A knowledge-augmented neural network model for sarcasm detection

Automatic sarcasm detection from text is one important research task in text mining and natural language processing and has attracted extensive attention from researchers. Most approaches focus on designing various models and features according to the original text, without making use of the knowledge and information from external knowledge source such as Wikipedia, which is publicly available. In this paper, we investigate a knowledge-augmented neural network model that leverages the contextual information of the original text from external knowledge source, for sarcasm detection. We first extract the context from external knowledge source for the original text. Then, the original text and its context are fed sequentially into the embedding layer and encoding layer, to automatically learn high-level semantic representation. Finally, we use the softmax layer to output the classification probability. Based on the Semeval-2018 Task 3 dataset, results show that our proposed model gives 82.79% F1 score, outperforming the existing models and strong neural baselines with significant margins. Experimental analysis also indicates that the contextual information is highly important for sarcasm detection.