Research map / Label-efficient and robust learning / Recognition and long-tail
RoPDA: Robust Prompt-Based Data Augmentation for Low-Resource Named Entity Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes RoPDA, a T5 soft-prompt-based augmentation method for low-resource NER that uses five masking-and-regeneration operations (label-flipping and label-preserving), Self-Consistency Filtering, and hidden-space mixup.
Problem
Low-resource named entity recognition suffers from data sparsity; prior augmentation methods disrupt syntax, cause token-label mismatch, or need external knowledge or manual effort.
Approach
Prompt-tuned T5 (soft prompts only, trained on few-shot labeled data) regenerates masked entity and context spans under five operations to produce label-flipping (adversarial) and label-preserving examples; a bidirectional Type2Word/Word2Type self-consistency filter removes inconsistent samples, and mixup interpolates adversarial hidden states with original examples.
Key results
- RoPDA reaches 74.1 F1 on CoNLL03 at shot-5 versus 71.1 for PromDA and 65.9 for the baseline.
- With unlabeled data, RoPDA* reaches 75.0 F1 on CoNLL03 at shot-5 and outperforms MetaST across all benchmarks and shots.
- Removing the soft prompt drops F1 by 1.4 on CoNLL03 at shot-20 (80.5 vs 81.9).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RoPDA instead of PromDA | Low-resource NER · CoNLL03 shot-20 | F1: 81.0 -> 81.9 | Helped |
| Use RoPDA instead of PromDA | Low-resource NER · MIT Restaurant shot-20 | F1: 66.7 -> 67.6 | Helped |
| Use RoPDA instead of PromDA | Low-resource NER · MIT Movie shot-20 | F1: 78.5 -> 79.8 | Helped |
| Add augmentation for unlabeled data to RoPDA | RoPDA · CoNLL03 shot-20 | F1: 81.9 -> 85.2 | Helped |
| Add soft prompt (ablation) | RoPDA · CoNLL03 shot-20 | F1: 80.5 -> 81.9 | Helped |
| Add soft prompt (ablation) | RoPDA · MIT Restaurant shot-20 | F1: 67.1 -> 67.6 | Helped |
| Add Self-Consistency Filtering (ablation) | RoPDA · CoNLL03 shot-20 | F1: 81.5 -> 81.9 | Helped |
| Add mixup (ablation) | RoPDA · MIT Movie shot-20 | F1: 79.4 -> 79.8 | Helped |
| Add Entity Adding (EA) (ablation) | RoPDA · CoNLL03 shot-20 | F1: 81.0 -> 81.9 | Helped |
| Add Entity Replacing (ER) (ablation) | RoPDA · MIT Restaurant shot-20 | F1: 67.0 -> 67.6 | No clear effect |
| Use F-Mixup instead of P-Mixup (ablation) | RoPDA · CoNLL03 shot-20 | F1: 81.2 -> 81.9 | Helped |
| Choose flipped entity types randomly instead of using similarity-based schemes (ablation) | RoPDA · CoNLL03 shot-20 | F1: 81.4 -> 81.9 | Helped |
Surprising findings
- Random label-flipping outperforms all entity-type-similarity-based flipping schemes.
- ELC (direct entity type change) contributes least among the four augmentation strategies when combined with the others.
Limitations
- Gains shrink as data size grows; data-rich gains are small (about 0.3-1.6 points).
- Label-flipping samples require mixup and filtering to avoid degradation; ELC introduces noise.
- Relies on T5-Large generation quality and on hyperparameters such as K, M, N chosen empirically.
Experimental setup
Datasets: CoNLL03 MIT Restaurant MIT Movie
Benchmarks: CoNLL03 MIT Restaurant MIT Movie
Metrics: micro-F1
Baselines: Baseline (BERT trained on original data) SDANER MELM PromDA MetaST RoPDA+self-training
Builds on: PromDA MELM SDANER MetaST mixup
Closest papers by meaning
- Adversity-aware Few-shot Named Entity Recognition via Augmentation Learning (AAAI 2025 · similarity 0.87)
- GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition (AAAI 2025 · similarity 0.84)
- Small Language Model Makes an Effective Long Text Extractor (AAAI 2025 · similarity 0.84)
- UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · similarity 0.84)
- Unify Named Entity Recognition Scenarios via Contrastive Real-Time Updating Prototype (AAAI 2024 · similarity 0.84)
- PMRC: Prompt-Based Machine Reading Comprehension for Few-Shot Named Entity Recognition (AAAI 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.