Methods of Unsupervised Semantic Analysis of Small and Medium-Sized Corpora

Abstract

Analysis and description of text corpora can present a number of technical challenges, especially in the case of corpora built by automated content extraction that may not allow for readily available annotations and other semantic information about the texts. In this work we describe and test an approach to analysis of the semantic content of corpora based on the methods of unsupervised feature extraction, dimensionality reduction and concept learning. With model corpora represented by texts in English newsgroups, we demonstrate how characteristic semantic types can be identified with methods of unsupervised machine learning and clustering. The results can be an instrumental addition to methods of analysis of semantic context of text corpora where the prior description such as annotations may not be available or is scarce. The approach and methods demonstrated in this work are in no way limited to the English language and can be applied to corpora in any language where the appropriate vectorization and preprocessing methods are available.