DSGAN: Generative Adversarial Training for Distant Supervision Relation Extraction

TL;DR

This paper introduces DSGAN, an adversarial training framework that improves distant supervision relation extraction by accurately identifying true positive samples at the sentence level, significantly enhancing performance.

Contribution

The paper proposes a novel adversarial learning approach, DSGAN, for sentence-level noise reduction in distant supervision relation extraction, outperforming existing soft bag-level methods.

Findings

Significant performance improvement over state-of-the-art systems

Effective filtering of false positive instances

Enhanced dataset quality for relation classification

Abstract

Distant supervision can effectively label data for relation extraction, but suffers from the noise labeling problem. Recent works mainly perform soft bag-level noise reduction strategies to find the relatively better samples in a sentence bag, which is suboptimal compared with making a hard decision of false positive samples in sentence level. In this paper, we introduce an adversarial learning framework, which we named DSGAN, to learn a sentence-level true-positive generator. Inspired by Generative Adversarial Networks, we regard the positive samples generated by the generator as the negative samples to train the discriminator. The optimal generator is obtained until the discrimination ability of the discriminator has the greatest decline. We adopt the generator to filter distant supervision training dataset and redistribute the false positive instances into the negative set, in which way to provide a cleaned dataset for relation classification. The experimental results show that the proposed strategy significantly improves the performance of distant supervision relation extraction comparing to state-of-the-art systems.