Few-shot Data Augmentation for Named Entity Recognition base on Random Sample Partition Verification

Few-shot Named Entity Recognition (FS-NER) aims to identify named entities with only a small amount of annotated data. Existing FS-NER methods can be broadly categorized into two types: model-centric methods, which focus on designing more effective model architectures, and data-centric methods, which attempt to generate additional data beyond the few-shot samples. However, both approaches overlook the hidden information within the labeled few-shot data. To address this limitation, we propose Few-shot Data Augmentation (FSDA). Our method first performs zero-shot predictions and then compares them with the labeled few-shot samples to identify hard negative samples. These hard negatives are fed back into the model, effectively reducing hallucinations and enhancing performance. Furthermore, current methods validate models based on the meta-learning paradigm, where Q samples are randomly selected from each class for validation. The number of samples per class Q impacts the validation results. However, using the full dataset for validation is computationally expensive in terms of both time and GPU resources. To overcome this issue, we introduce Random Sample Partition (RSP) for model validation, which replaces full-data validation with sampled RSP data blocks. The experimental results demonstrate that FSDA significantly enhances the performance of FS-NER. Meanwhile, the RSP-based validation method enables fast and accurate model validation results while significantly reducing computational costs.