Source-linked AI summary
DSGAN: Generative Adversarial Training for Distant Supervision Relation Extraction
Pengda Qin, Weiran Xu, William Yang Wang
TL;DR
Distant supervision scales relation-extraction labeling but introduces noisy sentences, while existing bag-level denoising can miss sentence-level false positives. DSGAN adversarially learns a sentence-level true-positive generator, uses it to clean the training data, and reports improved distant-supervision relation extraction performance.
Problem
Distant supervision efficiently labels relation-extraction data but introduces noisy samples, and bag-level denoising can overlook entity pairs whose sentences are all false positives.
Method
DSGAN adversarially trains a sentence-level generator to identify true positives, using generated positives as discriminator negatives and selecting the generator when discriminator performance drops most.
Results
The authors report consistent performance gains and improved distant-supervision relation extraction after filtering the training data with DSGAN.
Takeaways & Limitations
The model-agnostic generator can provide a cleaned distant-supervision dataset for downstream relation extractors without supervised information.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper addresses noisy distant-supervision labels by introducing sentence-level adversarial denoising for relation extraction. DSGAN identifies true positives, filters the training data, and improves downstream relation extraction.
- Distant supervision efficiently generates relational data but often produces noisy training samples that impede relation-extraction performance.
- Bag-level denoising methods can miss cases where every sentence for an entity pair is a false positive.
- DSGAN uses adversarial learning to train a sentence-level generator that recognizes true positives without supervised information.
- The generator challenges a discriminator by assigning generated samples negative labels and the remaining samples positive labels.
- The method is sentence-level and model-agnostic, enabling plug-and-play use with relation extractors.
- The cleaned dataset improves recently proposed neural relation extractors without supervised information.
2 Related Work
Prior distant-supervision research models noisy labels with multi-instance learning, attention, or feature-based filtering. These approaches may suppress noise, miss valuable sentences, propagate errors, or operate less directly than sentence-level filtering.
- Early methods formulate distant-supervision relation extraction as multi-instance single-label or multi-instance multi-label learning.
- Sentence-selection methods may miss valuable information, while soft attention methods down-weight incorrect sentences within an entity pair.
- Feature-based noise filtering can suffer error propagation from tools such as named-entity recognition and dependency parsing.
- The paper positions recognizing true positives before relation extraction as preferable to integrating soft denoising into relation prediction.
3 Adversarial Learning for Distant Supervision
DSGAN trains a sentence-level generator and discriminator adversarially to identify true positives within noisy distant-supervision bags. The resulting generator is used to filter noise and produce a cleaned dataset for relation classification.
- Adversarial learning pipeline: DSGAN obtains a sentence-level generator that discovers true positive samples from noisy distantly supervised data without supervised information.The generator assigns each sentence a probability of being a true positive and samples sentences from each bag.
- Adversarial learning pipeline: Generated high-confidence sentences form T and are treated as negative samples for discriminator training, while the remaining sentences F are treated as positive.This adversarial labeling reverses the generator's and discriminator's views of the sampled sentences.
- Generator optimization: The generator is optimized with policy-gradient reinforcement learning because its discrete sampling objective cannot be directly optimized by gradient-based methods.Its reward includes discriminator confidence on generated samples and a second component based on discriminator predictions for ND, with baselines used for variance reduction.
- Adversarial training strategy: The discriminator is reset to the same pre-trained parameters at each epoch, and the generator is selected when discriminator performance drops most within an epoch.The bag sequence and sentences are held identical across epochs to create an equal comparison condition.
- Noise filtering: After training, one generator is obtained for each relation type and used as a binary classifier to filter distant-supervision noise.The algorithm splits the positive set into bags, samples T according to generator probabilities, and retains the rest as F during adversarial training.
4 Experiments
Experiments evaluate DSGAN through dataset construction, adversarial convergence, generator quality, and downstream distant-supervision relation extraction. The results indicate that DSGAN produces stronger positive sets and improves several CNN- and PCNN-based extractors.
- Evaluation and implementation: Experiments use the Riedel distant-supervision dataset, whose automatically labeled instances lack a ground-truth test set, so evaluation relies on held-out measurement.The dataset pairs Freebase relations with sentences from the New York Times corpus; held-out evaluation provides an approximate precision measure without costly human annotation.
- Training process: DSGAN training is examined on three relation types by tracking discriminator performance on ND across epochs and comparing generators through positive-set training difficulty.Figure 3 compares DSGAN-selected, randomly selected, and pre-trained-generator positive sets under equal-size positive and shared negative sets.
- Training process: The discriminator’s accuracy on ND declines as the generator improves, and training stops at the critical point where this accuracy no longer decreases.The decline is interpreted as the generator producing increasingly accurate true-positive samples and becoming strong enough to collapse the discriminator.
- Quality of generator: The DSGAN-generated positive set yields the best training performance among equal-sized DSGAN, random, and pre-trained-generator positive sets.The comparison supports the claim that adversarial learning produces a robust true-positive generator; the pre-trained generator also performs well but does not provide the same false-positive boundary.
- Performance on distant supervision relation extraction: All four evaluated extractors—CNN+ONE, CNN+ATT, PCNN+ONE, and PCNN+ATT—achieve further improvement after training on the redistributed dataset.The comparison covers CNN- and PCNN-based models, with aggregate PR curves shown for each model family.
- Performance on distant supervision relation extraction: DSGAN addresses entity pairs whose entire sentence bag is falsely labeled, a case overlooked by prior bag-level filtering methods; Table 2 reports AUC comparisons and t-test results.The paper states that all reported p-values are below 5e-02 and attributes the downstream gains to filtering before relation extraction.
5 Conclusion
DSGAN uses adversarial training to jointly generate true positives and discriminate positive and negative samples for distant supervision relation extraction. The model-agnostic approach significantly improves competitive baselines on the widely used New York Time dataset.
- DSGAN combines a true-positive generator with a discriminator that classifies positive and negative samples through adversarial training.The generator improves true-positive prediction while the discriminator's performance declines toward equilibrium.
- The model-agnostic framework can be applied to any distant supervision relation extraction model.
- DSGAN significantly improves the performance of many competitive baselines on the widely used New York Time dataset.