Abstract
Although distant supervision automatically generates training data for relation extraction, it also introduces false-positive (FP) and false-negative (FN) training instances to the generated datasets. Whereas both types of errors degrade the final model performance, previous work on distant supervision denoising focuses more on suppressing FP noise and less on resolving the FN problem. We here propose H-FND, a hierarchical false-negative denoising framework for robust distant supervision relation extraction, as an FN denoising solution. H-FND uses a hierarchical policy which first determines whether non-relation (NA) instances should be kept, discarded, or revised during the training process. For those learning instances which are to be revised, the policy further reassigns them appropriate relations, making them better training inputs. Experiments on SemEval-2010 and TACRED were conducted with controlled FN ratios that randomly turn the relations of training and validation instances into negatives to generate FN instances. In this setting, H-FND can revise FN instances correctly and maintains high F1 scores even when 50% of the instances have been turned into negatives. Experiment on NYT10 is further conducted to shows that H-FND is applicable in a realistic setting.
Introduction
Relation Extraction (RE) is critical for identifying semantic connections between entities in text. To overcome the high cost of manual annotation, Distant Supervision (DS) aligns knowledge bases with text to automatically generate labels. However, DS is inherently noisy, producing False Positives (FPs)—incorrectly assigned relations—and False Negatives (FNs)—instances labeled as "No Relation" (NA) despite expressing a valid relation. While FP denoising has seen significant research, FNs are frequently overlooked, leading to substantial recall degradation. H-FND (Hierarchical False-Negative Denoising) is proposed to specifically target and rectify these missing labels.
Methodology: The H-FND Framework
The H-FND framework employs a reinforcement learning-based approach to denoise NA instances. It consists of two primary components: a Denoising Agent (DA) and a Relation Classifier (RC).
Hierarchical Policy
For every instance labeled as NA, the Denoising Agent executes a hierarchical decision process:
- Keep: The NA label is deemed correct and used as a negative example.
- Discard: The instance is identified as ambiguous or noisy and is removed from the training set.
- Revise: The instance is recognized as a false negative. It is then passed to the Relation Classifier to be reassigned a correct relation label.
Training Strategy
The framework utilizes a multi-stage training process. Initially, a transfer learning scheme pretrains the RC on noisy data, using its confidence scores to generate pseudo-labels for pretraining the DA. Subsequently, the DA and RC are co-trained. The DA cleans the data, the RC trains on this cleansed set, and the resulting F1 score on a validation set serves as a reward signal to update the DA via Policy Gradient (REINFORCE).
Experimental Setup
H-FND was evaluated on three benchmarks:
- SemEval-2010 & TACRED: Human-annotated datasets where FN noise was simulated by randomly flipping 10% to 50% of positive relations to "No Relation."
- NYT10: A standard distantly supervised dataset representing real-world noise.
Results and Discussion
H-FND demonstrates remarkable robustness in high-noise environments. As shown in the experimental results, H-FND consistently maintains superior F1 scores compared to baselines like IRMIE, Co-teaching, and Cleanlab, particularly as the FN ratio increases.
Key Findings
- Robustness to Extreme Noise: On SemEval and TACRED, H-FND's performance remains stable even when 50% of training instances are false negatives.
- High Revision Accuracy: The model effectively recovers lost signals, with the "Revise" action achieving 90-93% accuracy on TACRED.
- Real-World Application: On the NYT10 dataset, H-FND significantly improves precision at higher recall levels (>50%) over standard CNN/PCNN models.
- Dynamic Learning: Unlike methods requiring predefined noise rates, H-FND learns the noise distribution dynamically through RL.
Conclusion
H-FND provides a powerful solution for the often-ignored problem of false negatives in distant supervision. By incorporating a "Revise" action, it preserves valuable training data that other methods might simply discard. Future research will focus on unified frameworks capable of handling both FP and FN noise simultaneously and integrating advanced encoders like BERT to further enhance performance.
Summary updated