approaches, hierarchical clustering and K-means, that allow us
to better understand the phenomenon and types of citations and to explore the
multidimensional nature of the elements composing the contexts of citations. To
obtain the clusters we use the method of K-means (see [
From a methodological point of view, we have used popular partitioning method: K-means clustering. However, this approach requires to know the exact number of clusters. To determine this number, we have used a graph that shows the relation between the number of extracted clusters and their distance. This method allows us to obtain the value and to generate the clusters using the Kmeans method. The correct choice of k can be ambiguous. The di culty arises from the fact that we have to nd a balance between the shape and scale of the data set and the number of clusters that the user wants to obtain. We propose in this paper two approaches for diagnosing the number of clusters suitable for the data (see gure 1 and 2). The result on gure 1 uses elbow with the sum of squared error and gure 2 uses Calinsky criterion with an interval for groups between one and ten. The Elbow method is a method for interpreting and validating coherence in clusters to nd the appropriate number of clusters in a data set. 3
The results obtained from the K-means clustering and Hierarchical Clustering
are respectively presented in the gures 3 and 4. Figure 3 shows an analysis
of the data set with an arbitrary choice of the number of clusters k = 4. The
names of the journals are coded using 4 characters, followed by one character (i,
m, r or d) that corresponds to the section type. This gure shows the speci c
character of the Methods section (at the left), con rming earlier works [
The hierarchical clustering allows us to illustrate the hierarchical organisation of groups as shown on the gure 4. This visualization con rms the previous result, but o ers also a hierarchical view of the clusters. The hierarchical analysis highlights the speci c character of the Results sections in pmed and pntd that are in the same group as the Introduction and Discussion sections in these journals. Again, ppat appears in a separate cluster. The journal pcbi (PLOS Computational Biology) is also in a separate sub-cluster that suggests that the citation contexts in this journal are quite di erent from those in the other journals. 4
This work emphasizes the need for a tool to identify and analyze the contexts of
citations, while being aware of the multidimensional nature of these phenomena.
Indeed, we have not yet addressed the problem of the categorization of
quotations. Our approach aims to determine the clusters of citation contexts, so that
at a di erent stage the topics will have to be identi ed and processed for a ner
analysis. In order to do this, we can consider for example the syntax of citation
contexts and the topics, as already proposed in [
Studying the structure of scienti c papers and observing the regularities in the contexts of in-text citations is an important step towards understanding the phenomenon of citation which is central in the process of building scienti c knowledge. Di erent types of citations exist based on the motivation to cite and the relation between the citing authors and the cited work. To be able to create an ontology of citations that re ects the types of citations found in articles it is necessary to process existing corpora and study the properties of citation contexts on a large scale.
We thank Benoit Macaluso of the Observatoire des Sciences et des Technologies (OST), Montreal, Canada, for harvesting and providing the PLOS data set.