Hierarchical Rubrication of Text Documents

Abstract

Topic modeling is an important and widely used method in the analysis of a large collection of documents. It allows us to digest the content of documents by examination of the selected topics. It has drawbacks such as a need to specify the number of topics. The topics can become too local or too global, depending on that number. Also, it does not provide a relation between local and global topics. Here we present an algorithm and a computer program for the hierarchical rubrication of text documents. The program solves these problems by creating a hierarchy of automatically selected topics in which local topics are connected of the global topics. The program processes PDF documents split them into text segments, builds vector representations using word2vec model and stores them in a database. The vector embeddings are structured in the form of a hierarchy of automatically constructed categories. Each category is identified by automatically selected keywords. The result is visualized in an interactive map. Traversing the hierarchy of topics is done by zooming the map. An analysis of the constructed hierarchy of categories allows us to evaluate the minimum and maximum depth of the hierarchy corresponding to a minimum and a maximum number of different topics contained in the collection of documents. The program was tested on documents on deep nuclear waste disposal.