Improve topic modeling algorithms based on Twitter hashtags

• Published in: Journal of physics. Conference series Vol. 1660; no. 1; pp. 12100 - 12108
• Main Authors: Alash, Hayder M, Al-Sultany, Ghaidaa A
• Format: Journal Article
• Language: English
• Published: Bristol IOP Publishing 01.11.2020
• Subjects: Algorithms; Clustering; Coherence; Data mining; Dirichlet problem; Hashtag; Latent Dirichlet Allocation (LDA); Latent semantic analysis (LSA); Modelling; Physics; Social networks; Topic Derivation; Twitter; Unstructured data
• ISSN: 1742-6588; 1742-6596
• DOI: 10.1088/1742-6596/1660/1/012100

Today with increase using social media, a lot of researchers have interested in topic extraction from Twitter. Twitter is an unstructured short text and messy that it is critical to find topics from tweets. While topic modeling algorithms such as Latent semantic analysis (LSA) and Latent Dirichlet Allocation (LDA) are originally designed to derive topics from large documents such as articles, and books. They are often less efficient when applied to short text content like Twitter. Luckily, Twitter has many features that represent the interaction between users. Tweets have rich user-generated hashtags as keywords. In this paper, we exploit the hashtags feature to improve topics learned from Twitter content without modifying the basic topic model of LSA and LDA. Users who share the same hashtag at most discuss the same topic. We compare the performance of the two methods (LSA and LDA) using the topic coherence ( with and without hashtags). The experiment result on the Twitter dataset showed that LSA has better coherence score with hashtags than that do not incorporate hashtags. In contrast, our experiments show that the LDA has a better coherence score without incorporating hashtags. Finally, LDA has a better coherence score than LSA and the best coherence result obtained from the LDA method was (0.6047) and the LSA method was (0.4744) but the number of topics in LDA was higher than LSA. Thus, LDA may cause the same tweets to discuss the same subject set into different clustering.