Reproduce “LLM-DA: Data Augmentation via Large Language Models for Few-Shot Named Entity Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Junjie Ye, Nuo Xu, Yikun Wang, Jie Zhou, Qi Zhang, Tao Gui, Xuanjing Huang
Researchers from Fudan University and East China Normal University introduce LLM-DA, a data augmentation technique for few-shot Named Entity Recognition that leverages Large Language Models. This method systematically generates high-quality, diverse augmented data at both contextual and entity levels using 14 distinct strategies and noise injection, leading to improved performance over existing baselines across various datasets in low-resource settings.
Key insights
- LLMs are more effectively utilized as intelligent data generators for downstream NER models rather than for direct task execution, due to their superior rewriting capabilities.
- Maintaining high linguistic quality (semantic coherence, grammatical correctness) and balanced entity distribution in augmented data is critical for improving downstream model performance, often more so than just increasing data quantity.
- The effectiveness of context-level vs. entity-level augmentation varies with data scarcity: context diversification is more impactful in extreme low-shot settings, while entity-level augmentation gains importance with larger initial datasets.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.