Adversarial Discriminative Denoising for Distant Supervision Relation Extraction

Distant supervision has been widely used to generate labeled data automatically for relation extraction by aligning knowledge base with text. However, it introduces much noise, which can severely impact the performance of relation extraction. Recent studies have attempted to remove the noise explicitly from the generated data but they suffer from (1) the lack of an effective way of introducing explicit supervision to the denoising process and (2) the difficulty of optimization caused by the sampling action in denoising result evaluation. To solve these issues, we propose an adversarial discriminative denoising framework, which provides an effective way of introducing human supervision and exploiting it along with the potentially useful information underlying the noisy data in a unified framework. Besides, we employ a continuous approximation of sampling action to guarantee the holistic denoising framework to be differentiable. Experimental results show that very little human supervision is sufficient for our approach to outperform the state-of-the-art methods significantly.