Adversarial and Auxiliary Features-Aware BERT for Sarcasm Detection
Avinash Kumar, Vishnu Teja Narapareddy, Pranjal Gupta, Veerubhotla Aditya Srikanth, Lalita Bhanu Murthy Neti, Aruna Malapati · 2020
Sarcasm is a way to express a negative viewpoint using positive or intensified positive words in social media. This intentional ambiguity makes sarcasm detection, a key task of sentiment analysis. Sarcasm detection is modeled as a binary classification problem wherein both feature-rich traditional models and deep learning models have been successfully built to detect sarcastic comments. To help deep learning models generalize better huge labeled data collection is required, which can be laborious and time-consuming. An adversarial process which is carried out in the embedding space provides an alternative mechanism to generate real-world like examples. However, these examples are not real sentences, but act as a regularization method and can make neural networks more robust. In this paper, we present a novel model, Adversarial and Auxiliary Features-Aware BERT (AAFAB), which utilizes contextual word embedding (BERT) for encoding the semantic meaning of the sentence and then combine it with high quality manually extracted auxiliary features for sarcasm detection. Further, adversarial training is performed by adding perturbations to the input word embedding which enables AAFAB to generalize parameters in a better way. The experiment results show that the inclusion of auxiliary features and adversarial training enhances the performance of AAFAB, and it performs better than various baseline traditional models and deep learning models.