Maintaining topic models for growing corpora

Abstract

A reference library can be described as a corpus of an individual composition of documents. Over time, the corpus might grow because an agent decides to extend its corpus with additional documents, e.g., new publications, or new articles. Existing approaches use topic modelling techniques to compare documents with each other within the same corpus by the documents' topic distribution. However, for new documents, only the text, and no topic distribution is available. Thus, this paper describes three techniques for estimating topic distributions of new unseen documents considering the initial documents in a corpus. Additionally, we present an extensive evaluation about the performance and runtime of the three topic modelling techniques for various scenarios and different sized corpora.