Bangla Document Categorisation using Multilayer Dense Neural Network with TF-IDF

Abstract

Document categorisation is a quintessential example of a natural language processing quest which includes sorting documents by their content into one or more predefined classes. This thesis proposes a model which consists of multilayer Dense Neural Network with Term Frequency - Inverse Document Frequency (TF-IDF) as feature selection technique in terms of Bangla text document categorisation. This proposed system is divided into three consecutive steps: i) preprocessing raw text data and extracting feature using TF- IDF, ii) designing the model architecture and fitting the model to training set, and iii) evaluating model performance on test set by measuring accuracy and weighted average of F1-score. It is observed from experiments that the proposed method exhibits higher accuracy (85.208%) and weighted F1 score (0.85) compared to the other well-known classification algorithms (K Nearest Neighbor, Decision Tree, Support Vector Machine, Stochastic Gradient Descent, Multinomial Na¨ıve Bayes, and Logistic Regression) for Bangla text document classification.