Systematic Comparison of Vectorization Methods in Topic Classification for Myanmar Language

Text classification which is a research area of text data mining is widely used and one of the popular tasks of Natural Language Processing (NLP). But very few had focused on the effective use of vectorization methods which are vital in classification systems. In this paper, three widely used vectorizers for topic classification are discussed. The main goal is to evaluate the effectiveness of three different vectorization techniques such as TF-IDF Vectorizer (TFID), Count Vectorizer, and Word2Vec Vectorizer in the context of Myanmar news topic classification. The study focuses on a dataset comprising Myanmar news articles categorized into four distinct topics: Politics, Entertainment, Crime, and Business. Each category has 180 unique articles and in total 720 articles are prepared. Furthermore, these articles undergo segmentation and preprocessing. Each vectorization technique is applied through experimentation on an RBF (Radial Basis Function) Support Vector Machine (SVM) model. Moreover, feature selection via Chi Square algorithm is applied to optimize performance. Results indicate that TFID outperforms the other vectorizers, as seen by its accuracy of 83% on the test set, compared to 75% for the Count vectorizer and 48% for the Word2Vec. Overall, this research contributes insights into the effective utilization of NLP techniques for Myanmar news classification by emphasizing the advantage of TFID Vectorization compared to other vectorization techniques.