Arabic Short Text Analytics using Multivariate Filter Methods and ProdLDA Model

Research in the domain of Arabic short text analytics remains constrained. Although there have been several research conducted on other languages, there is still a lack of comprehensive exploration of short texts in Arabic. However, it is crucial to derive practical and useful insights from these Arabic manuscripts. This study delves into investigating the efficacy of diverse topic modeling techniques, particularly the ProdLDA model, in clustering Arabic short document datasets. The study compares the performance of ProdLDA alongside BTM, GS-DMM, and BERT, revealing intriguing findings regarding their effectiveness. Firstly, the integration of term weighting models significantly improves the clustering performance of the Latent Dirichlet Allocation (LDA) model, underscoring the importance of word importance in short text analysis. Secondly, ProdLDA emerges as the leading model, consistently surpassing others across various evaluation metrics such as Purity and Normalized Mutual Information (NMI). For example, ProdLDA demonstrates remarkable improvements of approximately 1%-9% in Purity over BTM and 1%-6% over GSDMM, showcasing its proficiency in handling short text datasets for analytical purposes. These findings underscore the pivotal role of advanced text analytics techniques in uncovering meaningful patterns and insights within Arabic short text data.