Incremental few-shot relation classification aims to train a model on minimal training instances to dynamically learn novel relations while retaining prior knowledge of base relations. Current approaches still suffer from two persistent challenges: novel-class overfitting exacerbated by prototype drift, and severe catastrophic forgetting of base-class knowledge caused by inadequate discriminability among similar classes. In this paper, we design a Distance-Weighted merged prototypical network with Calibration (DWproto-C) to address these issues. DWproto-C architecturally integrates three modules: (1) a Distance-Weighted (DW) merged prototype network that provides a simpler and more effective solution to the incompatible feature embedding problem in prototypical networks through exclusive use of scalar distance measures; (2) a Base-Attentive Novel Prototype Calibration (BANPC) module designed to mitigate prototype drift in novel class embeddings induced by distributional bias in support samples, thereby obtaining more accurate novel prototypes; (3) a Hardest Negative Mining (HNM) module designed to mitigate catastrophic forgetting caused by similar-class confusion, which explicitly enforces the distinction between the ground truth class and its most confusable counterpart for each query sample. Our model demonstrates superior performance over current state-of-the-art (SOTA) methods across both the FewRel 1.0 and 2.0 benchmarks. Particularly noteworthy is the significant improvement in novel relation classification accuracy on FewRel 2.0, achieving a 11.75% gain compared to baseline approaches in 5-shot setting, which substantiates the model’s exceptional capability for incremental few-shot domain adaptation learning.
Na et al. (Fri,) studied this question.