Automatic label curation from large-scale text corpus

Abstract

The topic modeling technique extracts themes based on their probabilistic measurements from any large-scale text collection. Even though topic modeling pulls out the most important phrases that describe latent themes in text collections, a suitable label has yet to be found. Interpreting the topics extracted from a text corpus and identifying a suitable label automatically reduces the cognitive load for the analyst. Extractive methods are used typically to select a label from a given candidate set, based on probability metrics for each candidate set. Some of the existing approaches use phrases, words, and images to generate labels using frequency counts of different words in the text. The paper proposes a method to generate labels automatically to represent each topic based on a labeling strategy to filter candidate labels and then apply sequence-to-sequence labelers. The objective of the method is to get a meaningful label for the result of the Latent Dirichlet Allocation algorithm. The BERTScore metric is used to evaluate the effectiveness of the proposed method. The proposed method generates good interpretative labels as compared to baseline models for topic words or terms automatically. The comparison with the label generated through ChatGPT API shows the quality of the generated label with the experiment performed on Four Datasets NIPS, Kindle, PUBMED, and CORD-19.