Research on continuous relation extraction by integrating multi-level features

Continuous relation extraction (CRE) is a key continuous learning task that aims to continuously learn and identify new entity relations from text. However, existing research mainly focuses on memory replay methods to avoid the catastrophic forgetting of old relations by the model. However, these methods often ignore the rich semantic information in the pre-trained language model (PLM) and do not fully consider the differences between samples, resulting in insufficient ability of the model to handle complex relations. The model also often over-relies on the features of the memory samples, which makes it unstable when facing new samples with large heterogeneity. To solve these problems, this paper proposes a new CRE method (MHS-CRE). Based on the relational knowledge of PLM, this method obtains richer contextual information by fusing multi-level features in the sentence encoding process. The hierarchical clustering idea is introduced to make the granularity of the model's relation classification finer. Through the improved typical sample refinement module, combined with residual connection and gating function, the model's adaptability to new relations is enhanced. Experimental results show that the MHS-CRE model shows stronger adaptability and generalization ability when dealing with complex relations and heterogeneous samples. Experimental results on the public datasets FewRel and TACRED show that the classification accuracy of the MHS-CRE method is improved by 1.29% and 5.77% respectively.