An Exploratory Analysis of GSDMM and BERTopic on Short Text Topic Modelling

Abstract

Topic models may be a useful tool for locating latent subjects in collections of documents. Short text clustering has become a more important task as social networking sites like Twitter have gained popularity. Short text is characterised by high sparsity, high-dimensionality, and large-volume. These characteristics are challenging to overcome. Two of the most well-known short text modelling algorithms are BERTopic and the Gibbs Sampling Dirichlet Multinomial Mixture Model (GSDMM). GSDMM is a topic model which can infer the count of topic clusters automatically with a good compromise between the fullness and uniformity of the clustering results, and is fast to converge. BERTopic is a neural topic model that extracts coherent topic representations based on the semantic similarity of words and phrases in the and clustering with the help of a class-based form of TF-IDF. We compare these two algorithms in this paper to determine which model is more effective in short text topic modelling.