Enhancing Twitter Sentiment Analysis using a Dual Perspective Robust Optimized BERT with Deep Learning

Abstract—Twitter is among the most common social media platforms on which the users can be seen sharing opinions, emotions and experiences, most of the times, using short texts. Twitter Sentiment Analysis (TSA) has gained relevance as the user-generated content has grown exponentially, and it is crucial to find out what users feel and classify their views as positive, negative, or neutral. This is important in most applications of Natural Language Processing (NLP) including opinion mining, social monitoring and market analysis. Nevertheless, most of the current TSA methods primarily depend on textual characteristics and usually encounter problems in the analysis of short, informal, and ambiguous tweets. Moreover, the application of conventional methods of sentiment analysis often omits emojis data despite the popularity of the latter in conveying feelings and contextual meaning in social media communication. These factors can be overlooked, which can decrease the precision of sentiment classification. In an effort to evade these constraints, this paper suggests a two-perspective robust optimized bert (DP-ROBERTa) that combines textual and emoji based data in enhancing the sentiment classification. Twitter sentiment data is retrieved in the Kaggle repository and processed with the Lexicon Text Terms Normalization (LT2N) feature in order to eliminate punctuations, redundant tags and noisy components. Aspect-based Emoji Text Affinity Rate (AETAR) method is then applied to calculate the relationship between emojis and textual aspects by giving affinity weights to them. The proposed DP-ROBERTa model will process these affinity-weighted representations of the tweets, as it will focus on local fine-grained sentiment patterns and global semantic representations. The experimental outcomes show that the proposed framework has a better performance than the baseline models in terms of accuracy, precision, recall, and F1-score to effectively classify the emotions of the Twitter posts.