RSED: Zero-Shot Relation Triplet Extraction via Relation Selection and Entity Boundary Detection

Citations Over TimeTop 21% of 2024 papers

Abstract

Zero-shot relation triplet extraction (ZeroRTE) aims to extract relation triplets of unseen relation types from unstructured texts, with a core challenge of training models to recognize new relations without labeled data. The seminal work handles this task by leveraging pre-trained language models to generate synthetic data for new relations, which requires a laborious training procedure. Moreover, the quality of the synthetic data significantly influences the performance. We propose a model named RSED to tackle the ZeroRTE task via Relation Selection and Entity boundary Detection, which considerably simplifies the procedure and is agnostic to the synthetic data. The whole model is composed of three modules. The first relation selection module utilizes a novel relation selection modeling approach to enable the recognition of unseen relations. Such a modelling approach breaks free from the constraints of specific relation types, naturally infuses the relation semantics, and eliminates the need for synthetic data. The second relation filtering module removes irrelevant relations. The third entity boundary detection module extracts multiple entities following a set prediction paradigm. We evaluate our model on two ZeroRTE datasets. The experiment results show that our method consistently outperforms previous works.

Related Papers

• → Alternated Training with Synthetic and Authentic Data for Neural Machine Translation(2021)9 cited
• → Transfer Learning for Driver Pose Estimation from Synthetic Data(2023)4 cited
• → Generation and study of the synthetic brain electron microscopy dataset for segmentation purpose(2023)2 cited
• → Generation and Study of the Synthetic Brain Electron Microscopy Dataset for Segmentation Purpose(2022)