Source-linked AI summary
OneRel:Joint Entity and Relation Extraction with One Module in One Step
Yu-Ming Shang, Heyan Huang, Xian-Ling Mao
TL;DR
Joint extraction methods can overlook the interdependence of triple elements, leading to cascading errors and redundant information. OneRel addresses this by using fine-grained triple classification with a scoring-based classifier and relation-specific horns tagging. Experiments report better performance than state-of-the-art baselines across datasets and complex overlapping scenarios.
Problem
Decomposition-based joint extraction overlooks the interdependence and indivisibility of head entities, relations, and tail entities, creating risks of cascading errors and redundant information.
Method
OneRel casts joint extraction as fine-grained triple classification using a scoring-based classifier and relation-specific horns tagging in one module and one step.
Results
OneRel outperforms state-of-the-art baselines on two public datasets, with consistent gains across overlapping patterns and multiple-triple scenarios.
Takeaways & Limitations
OneRel supports joint extraction by simultaneously modeling triple elements and using efficient decoding for complex overlapping cases.
Takeaways & Limitations
The model uses a simplified HOLE score function, and the authors plan to develop a more efficient and powerful alternative.
Abstract
from arXiv · showhide
Joint entity and relation extraction is an essential task in natural language processing and knowledge graph construction. Existing approaches usually decompose the joint extraction task into several basic modules or processing steps to make it easy to conduct. However, such a paradigm ignores the fact that the three elements of a triple are interdependent and indivisible. Therefore, previous joint methods suffer from the problems of cascading errors and redundant information. To address these issues, in this paper, we propose a novel joint entity and relation extraction model, named OneRel, which casts joint extraction as a fine-grained triple classification problem. Specifically, our model consists of a scoring-based classifier and a relation-specific horns tagging strategy. The former evaluates whether a token pair and a relation belong to a factual triple. The latter ensures a simple but effective decoding process. Extensive experimental results on two widely used datasets demonstrate that the proposed method performs better than the state-of-the-art baselines, and delivers consistent performance gain on complex scenarios of various overlapping patterns and multiple triples.
Introduction
Joint entity and relation extraction is important for NLP and knowledge graph construction, but decomposition-based methods can suffer cascading errors and redundant information. OneRel instead performs fine-grained triple classification with one module in one step.
- Introduction: Joint extraction identifies head, relation, and tail triples from unstructured text for NLP and knowledge graph construction.
- Introduction: Existing methods decompose joint extraction into multiple modules or processing steps, including cascading classification and text generation.
- Introduction: Cascading approaches risk errors propagating from early predictions, while separate recognition can produce redundant information because entity-relation constraints are insufficient.
- Introduction: Triple elements are interdependent and indivisible, so OneRel casts extraction as judging fine-grained candidate triples directly.
- Introduction: OneRel combines a scoring-based classifier with relation-specific horns tagging to extract triples using one module in one step.
- Introduction: Experiments on two public datasets report better performance than state-of-the-art baselines, especially for overlapping triples and complex scenarios.
Related Work
Existing joint extraction methods either process triple elements serially with multiple modules or recognize them in parallel before recombining them. OneRel uses fine-grained triple classification in a one-module, one-step design and addresses overlapping cases that a prior one-step method cannot handle.
- Related Work: Multi-module multi-step methods use different modules and interrelated steps to extract entities and relations serially.
- Related Work: Some serial methods identify entities before classifying relations between entity pairs, while others detect relations before predicting head and tail entities.
- Related Work: Multi-module one-step methods extract entities and relations in parallel, then combine them into triples through mechanisms such as table filling or set prediction.
- Related Work: OneRel formulates joint extraction as fine-grained triple classification in a one-module, one-step manner to address cascading errors and redundant information.
- Related Work: Unlike Novel-Tagging, OneRel is designed for overlapping cases rather than assuming each entity pair has at most one relation.
Method
OneRel formulates joint extraction as fine-grained triple classification, using a scoring-based classifier and Rel-Spec Horns Tagging to assign and decode relation-specific entity-boundary tags.
- Task Definition: OneRel assigns tags to every (wi, rk, wj) combination in a three-dimensional matrix M and decodes triples from the resulting classifications.The matrix has dimensions L × K × L, covering token pairs and predefined relations.
- Rel-Spec Horns Tagging: Rel-Spec Horns Tagging marks paired entity boundaries with HB-TB, HB-TE, and HE-TE tags, leaving all other cells as “-”.Only three rectangle corners require labels, producing a sparse matrix.
- Rel-Spec Horns Tagging: The tagging strategy supports overlapping triples by placing entity pairs in relation-specific sub-matrices.EntityPairOverlap uses different relation sub-matrices, while same-relation overlaps are represented within the corresponding matrix.
- Scoring-based Classifier: BERT provides token embeddings, after which OneRel enumerates token-pair and relation combinations for classification.The classifier uses token embeddings ei and ej together with a relation representation rk.
- Scoring-based Classifier: OneRel projects concatenated head and tail representations through a nonlinear function, preserving order to model asymmetric relations.The projection uses trainable W and b with ReLU activation; concatenation is non-commutative.
- Scoring-based Classifier: Relation representations simultaneously score all four tags for each token pair, enabling parallel scoring and reducing implementation processing to L × 1 × L.The score vector is passed through dropout and a softmax to predict tags, while the objective uses gold tags from annotations.
Experiments
The experiments evaluate OneRel and analyze its properties to validate the proposed model’s effectiveness.
- The paper conducts extensive experiments to validate OneRel’s effectiveness and analyze its properties.
Experimental Settings
OneRel is evaluated on NYT and WebNLG using partial- or exact-match metrics, with comparisons against ten baselines and an ablation model.
- Datasets: NYT and WebNLG are evaluated in versions annotating either entity endpoints or complete entity spans.The endpoint-annotation versions are denoted NYT∗ and WebNLG∗, while the span versions are NYT and WebNLG.
- Experimental Analysis: The test sets are additionally split by overlapping patterns and triple number to assess performance in complex scenarios.Table 1 reports dataset statistics, with N denoting the number of triples in a sentence.
- Evaluation Metrics: Evaluation uses micro Precision, Recall, and F1-score with Partial Match for NYT∗/WebNLG∗ and Exact Match for NYT/WebNLG.Partial Match checks relation and final head and tail words; Exact Match requires complete entity spans and relation.
- Baselines: The study compares OneRel with ten state-of-the-art baselines, including GraphRel, CasRel, TPLinker, SPN, and PRGC.
- Ablation: The experiments also include OneRel−, which replaces OneRel’s classifier with a linear function over entity and relation representations.Baseline sentence encoders differ across comparison systems, and reported baseline results come from the original literature.
Results and Analysis
Across NYT and WebNLG, OneRel outperforms baselines overall and remains strongest in overlapping, multi-triple, and sub-task evaluations. Its simpler one-module, one-step design also improves efficiency as relation-set size grows.
- Main Results: OneRel achieves state-of-the-art F1-score on all NYT and WebNLG datasets, with all three WebNLG metrics exceeding 90%.The comparison covers ten baselines under Partial Match and Exact Match evaluation.
- Main Results: 1.3 and 2.5 absolute F1-score gains over PRGC occur on WebNLG∗ and WebNLG, respectively.OneRel also exceeds TPLinker by 0.9, 2.4, 0.9, and 4.3 absolute gains across the four datasets.
- Detailed Results on Complex Scenarios: OneRel achieves the best F1-score on 13 of 18 overlapping-pattern and triple-number subsets, including HTO and N ≥5 on both datasets.These are described as the most complex cases, involving nested, shared-word, and combined overlap patterns.
- Results on Different Sub-tasks: OneRel outperforms baselines on most NYT∗ entity-pair and relation-classification instances and on every reported WebNLG∗ indicator.The analysis identifies dataset differences in EPO prevalence and relation inventory as relevant to sub-task difficulty.
- Model Efficiency: As relations increase from 24 in NYT∗ to 171 in WebNLG∗, OneRel’s training-time advantage over TPLinker rises from 1.3× to 6.8×.Both models have theoretical complexity O(KL^2), but OneRel processes K relations in parallel while TPLinker processes one relation per step.
- Topology Structure of Relations: The t-SNE relation topology places people-related and location-related relations in distinct regions, with place lived, place of birth, and place of death close together.The paper links this learned relational structure to predicting EPO triples.
Conclusion
OneRel reformulates joint extraction as fine-grained triple classification and uses one module with scoring and Rel-Spec Horns Tagging, achieving better performance across scenarios. Future work targets a more efficient and powerful score function and applying triple classification to other extraction problems.
- OneRel transforms joint extraction into fine-grained triple classification with a scoring-based classifier and Rel-Spec Horns Tagging.The model obtains triples with one module in one step and is designed to alleviate cascading errors and redundant information.
- OneRel performs better than state-of-the-art approaches across different extraction scenarios on public datasets.
- The authors plan to design a more efficient and powerful score function to strengthen the model’s ability to capture connections between entities and relations.The current model uses a simplified version of HOLE as its score function.
- The authors also plan to explore triple classification in other information extraction problems, such as event extraction.