Sentiment classification with modified RoBERTa and recurrent neural networks

The unprecedented growth in the use of social media platforms, where opinions and decisions are made and updated within seconds. Hence, Twitter is becoming a huge commercial interest for brands and companies to assess the sentiment of customers. Sentiment analysis tries to extract subjective opinions and sentiments from opinionated data using Natural Language Processing (NLP). Ontology-based analysis was primarily used to disambiguate the terms and get a high precision score for the emotive terms. In this paper, to improve the accuracy of sentiment analysis we modified the RoBERTa model to extract the more relevant contextualized information. Moreover, this modified RoBERTa is combined with RNN for effective sentiment classification. The proposed work attempts to find which words or phrases actually contribute to the particular sentiment as output by a modified RoBERTa model and this output of the RoBERTa model is fed as input for RNNs. The proposed model is experimented on the Twitter comment dataset. The proposed model experimented on various models such as single RNN, single layer LSTM, and Bi-directional LSTM and evaluated performance measures in terms of accuracy, precision, recall, and F1-score. Our proposed model performance significantly improved with respect to all other models in terms of accuracy, precision, recall, and F1-score. The experiments show that the proposed model not only increases the Jaccard similarity score but also improves different RNN performance when compared to existing state-of-the-art models. The proposed approach obtained a maximum accuracy of 84.6% which is a huge improvement and also evaluated comparison analysis of Simple RNN, Single-LSTM, and Bi-LSTM on full text and selected test. Our proposed modified ROBERTa performance is superior with selected text and full text. Finally, the statistical paired T-test is performed between the proposed model, and other models such as simple RNN, One layer RNN is giving evidence that the proposed model performance is superior with 95% confidence and p<0.05.

Automated summary

The paper proposes an improved model for sentiment analysis on social media platforms, specifically Twitter, using a modified RoBERTa model combined with RNN. The proposed model achieves higher accuracy, precision, recall, and F1-score compared to existing state-of-the-art models. Experimental results demonstrate significant improvements in performance. The proposed approach has a maximum accuracy of 84.6% and outperforms other models in terms of various evaluation measures.