Fine-Grained Prototype-Augmented Named Entity Recognition

Recently, k-nearest-neighbors named entity recognition (kNN-NER) successfully improves the performance of the vanilla NER model by using nearest neighbor search as the retrieval-augmented method. As a non-parametric solution, it has achieved excellent performance in natural language processing. However, due to the complicated process of nearest neighbor search, kNN-NER usually necessitates substantial inference cost, along with extra cost of extensive storage. In addition, for given input tokens, the two crucial hyperparameters \(\lambda \) and \(\mathcal {T}\) in the kNN-NER method are fixed, which limits flexibility and potentially underutilizes retrieved information. To address these limitations, we propose a fine-grained prototype-augmented NER method. Specifically, we first construct the datastore in the kNN-NER method based on the prototype vector of each entity type. Moreover, we explore multi-level strategies for learning two hyperparameters in a dynamically trainable mode to fully utilize the retrieved knowledge. Since our method is model-agnostic, we conduct extensive experiments on three benchmark NER datasets under different base model. The result demonstrates that the proposed method significantly outperforms the vanilla kNN-NER with reduced memory space and increased inference speedup.