A Study of fastText Word Embedding Effects in Document Classification in Bangla Language

Natural language processing is the current topic due to many important tasks like document classification, named entity recognition, opinion mining, sentiment analysis, textual entailment, etc. Such types of task in the Bangla language is also important. This research work endeavored to find out the word embedding of the Bengali language. Leveraging the fastText word embedding, it has shown significant performance in Bangla document classification without any prepossessing like lemmatization, stemming, and others. For the extrinsic evaluation of our word vectors, a classification problem-solving strategy has been used which showed an outstanding result. In the classification module, attempts have been made to classify 40 thousand News samples into 12 categories. For this purpose, three deep learning techniques have been used: Convolutional Neural Network (CNN), Bi-Directional LSTM (BLSTM) and Convolutional Bi-Directional LSTM (CBLSTM) alongside fastText. From the analogous study of all the parameters of every classifier implemented here, we found that the BLSTM technique is the most promising technique for this task. This technique achieved 91.49%, 87.87%, and 85.5% accuracies for Training, Testing, and Validation set, respectively.