DEEP LEARNING-BASED SENTIMENT ANALYSIS USING A HYBRID BERT AND BI-LSTM MODEL

In the digital world, large amounts of content—ranging from social site posts and product response to news articles—contain valuable sentiments data that, if properly analysed, can provide significant insights for businesses, governments, and individuals. Sentiment Analysis, also known as opinion mining. Sentiment analysis mainly involves to have a good look on their thoughts, ideas, behaviours, opinions, sentiments. Traditionally we have used machine learning things to do the job but these models could not able to get what exactly we tries to achieve so we further extended our research to try something that can handle the other points like sarcasm, idiomatic phrases and negation. To address these issues, this research deep dive to deep learning. A merged model for sentiment extraction by integrating the Bidirectional Encoder Representations from Transformers (BERT) with a Bidirectional Long Short-Term Memory network. BERT is better at contextual word embeddings. If we talk about Bi-LSTM it can be effective on long-range dependencies data that is sequential in nature. This proposed model is trained on Sentiment dataset. In this model for the sake of preprocessing we have used tokenization, padding and stop word removal so that it can used by the model. Evaluation parameters like accuracy, precision, recall, and F1-score are used to contrast the proposed model with baseline models including standard LSTM, CNN, and classical ML algorithms. For an example the sentence “I didn’t expect much, but this movie was surprisingly good” is correctly classified as positive, that showcase the contextual ability of model. The results demonstrate that the BERT-BiLSTM hybrid approach achieves better performance, particularly in detecting sentiment polarity in complex or ambiguous sentences. This hybrid approach contributes to the field by providing a scalable architecture suitable for real-time tasks such as brand monitoring, customer feedback analysis and political opinion but only by improving emotion classification accuracy. If we talk about future scope of model its functionality can be extended for other languages also.