Sentiment Analysis and Opinion Mining using Decoding Enhanced BERT with Disentangled Attention and Light Gradient Boosting Machine

In recent years, consumers are progressively sharing comments and revies on several consumer industries like e-commerce, restaurants and entertainment due to rapid growth in social media. The analysation of these reviews and comments are important for understanding consumer sentiment and improving products as well as services. Even though, previous researchers have suggested various advanced model, but still there are challenges like contextual misinterpretation and negation that lead to inaccurate sentiment classification. To address these challenges, this research proposes Decoding enhanced Bidirectional Encoder Representations from Transformers with Disentangled Attention and Light Gradient Boosting Machine (DeBERTa- LightGBM) for effective sentiment analysis and opinion mining. Initially, data is collected from three datasets namely, SemEval-2014, Sentiment140 and STS-Gold which incorporates restaurant and various tweets review data. This collected data is pre-processed by using tokenization, stemming and stop word removal for cleaning raw text data. After that, feature extraction is done by using Review-Related Features (RRF) and Aspect-Related Features (ARF) to generate Hybrid Feature Vector (HFV). Finally, DeBERTa- LightGBM is employed in which, DeBERTa extracts contextual embeddings and LightGBM classify sentiments into positive, negative, or neutral.