Incre-ICAPQ: Iterative Cross Alignment and Prototype Quadruplet Loss for Incremental Few-Shot Relation Classification

Anonymous

Paper Link: https://openreview.net/forum?id=0YP91SXU9id

Paper Type: Long paper (up to eight pages of content + unlimited references and appendices)

Abstract: In incremental few-shot relation classification task, model performance is always limited by incompatibility between base feature embedding space and novel feature embedding space. To tackle the issue, we present a novel method named Incre-ICAPQ with Iterative Cross Alignment and Prototype Quadruplet loss. Specifically, we incorporate the query instance representation into the encoding of novel prototypes and meanwhile utilize the query-aware prototypes to acquire the query instance representation. To achieve better interaction, we further implement the above dual encoding iteratively. Moreover, prototype quadruplet loss enlarges the distance between different types of prototypes, especially the relative distance between base and novel classes, and makes the distance between query and prototype of the same class as close as possible. Experimental results on two benchmarks demonstrate that Incre-ICAPQ significantly outperforms the state-of-the-art baseline model.