Aspect-based Sarcasm Detection in COVID-19 Vaccination Tweets using Transformers-based Approaches

Abstract

Sentiment analysis is the technique that analyzes our attitudes and emotions toward an entity. Sarcasm is a factor that can affect sentiment analysis and eventually results in factually incorrect results. Several sentimental analysis studies have been conducted on Coronavirus and the measures taken against it. All of them are based on an over all analysis strategy, and none of them goes deeper to extract the significant aspects about which users are being sarcastic. This research aims to perform an aspect-based sarcasm detection on Twitter data to get real insights into the public sentiments to ward the vaccination program initiated by the Government of Pakistan. The under lying study has three phases; first, four major topics are extracted from the data for which the Latent Dirichlet Allocation and BERTopic techniques are used. Secondly, four state-of-the-art transformers-based models are used for sarcasm detection: BERT, DistilBERT, RoBERTa, and XLNet. All of them are fine-tuned and trained on a labeled dataset. Outperforming the others, BERT showed a validation accuracy of almost 92%. Lastly, sentiment analysis is performed on each topic using two popular lexicon-based techniques, VADER and TextBlob. It is observed that the text classified by sentiment analyzers as positive, negative, or neutral is substantially sarcastic actually and may lead to factually incorrect conclusions. This study can help analyze any future event, trend, or product on a deeper aspect-based level and explore each aspect by deter mining the sarcasm in it. This experimental process can reduce the risk of misleading decisions based on incorrect facts.