Continual Relation Extraction (CRE) aims to incrementally learn relation knowledge from a non-stationary stream of data, maintaining the stability of the model when new training data is added. Big Data Department of CNIC has proposed a DP-CRE framework that decouples the process of prior information preservation and new knowledge acquisition. By analyzing the change amount of embedding space when new relations appear, the framework improves the process of knowledge preservation and acquisition, enabling incremental learning of relation knowledge from non-stationary stream of data, and providing a new perspective for continual relation extraction. Experiment results show that the model has achieved significant improvements in multiple indicators.

This study has been accepted by The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LERC-COLING24, CCF B). Master student Huang Mengyi and postdoctoral researcher Xiao Meng are the co-first authors of the paper. Researcher Du Yi and Senior Engineer Wang Ludi are the co-corresponding authors.

This achievement is supported by the National Key Research and Development Plan of China, the Natural Science Foundation of China, etc.

DecouPled Framework of Continual Relation Extraction 

Related Achievements:
Mengyi Huang, Meng Xiao, Ludi Wang, Yi Du. DP-CRE: Continual Relation Extraction via Decoupled Contrastive Learning and Memory Structure Preservation[C].// LREC-COLING 2024: The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation, 2024.（Accepted）