Interpreting sarcasm on social media using attention-based neural networks

摘要

Posting sarcastic comments is a common trend on social media in which the intended meaning of the expression by the user is opposite to the literal meaning. Most existing approaches have mainly focused on sarcasm detection and left sarcasm interpretation rather underexplored, whereas what can really improve the efficiency of social text analysis techniques are generating the correct interpretation of sarcastic posts. In this paper, we present a deep learning neural network architecture for sarcasm interpretation in which the effect of adapting the idea of different attention mechanisms with the proposed architecture is explored to generate a non-sarcastic post conveying the same meaning as the original sarcastic one. This is based on the idea that, in interpreting a sarcastic post, the incongruity between a positive word and a negative situation plays a key role and consequently making it necessary to pay more attention to some words in the sentence to interpret it accurately. In addition, through extensive experiments, we demonstrate the effectiveness of our methods in interpreting sarcastic posts compared to state-of-the-art baselines. Moreover, we further show that the impact of our methods is especially meaningful when considered for long posts.