Sarcasm detection in product reviews using textual entailment approach
Shubham Sinha, Tarun Vijeta, Pratik Kishor Kubde, Ayush Pradeep Gajbhiye, Mansi Anup Radke, Christopher B. Jones · 2023
Sarcasm is a form of sentiment characterized by the use of words that express the opposite of what is meant. Sarcasm detection has applications in multiple domains ranging from sentiment analysis in product reviews to user feedback, and online forums. Sarcasm detection is important to understand user opinions and intentions in areas such as sentiment-based classification and opinion mining. This can result in better product development and customer service. Sarcasm detection can be a challenging task because sarcastic sentences may use positive expressions to convey negative meanings or may use negative sentences to convey positive meanings. Also, sarcastic sentences form a very small component of the entire communication. The increasing use of sarcasm in various social media such as Twitter, Reddit, Amazon product reviews, etc. has highlighted the importance of detecting and understanding sarcasm in various contexts. Sarcasm detection is a challenging problem for NLP systems that often rely on statistical models for performing sentiment analysis. In this research, the focus is on the use of a textual entailment approach for detecting sarcasm. Textual entailment is a natural language inference task that involves determining whether one text (hypothesis) can be derived from another text (premise). The underlying assumption behind this approach is that - if there is a contradiction between the premise and hypothesis, we can say that the hypothesis is sarcastic. To test our approach, an annotated corpus of 3000 product reviews was developed methodically from the Amazon Reviews dataset and tested using the textual entailment approach. The proposed approach achieved an F1 score of 0.76 on this dataset. The result is better than the baseline considered which is the BERT binary classifier which gives an F1 score of 0.48 on the same dataset.