Comparing Topic Modeling Techniques for Identifying Informative and Uninformative Content: A Case Study on COVID-19 Tweets

Abstract

It is essential to understand what topics related to the COVID19 pandemic forms informative and uninformative content on social networks instead of general information (which contains both informative and uninformative). Uninformative content is mainly based on personal opinions and is more suitable for sentimental analysis. Whereas informative content is based on facts, figures, and reports; therefore, it is beneficial to gain a more in-depth understanding for a better strategic response to COVID-19. Despite knowing this fact, there is still a lack of study performed to investigate the aspects of informative content to gain an in-depth understanding of COVID-19 discussed topics. We aim to fill this gap through the study presented in this paper. We used the dataset containing 4719 “informative” and 5281 “uninformative” labeled tweets to realize informative aspects. Latent Dirichlet Allocation (LDA) and Latent Semantic Analysis (LSA) are popular topic modeling techniques. However, since both are based on an unsupervised approach, it is still unknown whether LDA or LSA effectively categorizes documents and how an appropriate number of topics can be determined. Therefore, we used both techniques to analyze tweets' content. Results show that LDA outperforms LSA by achieving a topic coherence score of 0.619 on uninformative and 0.599 on informative. In addition, based on LDA's results, it is also observed that most of the words that form informative content are death, case, coronavirus, people, confirmed, total, positive, tested, number, reported indicating tested, and death cases are the most concerned topics. On the other hand, words like immunity, fatality, protocol, thread, tourist, queue, blockade, eradication, prediction, detention, concerned are most likely to form uninformative content.