Research map / Adversarial robustness and security / Certified and empirical robustness
Adversity-aware Few-shot Named Entity Recognition via Augmentation Learning
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 AAL, a two-stage few-shot NER method that uses prompt-based data augmentation, variational prototype learning and uncertainty-injected domain-oriented prototypes to stay robust to synonym-based adversarial attacks.
Problem
Few-shot named entity recognition degrades under cross-domain transfer and synonym-based adversarial attacks on target-domain text, where existing prototype-based methods are fragile.
Approach
Augments the support set with two prompt schemes (context-enhancing and oriented-inducing), learns Gaussian-based (Jensen-Shannon divergence) prototypes over augmented and original support/query tokens, and injects random noise into domain-oriented prototypes to improve robustness to BERT-Attack perturbations.
Key results
- On SNIPS under certain conditions, AAL averages 6.27% (1-shot) and 3.04% (5-shot) overall gains over MANNER.
- On attacked SNIPS 1-shot, AAL reaches 67.37 F1 on We vs MANNER 62.11, with MANNER dropping ~7.94% average on SNIPS under attack.
- On attacked Cross-Dataset 1-shot News, AAL achieves 41.25 vs MANNER 38.57 F1 on Cross-Dataset under attack.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AAL instead of MANNER | Few-shot NER · SNIPS (1-shot and 5-shot) and Cross-Dataset (1-shot and 5-shot) | F1: +6.27%; +3.04% | Helped |
| Use AAL instead of MANNER | Few-shot NER · SNIPS Se, 1-shot | F1: +12.68% | Helped |
| Use AAL instead of MANNER | Few-shot NER · Cross-Dataset Wiki, 1-shot | F1: +8.05% | Helped |
| Use AAL instead of MANNER | Few-shot NER under synonym attacks · SNIPS and Cross-Dataset | F1 | Helped |
| Add domain-oriented prototypes with residual connections (ablation) | AAL · Cross-Dataset | F1 | Helped |
| Add prototype learning to capture common entity knowledge (ablation) | AAL · Cross-Dataset | F1 | Helped |
| Add context-enhancing prompt augmentation (ablation) | AAL · Cross-Dataset | F1 | Helped |
| Add the full augmentation learning module (ablation) | AAL · Cross-Dataset | F1 | Helped |
Surprising findings
- MANNER, a leading few-shot NER model, loses about 7.94% (SNIPS) and 9.07% (Cross-Dataset) average F1 under BERT-Attack, and Decp-MetaNER and BDCP drop sharply on attacked SNIPS.
Limitations
- Robustness evaluated only with one attack type (BERT-Attack synonym substitution) on randomly selected adversarial examples.
- Adversarial examples are generated on target-domain support and query sets, and the uncertainty mechanism is only briefly motivated.
- Relies on a BERT base backbone; no evaluation with larger PLMs or LLMs.
Experimental setup
Datasets: SNIPS Cross-Dataset (CoNLL-2003, GUM, WNUT-2017, OntoNotes)
Benchmarks: SNIPS Cross-Dataset attacked SNIPS attacked Cross-Dataset
Metrics: F1-score
Baselines: TransferBERT M-Network L-TapNet+CDT ProtoBERT ESD MANNER SimBERT CONTaiNER Decp-MetaNER SpanProto MSDP BDCP
Builds on: MANNER prototypical networks Decp-MetaNER BERT
Closest papers by meaning
- Robust Few-Shot Named Entity Recognition with Boundary Discrimination and Correlation Purification (AAAI 2024 · similarity 0.91)
- RoPDA: Robust Prompt-Based Data Augmentation for Low-Resource Named Entity Recognition (AAAI 2024 · similarity 0.87)
- UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · similarity 0.86)
- MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning (AAAI 2024 · similarity 0.85)
- ConsistNER: Towards Instructive NER Demonstrations for LLMs with the Consistency of Ontology and Context (AAAI 2024 · similarity 0.85)
- Multi-Grained Query-Guided Set Prediction Network for Grounded Multimodal Named Entity Recognition (AAAI 2025 · similarity 0.84)
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.