Multi-Language Sentiment Analysis Using BERT and BiLSTM

The growing globalization of online communication requires the development of powerful multilingual sentiment analysis (MLSA) systems, which are able to decode the emotional settings of various languages and different cultures. In the presented paper, a hybrid architecture is proposed that involves Bidirectional Encoder Representations with Transformers (mBERT) and Bidirectional Long Short-Term Memory (BiLSTM) networks to improve multilingual sentiment classification. Instead of relying on contextual embeddings, or recurrent sequence modeling, the proposed BERT-BiLSTM fusion architecture uses the semantic richness of mBERT but pathways sequential dependencies with BiLSTM layers, which leads to cross-lingual emotion representation by using the weighted cross-entropy regularization and attention-based token alignment. Four benchmark multi-lingual datasets (Twitter-MLSA, MultiEmo, Amazon-Reviews-Mix, and Code-Mixed-SocialMedia) with ten languages, such as English, Hindi, Arabic, and Spanish are experimented on. Findings indicate a mean 6.4 per cent enhancement in F1-scores compared to state-of-the-art hybrid assistants like CNN-BiLSTM and XLM-RoBERTa-GRU.