Source-linked AI summary
Relabel the Noise: Joint Extraction of Entities and Relations via Cooperative Multiagents
Daoyuan Chen, Yaliang Li, Kai Lei, Ying Shen
TL;DR
Distant supervision creates noisy training labels that can differ from human-annotated test labels, and pipelined extraction can propagate errors. The paper uses cooperative agents to relabel instances and adjust extractor losses, reporting improved performance on two real-world datasets.
Problem
Distant-supervision labels can differ from human-annotated test labels, while pipelined entity-to-relation extraction can propagate errors.
Source: §1 · pp. 1–2Method
Cooperative agents assess noisy instances from different extraction views, combine their confidences to relabel training data, and use those confidences to adjust extractor losses.
Source: §1 · p. 2, §2.2 · p. 3, §2.3.1 · p. 4Results
Takeaways & Limitations
The method can help existing extractors achieve strong extraction performance with noisy labels, with task correlations supporting efficient agent training.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Distant supervision based methods for entity and relation extraction have received increasing popularity due to the fact that these methods require light human annotation efforts. In this paper, we consider the problem of \textit{shifted label distribution}, which is caused by the inconsistency between the noisy-labeled training set subject to external knowledge graph and the human-annotated test set, and exacerbated by the pipelined entity-then-relation extraction manner with noise propagation. We propose a joint extraction approach to address this problem by re-labeling noisy instances with a group of cooperative multiagents. To handle noisy instances in a fine-grained manner, each agent in the cooperative group evaluates the instance by calculating a continuous confidence score from its own perspective; To leverage the correlations between these two extraction tasks, a confidence consensus module is designed to gather the wisdom of all agents and re-distribute the noisy training set with confidence-scored labels. Further, the confidences are used to adjust the training losses of extractors. Experimental results on two real-world datasets verify the benefits of re-labeling noisy instance, and show that the proposed model significantly outperforms the state-of-the-art entity and relation extraction methods.
1 Introduction
The introduction identifies shifted labels in distant-supervision training and noise propagation in pipelined extraction as challenges. It proposes jointly relabeling noisy instances with cooperative agents and reports improved extraction on two real-world datasets.
1 Introduction
- Distant supervision aligns an external knowledge graph with text to create training labels, but the alignment introduces noisy labels that can degrade extractors.Prior approaches include probabilistic graphical models, attention-based neural networks, and reinforcement-learning instance selection.Source: §1 · p. 1
- Distant-supervision training labels differ from human-annotated test labels, with false positives and false negatives tied to knowledge-graph alignment.Existing denoising methods may down-weight noisy instances or discard false positives without recovering original labels, leaving the distribution gap unresolved.Source: §1 · p. 1
- In pipelined extraction, entity errors can propagate to relation extraction, while correlations and complementary information between the tasks may help identify noise more precisely.The introduction gives incompatible entity types for a relation as an example of such information.Source: §1 · pp. 1–2
- Cooperative agents evaluate noisy instances from different extraction views, and a consensus module redistributes them with confidence-scored labels for iterative extractor training.The confidences also adjust extractor training losses.Source: §1 · p. 2, Figure 1 · p. 2
- Evaluations on two real-world datasets show that the approach helps existing extractors achieve strong performance with noisy labels, while task correlations make agent training efficient.The reported evaluation also says the method significantly improves extractor performance.Source: §1 · p. 2
2 Methodology
The methodology uses cooperative reinforcement-learning agents to evaluate noisy instances from entity and relation views, relabel them with confidence scores, and iteratively refine extractors. It also uses confidence-weighted losses and curriculum learning during training.
2.1 Overview
- The framework treats relabeling as decentralized multiagent reinforcement learning, using validation performance as delayed feedback while agents exchange entity- and relation-extractor information through intermediate agents.The agents redistribute confidence-scored labels and retrain the extractors iteratively.Source: §2.1 · pp. 2–3
2.2 Confidence Evaluators as Agents
- Entity- and relation-view agents use task-specific states and continuous confidence actions; a consensus module combines their votes into confidence-scored labels.The state representations include sentence and type information, and agents are assigned subsets of types to scale to large state spaces.Source: §2.2 · p. 3
2.3.1 Loss Correction for Extractors
- Extractor losses are scaled by instance confidence, reducing the optimization impact of low-confidence noisy examples, especially when agent votes diverge.The adjusted loss is ℓ′ = C^λℓ.Source: §2.3.1 · p. 4
2.3.2 Training Algorithm
- Training initializes extractors and agent policies, then iteratively relabels batches, retrains extractors, and rewards agents; curriculum learning orders instances by reward to increase policy difficulty gradually.The policy networks are pretrained using relatively clean data, and PPO is used to learn policies after reward thresholds are met or the priority queue fills.Source: §2.3.2 · pp. 4–5
3 Experiments
Experiments on Wiki-KBP and BioInfer evaluate the approach with CoType and PCNN extractors. The results cover entity and end-to-end relation extraction, training strategies, and patterns in relabeled instances.
3.1 Experimental Setup
- The experiments use Wiki-KBP and BioInfer and compare MRL-CoType and MRL-PCNN with established relation- and entity-extraction baselines.PCNN uses a fixed pretrained CoType entity extractor; the datasets differ in domain and type-set scale.Source: §3.1 · pp. 5–6
3.2.1 Performance on Entity Extraction
- The approach refines base extractors and outperforms all baselines on every metric; on BioInfer, refinements are significant (t-test, p < 0.05) despite 2,200 entity types and CoType’s 0.74 S-F1.Table 2 reports three-time average results with standard deviations.Source: §3.2.1 · pp. 6–7, Table 2 · p. 6
3.2.2 Performance on Relation Extraction
- MRL-CoType and MRL-PCNN significantly refine their base extractors; MRL-PCNN achieves the best Wiki-KBP F1 and beats all baselines on every BioInfer metric.MRL-PCNN significantly improves over RRL-PCNN with the same base extractor and is comparable to BA-Fix-PCNN, which uses test-set information.Source: Table 3 · p. 6, §3.2.2 · p. 7
3.3 Ablation Analysis
- Curriculum and joint multiagent training outperform separate training, while curriculum training performs better than joint training, especially on BioInfer.The authors attribute curriculum’s advantage to gradually increasing instance difficulty, which supports exploration in the larger state space.Source: §3.3 · p. 7, Table 4 · p. 7
3.4 Re-labeling Study
- Relabeling covers noisy positive and negative instances; negative-to-positive relabelings are more common, especially on Wiki-KBP, and divergent agent judgments receive low confidence.The examples show consensus can relabel a noisy NONE instance as positive, while low-confidence disagreement is damped in the loss.Source: §3.4 · p. 8, Figure 3 · p. 8
4 Related Works
The related work spans distant- and weak-supervision extraction, denoising, and joint extraction, while identifying gaps in handling noisy labels and shifted label distributions. The proposed approach is positioned as a plug-in that uses multiagent confidence evaluation to address these issues.
4 Related Works
- Distant- and weak-supervision methods address the training-data bottleneck, while the proposed approach can post-process many such extractors because they incorporate noise.The passage contrasts supervised NER's annotation costs with DS and WS approaches.Source: §4 · p. 8
- CrossWeigh addresses mistakes in supervised gold labels, whereas this work targets automatically constructed silver labels and shifted label distributions.Ye et al. (2019) use test-set distribution information; this work uses RL to explore suitable label distributions without requiring knowledge of the test distribution.Source: §4 · p. 8
- Unlike joint extraction methods that leave noise unmodeled and RL methods that select instances or focus on false positives, this approach evaluates noisy instances with continuous confidence scores.It differs from Reinforced Co-Training by relabeling noisy instances rather than selecting unlabeled ones, and from HRL by using a group of agents as a pluggable helper.Source: §4 · p. 9
5 Conclusions
The paper proposes cooperative multiagents for joint entity and relation extraction, using confidence-scored relabeling to address distant-supervision noise. Experiments on two real-world datasets report improved extractor performance and effective learning.
5 Conclusions
- Cooperative multiagents evaluate noisy instances from different views, and a confidence consensus module relabels them with confidence scores for joint entity and relation extraction.The confidences are then used to adjust extractor training losses.Source: §5 · p. 9
- Confidence-scored labels are used to adjust extractor training losses, potentially alleviating harm from noisy instances.The paper describes RL agents as exploring a suitable label distribution.Source: §5 · p. 9
- Experiments on two real-world datasets report that the proposed method significantly improves extractor performance and achieves effective learning.The conclusion describes these results as confirming the method's effectiveness.Source: §5 · p. 9