PGLDA: enhancing the precision of topic modelling using poisson gamma (PG) and latent dirichlet allocation (LDA) for text information retrieval

• Main Author: Bakari, Ibrahim Bala
• Format: Thesis
• Language: English; English; English
• Published: 2021
• Subjects: QA76 Computer software; T Technology (General)
• Online Access: http://eprints.uthm.edu.my/4890/1/24p%20IBRAHIM%20BALA%20BAKARI.pdf; http://eprints.uthm.edu.my/4890/2/IBRAHIM%20BALA%20BAKARI%20COPYRIGHT%20DECLARATION.pdf; http://eprints.uthm.edu.my/4890/3/IBRAHIM%20BALA%20BAKARI%20WATERMARK.pdf; http://eprints.uthm.edu.my/4890/

The Poisson document length distribution has been used extensively in the past for modeling topics with the expectation that its effect will disintegrate at the end of the model definition. This procedure often leads to down Playing word correlation with topics and reducing retrieved documents' precision or accuracy. The existing document model, such as the Latent Dirichlet Allocation (LDA) model, does not accommodate words' semantic representation. Therefore, in this thesis, the PoissonGamma Latent Dirichlet Allocation (PGLDA) model for modeling word dependencies in topic modeling is introduced. The PGLDA model relaxes the words independence assumption in the existing Latent Dirichlet Allocation (LDA) model by introducing the Gamma distribution that captures the correlation between adjacent words in documents. The PGLDA is hybridized with the distributed representation of documents (Doc2Vec) and topics (Topic2Vec) to form a new model named PGLDA2Vec. The hybridization process was achieved by averaging the Doc2Vec and Topic2Vec vectors to form new word representation vectors, combined with topics with the largest estimated probability using PGLDA. Model estimations for PGLDA and PGLDA2Vec models were achieved by combining the Laplacian approximation of log-likelihood for PGLDA and Feed-Forward Neural Network (FFN) approaches of Doc2Vec and Topic2Vec. The proposed PGLDA and the hybrid PGLDA2Vec models were assessed using precision, micro F1 scores, perplexity, and coherence score. The empirical analysis results using three real-world datasets (20 Newsgroups, AG'News, and Reuters) showed that the hybrid PGLDA2Vec model with an average precision of 86.6%, and an average F1 score of 96.3%, across the three datasets is better than other competing models reviewed.