Abstract
Continual relation extraction (CRE) aims to extract relations towards the continuous and iterative arrival of new data. To address the problem of catastrophic forgetting, some existing research endeavors have focused on exploring memory replay methods by storing typical historical learned instances or embedding all observed relations as prototypes by averaging the hidden representation of samples and replaying them in the subsequent training process. However, this prototype generation method overlooks the rich semantic information within the label namespace and are also constrained by memory size, resulting in inadequate descriptions of relation semantics by relation prototypes. To this end, we introduce an approach termed Label-Guided Relation Prototype Generation. Initially, we enhance the representations of label embeddings through a technique named label knowledge infusion. Following that, we utilize the multi-head attention mechanism to form relation prototypes, allowing them to capture diverse aspects of typical instances. The embeddings of relation labels are utilized at this stage, leveraging their contained semantics. Additionally, we propose a feature-based distillation loss function called multi-similarity distillation, to ensure the model retains prior knowledge after learning new tasks. The experimental results indicate that our method has achieved competitive performance compared to the state-of-the-art baseline models in CRE.