Sarcasm Detection Using Graph Convolutional Networks with Bidirectional LSTM

Abstract

Sarcasm is a form of figurative language where the literal meaning of words does not hold, and instead, the opposite interpretation is intended. This deliberate ambiguity makes sarcasm detection an important task in sentiment analysis. Sarcasm detection is considered to be a binary classification problem. Traditional works rely heavily on feature-rich traditional models and deep learning models. However, in the model training, previous works worked with a sentence or comment as a unit, they may lose some important global features and ignore rich relational structures in the corpus. In contrast, if we use the whole sarcasm corpus to construct a graph network in the model, we can learn to get representations with global information under sarcasm background. In this work, we propose a new type of neural network model. Specifically, we use a graph convolutional neural(GCN) network to capture the features of global information in the satire context and jointly bidirectional LSTM(bi-LSTM) neural network to capture the sequence features of the comments respectively. And finally, we concatenate these two embeddings and put them into a traditional classifier for classification prediction. Experiment results include precision, accuracy, recall and F1 score on the publicly available dataset show that our model is significantly better than the standard evaluation model benchmarks.