Adversarial Training for Sarcasm Detection

摘要

Adversarial training has shown expressive performance in image classification task. However, there are few applications in natural language processing domain. In this paper, we propose to apply adversarial training strategy to sarcasm detection with small labeled samples. Several different neural network architectures are adopted including Con-volutional Neural Networks (CNN) and Hierarchical Recurrent Neural Networks (HRNN). The experimental results on three datasets show that adversarial training is effective to improve the performance on sarcasm detection.