A hierarchical topic modelling approach for tweet clustering

Abstract

While social media platforms such as Twitter can provide rich and up-to-date information for a wide range of applications, manually digesting such large volumes of data is difficult and costly. Therefore it is important to automatically infer coherent and discriminative topics from tweets. Conventional topic models and document clustering approaches fail to achieve good results due to the noisy and sparse nature of tweets. In this paper, we explore various ways of tackling this challenge and finally propose a two-stage hierarchical topic modelling system that is efficient and effective in alleviating the data sparsity problem. We present an extensive evaluation on two datasets, and report our proposed system achieving the best performance in both document clustering performance and topic coherence.