Unveiling the multifaceted dynamics of Banglish online communication: a comparative analysis of sentiment, Toxicity, Hate, and Threats

Abstract

Banglish, a Bengali-English language, is gaining popularity online. However, navigating its complexities presents challenges for Natural Language Processing (NLP) methods due to its linguistic fusion and lack of resources. This research explores the multifaceted analysis of Banglish text, including tasks like toxicity detection, identity hate prediction, threat assessment, and insult recognition. Using a dataset of 15,370 Banglish comments from social media platforms, the study investigates the effectiveness of four machine learning models: Support Vector Classifiers (SVCs), Random Forests Classifiers (RFCs), Long Short-Term Memory (LSTM) networks, and Bi-Directional LSTMs. Support Vector Machines (SVM) outperform other models in sentiment analysis, identifying Banglish text sentiment with 87% accuracy. This allows businesses and social media platforms to customize information and services based on this performance. With an 85% accuracy rate, SVCs are also excellent at anticipating potential toxicity. SVC also predicts insult and hatespeech with 75% and 77% accuracy for promoting safer online conversation. They also lead the industry with a 73% accuracy rate in identifying potential threat from Banglish text, ensuring a safer online environment. The study explores the effectiveness of Support Vector Machines (SVM) in Banglish text classification, highlighting their potential in handling intricate aspects. However, further research is needed on transfer learning strategies, domain-specific word embeddings, and ethical issues in code-mixed language processing. The research also addresses practical issues like danger assessment.