A survey and an experimental comparison of methods for text clustering: application to scientific articles

Text documents clustering is used in many applications such as information retrieval, exploratory search, spam detection. This problem is the subject of many scientific papers, but the specificity of scientific articles in regards to the clustering efficiency remains to be studied insufficiently; in particular, if all documents belong to the same domain or if full texts of articles are unavailable. This paper presents an overview and an experimental comparison of text clustering methods in application to scientific articles. We study methods based on bag of words, terminology extraction, topic modeling, word embedding and document embedding obtained by artificial neural networks (word2vec, paragraph2vec).