Abstract
Recent work for extracting relations from texts has achieved excellent performance. However, existing studies mainly focus on simple relation extraction; these methods perform poorly on the overlapping triple problem because the tags of shared entities conflict with each other. Especially, overlapping entities are common and indispensable in Chinese. To address this issue, this paper proposes PasCore, which utilizes a global pointer annotation strategy for overlapping relation extraction in Chinese. PasCore first obtains the sentence vector via a general pre-training model encoder and uses a classifier to predict relations. Subsequently, it uses a global pointer annotation strategy for head entity annotation, which uses global tags to label the start and end positions of the entities. Finally, PasCore integrates the relation, head entity, and its type to mark the tail entity. Furthermore, PasCore performs conditional layer normalization to fuse features, which connects all stages and greatly enriches the association between relations and entities. Experimental results on both Chinese and English real-world datasets demonstrate that PasCore outperforms strong baselines on relation extraction and, especially, shows superior performance on overlapping relation extraction.
Introduction
Relation Extraction (RE) is a fundamental task in natural language processing, transforming unstructured text into structured knowledge triples (subject, relation, object). While standard RE models perform well on simple sentences, they often struggle with "overlapping triples," where a single entity participates in multiple relationships. This phenomenon is categorized into types such as Single Entity Pair (SEP), Nested Entity Overlap (NEO), Subject Entity Overlap (SEO), and Entity Pair Overlap (EPO).
In Chinese, sentences are often longer and contain more complex entity overlaps than in English. Traditional BIO (Beginning, Inside, Outside) tagging schemes fail in these scenarios because a single token cannot easily represent multiple tags for different overlapping roles. PasCore addresses this by moving away from sequential tagging toward a global pointer-based strategy.
The PasCore Framework
PasCore (Pointer-based Chinese Overlapping Relation Extraction) employs a cascading, three-stage architecture designed to untangle complex entity webs:
- Relation Prediction: The model uses a pre-trained encoder (e.g., BERT or ERNIE) to predict which relations are present in the sentence. This "relation-first" approach provides a semantic anchor for subsequent steps.
- Head Entity Annotation: For each predicted relation, a Global Pointer strategy identifies head entities. Instead of per-token tagging, it predicts start and end positions in a global matrix, naturally handling nested or overlapping spans.
- Tail Entity Annotation: The model identifies corresponding tail entities. Crucially, it uses Conditional Layer Normalization (CLN) to fuse the features of the predicted relation and the identified head entity into the word vectors.
This cascading information flow ensures that tail entity extraction is strictly conditioned on the specific relation and head entity already identified, mitigating the tag conflict problem.
Experimental Evaluation
The authors evaluated PasCore on three diverse datasets: DuRED (a large-scale Chinese dataset), ICRED (a complex Chinese dataset with long-tailed overlaps), and NYT24 (a standard English benchmark).
As shown in the results, PasCore achieves state-of-the-art (SOTA) performance on Chinese datasets. On ICRED, PasCore (ERNIE-base) achieved an F1 score of 60.7, significantly outperforming previous models like GRTE (55.1) and CasRel (55.0). It also demonstrated strong generalizability on the English NYT24 dataset with an F1 of 92.5. Ablation studies confirmed that the feature fusion via CLN is critical; removing relation features from the tail entity stage caused the F1 score on ICRED to drop from 60.7 to 44.4.
Conclusion
PasCore effectively addresses the overlapping relation extraction problem by combining a global pointer strategy with a cascading architecture. By prioritizing relation prediction and utilizing conditional layer normalization for feature fusion, the model overcomes the limitations of traditional tagging schemes. Future work will focus on reducing error propagation in the pipeline and exploring unified single-step annotation methods.
Summary updated