Detecting Multilabel Sentiment and Emotions from Bangla YouTube Comments

Abstract

Sentiment analysis has become a key research area in natural language processing due to its wide range of practical applications that include opinion mining, emotions extraction, trends predictions in social media, etc. Though the sentiment analysis in English language has been extensively studied in recent years, a little research has been done in the context of Bangla language, one of the most spoken languages in the world. In this paper, we present a comprehensive set of techniques to identify sentiment and extract emotions from Bangla texts. We build deep learning based models to classify a Bangla sentence with a three-class (positive, negative, neutral) and a five-class (strongly positive, positive, neutral, negative, strongly negative) sentiment label. We also build models to extract the emotion of a Bangla sentence as any one of the six basic emotions (anger, disgust, fear, joy, sadness and surprise). We evaluate the performance of our model using a new dataset of Bangla, English and Romanized Bangla comments from different types of YouTube videos. Our proposed approach shows 65.97% and 54.24% accuracy in three and five labels sentiment, respectively. We also show that the performance of our model is better for domain and language specific texts.