Deep Generative Replay for Continual Relation Extraction

Continual relation extraction (CRE) aims to incrementally acquire new relational knowledge while retaining performance on previously learned relations. A major challenge in this setting is catastrophic forgetting, typically addressed by replay-based methods that store and revisit past data. While effective, these approaches raise privacy concerns and may lead to overfitting. Generative replay (GR) has emerged as a privacy-preserving alternative by synthesizing pseudo-samples that approximate prior task distributions. However, existing GR-based methods often fail to generate diverse and coherent relational instances, limiting their effectiveness. In this work, we propose a novel GR framework tailored for CRE. We first leverage large language models (LLMs) to generate task-specific explanations, improving the interpretability and contextual coherence of generated samples. To enhance generation quality, we introduce a multi-task fine-tuning framework that incorporates prompt engineering techniques for effective relation-specific prompting. Additionally, we propose a prototype-guided refinement strategy that aligns generated relations with learned relation prototypes, enforcing semantic consistency and improving sample fidelity. Extensive experiments on two standard CRE benchmarks demonstrate that our method outperforms prior GR-based and replay-based baselines in mitigating forgetting, while ensuring data privacy and reducing memory overhead.