Document Level Comparative Sentiment Analysis of Bangla News Using Deep Learning-based Approach LSTM and Machine Learning Approaches
Nuren Nafisa, Sabrina Jahan Maisha, Abdul Kadar Muhammad Masum · 2023
Living in the age of digitalization, people are empowered to easily express themselves as well as being manipulated by information. Online news is one of the powerful mediums to impact people&s;s thoughts and ideas. Vulnerable news may create disturbance and hamper mentally certain age groups or people. Less research works has been done in extracting sentiments from newspapers in a morphologically rich language like Bangla. In this paper, our aim is to automatically analyze the bipolar sentiments (i.e., positive or negative) from news articles and comparatively evaluate the behavior of traditional supervised Machine Learning (ML) classifiers with respect to the deep learning approach Long Short-Term Memory (LSTM) on our Bangla News dataset. In this respect, firstly the dataset is collected from different online news sources and with the help of four annotators from our acquaintance each news document is labelled. To build up a model with ML, performances of six supervised classifiers (i.e., Multinomial Naive Bayes (MNB), K-Nearest Neighbor (KNN), Random Forest (RF), Decision Tree (DT), Logistic Regression (LR) and Linear Support Vector Machine (SVM)) are analyzed along with Count Vectorizer and TF-IDF transformer techniques. Evaluated with percentage split method, it is seen that RF shows best results with 80\% accuracy. Conversely, another model is created employing Word2Vec technique with the LSTM approach. With 100 epochs, this method revealed an accuracy of 83\%. Our model has been able to perform well using LSTM rather than RF approach. Although in both cases accuracy can be considered quite good, increase in volume of dataset can stimulate the performance of model.