2023 Volume 31 Pages 353-364
Human-in-the-loop topic models allow users to encode feedback to modify topic models without changing the core machinery of the topic models. Basic refinement functions have been proposed in prior works in which the main focus was to modify the top word lists of topics, e.g., add a single word in a topic having distribution over a large vocabulary set. In this work, we point out that such refinements have very little to no effect on document-topic associations, which are rather important in practical applications, and propose keyphrase-based refinement functions that are designed to improve document-topic associations efficiently. In the proposed method, these keyphrases are extracted by using a neural keyphrase generation model that summarizes a document in a few keyphrases which are human-interpretable representations of each given document. The proposed refinement functions are as simple as word-based refinements but directly modify the document-topic association in functionality by referring to the keyphrase representations of documents. To examine the capability of the refinement functions for revising topic models, we conducted experiments based on a simulated user that has a fixed preference for the document-topic association using 20Newsgroups dataset. Our results showed that the proposed keyphrase-based refinements outperform the basic word-based refinements in terms of the F1 score computed on the fixed preference.
