Bangla Sentiment Analysis On Highly Imbalanced Data Using Hybrid CNN-LSTM & Bangla BERT

Sentiment analysis is a technique that combines machine learning and natural language processing to identify the emotional attitude of a text. This is a very active research area in recent years. Bengali is the fifth most spoken Indo-European language in the world. Many people in Bangladesh use news portals and social media to gather information on various topics. We used a publicly available dataset from Kaggle. This data set consists of more negative reviews than positive reviews. We try to experiment with this dataset with different models, such as traditional ML models and deep learning models like CNN, LSTM, and the transformer model (Bangla-BERT-base). The Bangla-BERT-base achieved a notable 96% accuracy through 10-fold cross-validation. Several other performance measures are also used to evaluate our model.

I. Introduction

Sentiment analysis(SA) is a popular research topic in Natural Language Processing(NLP) [1]. Most of the major approaches to sentiment analysis are based on machine learning techniques [2], [3]. This is a field study to determine the polarity and intensity of text containing human emotions, sentiments, and views regarding an entity [4]. Entities can be products, subjects, individuals, and services [5], [6]. Sentiment Analysis is also known as Opinion Mining, but some researchers say they are different because Opinion Mining extracts and analyzes people’s opinions, while Sentiment Analysis identifies sentiment in the text, then analyzes [7]. Sentiment analysis emerged as an analytical and predictive process [8]. Sentiments of the Internet, the repository of information where social networks, websites, web forums, and blogs are the means by which people express their opinions. Anyone can collect huge amounts of data from them. In the age of technology, correct information is worth more than millions of tons of gold. Since we are working on Bangla sentiment analysis, collecting various Bangali data is not such an easy task. In order to improve the quality of the data, it is important to be able to analyze how people feel about it. A machine trained by labeled data will give it an edge for future analysis [9]. Many works in this field, specifically Bengali texts, criticisms, and songs, have been made in the last few years. Our research is enhanced by approaching it with a combined technique of deep learning. The benefit of our research to publishers will be to reflect on their work when it comes to making the right decisions or publishing quality information to audiences based on analytics. The fact that this method is real and can purify more things shows that it is useful. Extracting text from different websites, classifying these texts, and polarizing them will give the best guess. Classification beyond polarization is also a key factor.