Source-linked AI summary
PRGC: Potential Relation and Global Correspondence Based Joint Relational Triple Extraction
Hengyi Zheng, Rui Wen, Xi Chen, Yifan Yang, Yunyan Zhang, Ziheng Zhang, Ningyu Zhang, Bin Qin, Ming Xu, Yefeng Zheng
TL;DR
Relational triple extraction remains limited by redundant relation prediction, weak span-based generalization, and inefficient alignment. PRGC decomposes the task into relation judgement, entity extraction, and subject-object alignment, using potential-relation prediction, relation-specific sequence tagging, and global correspondence. Experiments report state-of-the-art benchmark performance with higher efficiency and gains on complex overlapping triples.
Problem
Relational triple extraction is limited by redundant relation judgement, poor generalization of span-based extraction, and inefficient subject-object alignment.
Method
PRGC decomposes extraction into relation judgement, entity extraction, and subject-object alignment, implementing them with potential-relation prediction, relation-specific sequence tagging, and global correspondence.
Results
PRGC achieves state-of-the-art performance on public benchmarks with higher efficiency and fewer parameters, including complex overlapping scenarios.
Takeaways & Limitations
The framework successfully handles complex overlapping triples while alleviating redundancy, span-based extraction weaknesses, and inefficient alignment.
Takeaways & Limitations
The encoder is experimentally instantiated with BERT, although the framework can theoretically be extended to other encoders; training uses equal sub-loss weights for simplicity.
Abstract
from arXiv · showhide
Joint extraction of entities and relations from unstructured texts is a crucial task in information extraction. Recent methods achieve considerable performance but still suffer from some inherent limitations, such as redundancy of relation prediction, poor generalization of span-based extraction and inefficiency. In this paper, we decompose this task into three subtasks, Relation Judgement, Entity Extraction and Subject-object Alignment from a novel perspective and then propose a joint relational triple extraction framework based on Potential Relation and Global Correspondence (PRGC). Specifically, we design a component to predict potential relations, which constrains the following entity extraction to the predicted relation subset rather than all relations; then a relation-specific sequence tagging component is applied to handle the overlapping problem between subjects and objects; finally, a global correspondence component is designed to align the subject and object into a triple with low-complexity. Extensive experiments show that PRGC achieves state-of-the-art performance on public benchmarks with higher efficiency and delivers consistent performance gain on complex scenarios of overlapping triples.
1 Introduction
PRGC reframes relational triple extraction as Relation Judgement, Entity Extraction, and Subject-object Alignment, addressing redundancy, span-based extraction weaknesses, and inefficient alignment. Its experiments report state-of-the-art performance with higher efficiency, fewer parameters, and advantages in complex overlapping scenarios.
- Proposed framework: Its Potential Relation Prediction component restricts entity extraction to relations predicted for the sentence instead of all relations.This targets redundant relation judgement and reduces invalid operations.
- Proposed framework: Global Correspondence aligns subject-object pairs into triples, providing the alignment component of the framework.
- Proposed framework: Relation-Specific Sequence Tagging extracts subjects and objects separately for each potential relation, naturally handling overlapping entities.
- Experimental findings: PRGC achieves state-of-the-art performance on public benchmarks, especially for complex scenarios involving overlapping triples.Ablation studies and analyses confirm the effectiveness of each component.
- Problem perspective: PRGC decomposes relational triple extraction into Relation Judgement, Entity Extraction, and Subject-object Alignment.
- Experimental findings: PRGC shows advantages in accuracy, complexity, parameter count, floating point operations, and inference time compared with previous works.
2 Related Work
Prior relational triple extraction methods include pipelined, cascade, and single-stage approaches, but reported limitations remain in joint decoding, overlap handling, redundancy, and computational complexity. PRGC is presented as an alternative framework addressing these issues through potential-relation prediction, relation-specific tagging, and global correspondence.
- Prior approaches: Traditional approaches separated entity extraction from relation prediction, while later joint models sought to establish correlation between the two tasks.
- Prior approaches: Sequence-labeling approaches unified entity and relation roles but failed to solve overlapping problems.
- Prior approaches: Multihead-selection approaches extracted candidate entities before predicting relations for every entity pair but did not decode jointly.
- Prior approaches: CasRel applies span-based object extraction to every relation, causing redundant relation judgement and limiting parallelization during subject-object alignment.
- Prior approaches: TPLinker performs two O(n^2) matrix operations under each relation, creating redundancy in relation judgement and complexity in subject-object alignment.
3 Method
PRGC decomposes relational triple extraction into relation judgement, entity extraction, and subject-object alignment, implemented through three corresponding decoder components. It predicts potential relations, extracts entities with relation-specific tagging, and aligns candidate pairs using a global correspondence matrix.
- Problem decomposition: PRGC frames relational triple extraction as Relation Judgement, Entity Extraction, and Subject-object Alignment.The desired output is a set of triples (s, r, o), and each subtask addresses one stage of producing them.
- Global Correspondence: Global Correspondence scores start-token pairs in an n × n matrix, retaining enumerated subject-object pairs whose scores exceed threshold λ2.Higher matrix values indicate greater confidence that a subject-object pair belongs to a triple.
- Decoder overview: The PRGC decoder comprises Potential Relation Prediction, Relation-Specific Sequence Tagging, and Global Correspondence.These three components fulfill the three decomposed subtasks, respectively.
- Potential Relation Prediction: Potential Relation Prediction restricts entity extraction to relations predicted for the sentence instead of applying it to every relation.The component uses multi-label binary classification with threshold λ1 to select potential relations.
- Relation-Specific Sequence Tagging: Relation-specific sequence tagging extracts subjects and objects separately, allowing the model to handle Subject Object Overlap patterns.The paper notes that a single tagging operation can be used when SOO patterns are absent; it adopts a fully connected network rather than LSTM-CRF.
- Training Strategy: PRGC jointly trains the shared encoder and three components with total loss Ltotal = αLrel + βLseq + γLglobal.The paper assigns equal weights for simplicity: α = β = γ = 1.
4 Experiments
PRGC is evaluated on four public dataset variants using exact-match triple extraction and standard micro-averaged metrics. It outperforms strong baselines across most metrics and complex scenarios while improving efficiency.
- PRGC is evaluated on NYT*, NYT, WebNLG*, and WebNLG using exact-match triples and micro Precision, Recall, and F1-score.Entity annotation differs between starred and unstarred dataset versions.
- PRGC outperforms baseline methods on almost all evaluation metrics across the four datasets, including the strongest recent baseline.The comparison includes CasRel and TPLinker.
- 7% improvement over CasRelRandom shows that PRGC’s decoder remains competitive with randomly initialized BERT encoders.PRGCRandom is compared against CasRelRandom without pretrained BERT parameters.
- 10× improvement over TPLinker on WebNLG* is attributed to reducing redundant negative relations before entity extraction.The reduction is especially relevant to datasets containing hundreds of relations.
- PRGC exceeds all baselines across overlapping patterns in both datasets except SOO on NYT*, where only 45 of 8,110 cases occur.Its performance is also better in almost every subset grouped by the number of triples.
5 Analysis
Analysis examines PRGC’s efficiency, convergence, and component contributions. The results link its gains to potential-relation filtering, relation-specific tagging, and global correspondence.
- 5.1 Model Efficiency: PRGC decodes 2× faster than CasRel and 3× faster than TPLinker in single-thread settings.It also performs significantly better than TPLinker with parallel processing.
- 5.1 Model Efficiency: Against TPLinker, PRGC uses an order of magnitude less complexity, 200× fewer FLOPs, 3× faster inference, and 1.1% higher F1-score.These comparisons are reported for the WebNLG* efficiency evaluation.
- 5.2 Ablation Study: Without Potential Relation Prediction, precision decreases because extraction under absent relations produces redundant predicted triples.The effect is especially pronounced on WebNLG*, whose relation set is much larger.
- 5.2 Ablation Study: Span-based extraction significantly lowers performance, whereas relation-specific sequence tagging handles long entities and relation identification more robustly.The case study reports that span-based models may find entity pairs while missing their relation.
- 5.2 Ablation Study: Removing Global Correspondence lowers precision because heuristic pairing creates mismatched subject-object triples.The global correspondence component therefore outperforms nearest-neighbor alignment.
6 Conclusion
The paper presents PRGC as a joint relational extraction framework that addresses redundant relation judgement, weak span-based generalization, and inefficient alignment. Experiments report state-of-the-art performance, complex-scenario handling, and higher efficiency.
- PRGC decomposes relational triple extraction into Relation Judgement, Entity Extraction, and Subject-object Alignment within a unified framework.Its components are based on Potential Relation and Global Correspondence.
- Experimental results report state-of-the-art performance on public datasets, successful handling of complex overlapping scenarios, and higher efficiency.
A Overlapping Patterns
The paper distinguishes four overlap patterns: Normal, SEO, EPO, and SOO. SOO is identified as a special pattern addressed by the proposed approach.
- SOO is presented as a special overlap pattern identified and addressed in this work, alongside the previously studied Normal, SEO, and EPO patterns.
- The four patterns are Normal, Single Entity Overlap, Entity Pair Overlap, and Subject Object Overlap.Overlapping entities are highlighted in the illustrated examples.
B Implementation Details
The model is implemented and trained with PyTorch, Adam, BERT-Base-Cased English, and fixed hardware and training settings. NYT and WebNLG use different batch sizes, while the encoder and decoder use separate learning rates.
- PRGC is implemented with PyTorch and optimized using Adam.The batch size is 64/6 for NYT/WebNLG.
- The BERT encoder learning rate is 5 × 10−5, while the decoder learning rate is 0.001.Weight decay is set to 0.01.
- Experiments use BERT-Base-Cased English with a maximum input length of 100.The setup uses an Intel Xeon E5 2.40 GHz CPU, 128 GB memory, an NVIDIA Tesla V100 GPU, and CentOS 7.2.
- The model is trained for 100 epochs, and the last model is selected.The reported performance improves when the Potential Relation Prediction threshold is higher.
C Results on Different Subtasks
Detailed evaluations on NYT* and WebNLG* examine the three subtasks corresponding to PRGC's components. Potential relation prediction and entity extraction show strong recall, while subject-object alignment is useful but remains improvable, especially in recall; together, the components achieve solid results.
- Table 8 evaluates Relation Judgement, Entity Extraction, and Subject-object Alignment on NYT* and WebNLG*.Each subtask corresponds to one PRGC component, and the table reports evaluation percentages.
- Relation Judgement: Potential Relation Prediction achieves high recall, and its effectiveness is not affected by relation-set size.Recall is emphasized because missed true relations cannot be recovered in later steps.
- Entity Extraction: The Relation-Specific Sequence Tagging component extracts almost all correct subjects and objects.It is evaluated as named entity recognition, with recall emphasized because alignment can filter many false negatives.
- Subject-object Alignment: The Global Correspondence alignment scheme is useful but can still be improved, particularly in recall.Both precision and recall matter because the evaluation considers entity pairs in triples while ignoring the relation.
- The combination of all three components achieves better and solid relational triple extraction results.