Bengali Text Categorization Based on Deep Hybrid CNN–LSTM Network with Word Embedding
Samrat Alam, Md Afnan Ul Haque, Ashiqur Rahman
Abstract
With the evolution of the internet and digital network, the amount of textual data is increasing gradually. As a result, it becomes increasingly difficult to categorize vast amounts of data manually. But, with the benefit of the machine learning process, we can categorize these massive amounts of text data automatically according to their contents. Compared with other languages such as English, text categorization in Bangla is one of the challenging tasks. Because a lack of resources is the reason for this. In this research work, we take data from open resources which are containing the Bangla newspaper article corpus with 12 categories. For categorizing these article datasets, we have used word embedding methods to extract features from raw data and then used deep learning methods: Convolutional Neural Network (CNN), Artificial Neural Network (ANN), Long Short-term memory (LSTM) and Hybrid (CNN+Bi-LSTM) methods to classify them based on their contents. Next, we have analyzed the performance of these deep learning based classification models and looked into the reasons for it. Furthermore, the hybrid model has a greater accuracy of 88.56% in 10 categories and 84.93% in 12 categories. Lastly, the limitations of this study as well as future prospects have been discussed.