.png){kind=icon}
.png){kind=icon}
Return
Retrieving memory as prompts for continual relation extraction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2024.124542.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Continual relation extraction is an advanced form of traditional relation extraction that is specifically designed to handle continuously emerging new relations. Sharing the common challenge with continual learning, continual relation extraction encounters the problem of catastrophic forgetting. The task requires learning to extract new relations while retaining knowledge of previously learned ones. Although keeping some samples of previous tasks can alleviate this problem, the risk of overfitting memory samples remains. To tackle the challenges, we retrieve memory as prompts and propose a Memory Retrieval Enhanced Model (MREM) for continual relation extraction. We reinforce sample representations by retrieving the memory repository and employing memory as prompts instead of overfitting a limited number of memory samples. Accordingly, a gated prompt fusion decoder and a consistent relation classifier utilizing memory samples are developed for assistance. The comprehensive experiments on benchmark datasets indicate that our proposed model outperforms the baselines and demonstrates its ability to mitigate overfitting to memory samples and alleviate catastrophic forgetting.
Keywords:
Continual relation extraction
Catastrophic forgetting
Prompt learning
Transformer
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W