A sentiment analysis framework using RoBERTa with Bi‐LSTM classification on Dynamic twitter data

Abstract This study employs RoBERTa, a robustly optimized variant of Bidirectional Encoder Representations from Transformers (BERT), to capture the semantic and contextual features from text using self‐attention for enhanced feature representation. The extracted features are classified using a bidirectional long short‐term memory (Bi‐LSTM) based on sentiments into three classes: positive, neutral, and negative. The Bi‐LSTM is optimized with adaptive moment estimation and reduced‐learning‐rate‐on‐plateau in the training process to adjust the learning rate, maximize the convergence of the training process, and prevent overfitting. The proposed method is evaluated on three datasets: Sentiment‐140, COVID‐19, and Twitter Emoji. The experimental results show that the proposed method obtains an accuracy of 97.905%, precision of 97.921%, recall of 97.994%, and an F1 score of 97.983% on the Sentiment140 dataset, along with an accuracy of 96.93% on the COVID‐19 dataset. The proposed model demonstrates superior effectiveness compared to existing conventional methods, namely the gated attention recurrent network and fine‐tuned BERT models.