Assessment of sentiments: A performance evaluation on Bangla noisy text

Abstract

The fact that people have sentiments is perhaps the most significant distinction between robots and humans. Researchers have been working on ways to imitate sentimentality in computers for decades. The majority of recent Sentiment Analysis research in Natural Language Processing (NLP) has concentrated on the English language. Because of the rich grammatical structure of the text, a few notable studies have been conducted in the Bangla language sector. It should also be highlighted that Bangla lacks a comprehensive dataset. As a consequence, current research projects including Bangla have failed to yield findings that are similar to those produced by researchers in other languages and reusable for future study. In this work three categorical machine learning models namely classical, neural network, and transformers that are prevalent in sentiment analysis tasks have been evaluated on a recently introduced noisy Bangla dataset. The experimental outcome showed that the classical machine learning model Support Vector Machine trained with n-gram feature extractors from the category of classical methods performed preferably in contrast to the models in the same category and other categories of approaches implemented. The results acquired in this work can be subsidiary in terms of understanding the impact of the content and human perception from comments that include distorted words or regional dialects associated with different media domains.