Topic Modelling in Bangla Language: An LDA Approach to Optimize Topics and News Classification

Abstract

Topic modeling is a powerful technique for unsupervised analysis of large document collections. Topic models have a wide range of applications including tag recommendation, text categorization, keyword extraction and similarity search in the text mining, information retrieval and statistical language modeling. The research on topic modeling is gaining popularity day by day. There are various efficient topic modeling techniques available for the English language as it is one of the most spoken languages in the whole world but not for the other spoken languages. Bangla being the seventh most spoken native language in the world by population, it needs automation in different aspects. This paper deals with finding the core topics of Bangla news corpus and classifying news with similarity measures. The document models are built using LDA (Latent Dirichlet Allocation) with bigram.

LSA & LDA topic modeling classification: comparison study on e-books

“…The pre-processing operations were performed, which included cleaning the text from the characters and numbers, removing the stop signs and special symbols, removing the insignificant words, then reconfiguration data for representation unstructured data in a bag of words (corpus ). The second step is training the data by using two methods from topic modeling (LSA, LDA) method with different number of topics (10,15,20) in order to see which one will give the best performance from these two methods with our database. The keywords for each topic are used as features to classify the books.…”

“…(Corpus, dictionary, and a number of topics) are needed to train the (LDA and LSA) model, where each word in the corpus of vocabulary is then connected with one or more topics with a probability, as estimated by the model. (LDA, LSA) model is built with (10,15,20) various topics where each topic is a mixture of keywords and each keyword contributes a certain weight to the topic.…”

“…Rajasundari et al employed three different machine learning methods (Naive Bayes, K-NN and K-means) and topic modeling techniques Latent Dirichlet Allocation (LDA), Latent Semantic Analysis (LSA) for BBC news dataset, where topic modeling demonstrated its ability to detect hidden topics and bond between words and documents in addition to work better with many probability distributions [14]. LDA (Latent Dirichlet Allocation) was used by Mouhoub & Al Helal in [15] and bigram for obtaining the core topics of Bangla language news corpus and classifying news, where it proposed the first ever topic modeling tool for Bangla and this core is a challenge because the research in Bangla is not repeated due to insufficient datasets, unorganized grammar rules.…”