Using Full-text of Academic Articles to Find Software Clusters

Scientific software is making contributions to modern science. To meet huge academic demands such as data analysis, modelling, visualization and so on, various software has been developed to help different steps in scientific work. In order to reveal the connections between scientific software, we conduct cluster analysis among scientific software based on the full-text data of 23,120 articles published in PLOS ONE. Firstly, we select some popular software whose mention times are over 50 to be our candidate software list for clustering analysis. Secondly, Word2Vec is applied to learn distributed representation for each software. Then, we apply Affinity Propagation to cluster software and tune different parameters to obtain better results. Silhouette coefficient is computed here to evaluate clustering performance under each parameter setting. According to our optimal results, software clusters with specific functions can be found. And software which have strong linkage between each other are mainly have functions in common.