.png){kind=icon}
.png){kind=icon}
Return
SDZRE: A semantic distillation method for zero-shot relation extraction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.127609.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Zero-shot relation extraction (ZeroSRE) task aims to identify and extract new relations that have not appeared in the training process. The semantic matching based method has gained significant attention in recent years, which predicts relationships by matching sentences with relational descriptions. However, while most research methods have improved sentence representation quality, they have not fully eliminated interference from complex semantics. Moreover, these methods are prone to misclassifying different relations as the same type when they share similar contextual and entity information. To solve these issues, we propose an efficient Semantic Distillation Method for Zero-Shot Relation Extraction (SDZRE) based on fine-grained semantic matching for ZeroSRE tasks. Specifically, We innovatively design a bidirectional semantic distiller to overcome the limitations of solely removing irrelevant features. This approach enables more effective extraction of the core semantics of sentences, thereby mitigating the interference of complex semantics. Additionally, we propose a novel contrastive learning framework that integrates the bidirectional semantic distiller and employs a combination of random masking and feature truncation strategies for data augmentation. This framework effectively amplifies the differences between similar relations, helping the model learn more meaningful feature representations and reducing the impact of relation similarity. Furthermore, we introduce a multi-negative sample selection and training strategy to further refine the relational feature space, thereby enhancing the discriminative ability of model. Extensive experimental results show that SDZRE, while maintaining efficient inference, significantly outperforms existing methods in extracting core semantics, reducing complex semantic interference, distinguishing similar relations, and enhancing the discriminative ability of model. It achieves state-of-the-art (SOTA) performance, providing a novel approach for the ZeroSRE task that balances both performance and efficiency.
Keywords:
Relation extraction
Zero-shot
Semantic matching
Semantic distillation
Contrastive learning
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W