Hybrid Deep Neural Networks for Improved Sentiment Analysis in Social Media

Based on S-BERT pre-trained embeddings, this body of work suggests an approach to sentiment analysis using a convolutional neural network (CNN). GloVe and Word2Vec, utilizing the IMDB dataset. The results of our testing showed that the CNN algorithm we built had the highest accuracy at 89.8 percent, outperforming the GloVe and Word2Vec models, which were considered gold standards at the time. During our research on ablation, we noticed that replacing bigrams or trigrams with N-grams can result in improved model performance. In addition, we used a sentiment lexicon to provide context to the text data, which helped improve the model's accuracy. Our study has demonstrated that sentiment analysis can be performed using S-BERT pre-trained embeddings in combination with a CNN model. This strategy has the potential to outperform both standard machine learning approaches and commonly used word embedding models. When these factors are considered, our suggested strategy of using S BERT pre-trained embeddings shows significant potential in real-world applications where sentiment analysis is critical.