Sarcasm Detection with Self-matching Networks and Low-rank Bilinear Pooling

Sarcasm is sophisticated linguistic expression and is commonly observed on social media and e-commerce platforms. Failure to detect sarcastic expressions in natural language processing tasks, such as opinion mining and sentiment analysis, leads to poor model performance. Traditional approaches rely heavily on discrete handcrafted features and will incur enormous human costs. It was not until recent that scholars began to employ neural networks to address these limitations and have achieved new state-of-the-art performance. In this work, we propose a novel self-matching network to capture sentence ”incongruity” information by exploring word-to-word interactions. In particular, we calculate the joint information in each word-to-word pair in the input sentence to build a self-matching attention vector, based on which we attend the sentence and build its representation vector. Such a network allows sentence to match within itself word by word and cater to the words of conflict sentiments. In addition, we incorporate a bi-directional LSTM network into our proposed network to retain compositional information. We concatenate incongruity information and compositional information through a Low-rank Bilinear Pooling method to control for potential information redundancy without losing discriminative power. Experiment results on publicly available datasets demonstrate that our model significantly outperforms extant baselines on standard evaluation metrics including precision, recall, F1 score and accuracy.