A Deep Learning-Based Cyber Framework for Bi-Lingual Sentiment Analysis

Abstract

This study on cyber sentiment analysis on cross-lingual datasets is based on the BERT model and the RoBERTa model. Cyber sentiment reflects the emotional attitudes and sentimental tendencies expressed by users through online datasets. While earlier research has mostly looked at how to identify whether people feel happy or sad about something in one language, our study looks at how to do this in many languages. We joined together lots of web-based information about how people feel about things and made a system with five classes. This system can tell the difference between discrimination, hate speech, cyberbullying, sexual harassment, and normal speech. We want to push the limits of sentiment analysis beyond BERT and RoBERTa, which mostly deal with monolingual and binary classification tasks. To do this, we test traditional and cutting-edge deep learning structures on cross-lingual datasets. Our results provide valuable insights into how well bilingual web sentiment detection works and the challenges it faces. This will help to improve the development of AI-driven moderation systems that can be used widely and are easy to expand. We will test how well BERT and RoBERTa can identify emotions in different languages. We will use English (33,938), Chinese (30,376), and Chinese-English (64,313) data. On the English dataset, RoBERTa achieved an accuracy of 91.16% and an F1 score of 91.11%. This is better than BERT's accuracy of 89.26% and F1 score of 89.25%. So it has the best overall performance. On the other hand, BERT performed better on the Chinese dataset, achieving an accuracy of 87.28% and an F1 score of 84.01%, while Roberta only achieved an accuracy of 83.55% and an F1 score of 82.40%. When trained on all the data together, both models performed well, with RoBERTa (accuracy: 90.52%, F1 score: 90.28%) doing slightly better than BERT (accuracy: 90.36%, F1 score: 90.46%). These results show that how well a model performs can vary a lot based on the language it is used in. This means that choosing the right model is very important for applications that use more than one language.