Research map / Continual learning / Online and replay-based
Unify Named Entity Recognition Scenarios via Contrastive Real-Time Updating Prototype
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 CRUP, a unified NER framework for supervised, class-incremental, and online settings that trains a Gaussian-projected contrastive encoder and updates class prototypes in real time using only the current batch (batch-wise semantic drift compensation).
Problem
Continual named entity recognition suffers from catastrophic forgetting; existing distillation-based and memory-based methods degrade on long task sequences and are mostly limited to class-incremental settings rather than online streams.
Approach
Stores only class prototypes (not samples) and updates existing prototypes using semantic-drift estimates computed from the current batch's representation changes (batch-wise SDC), with a Gaussian projection head trained by a regularized supervised contrastive KL-divergence loss, and classifies by nearest class mean.
Key results
- On CoNLL-03 T4 class-incremental, CRUP reaches 73.2 F1 vs 71.4 for GDumb and 67.1 for ODDL; on OntoNotes T6 it reaches 58.4 vs 54.3 for GDumb.
- On FewNER-66 CI (T60) CRUP achieves 9.10 F1 in the Table 3 variant row, while the main Table 1 reports 21.6 at T40-equivalent sampling, with CRUP outperforming baselines on SO at most steps.
- In the supervised setting on FewNER-66 and SO, CRUP reaches 58.91 and 57.05 F1 versus 54.48 and 49.59 for Bert-Tagger.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CRUP instead of ChatGPT | NER continual learning · CoNLL-03 CI T4 | F1: 73.2 vs 75.6 | Hurt |
| Use CRUP instead of ChatGPT | NER continual learning · OntoNotes-5.0 CI T6 | F1: 57.3 -> 58.4 | Helped |
| Use CRUP instead of ChatGPT | NER continual learning · FewNER-8 CI T8 | F1: 48.6 -> 51.2 | Helped |
| Use CRUP instead of ChatGPT | NER continual learning · FewNER-66 CI T60 | F1: 11.0 -> 11.1 | Helped |
| Use CRUP instead of DSDM | NER continual learning · SO CI T18 | F1: 9.34 -> 9.73 | Helped |
| Use CRUP instead of Bert-Tagger | supervised NER · FewNER-66 | F1: 54.48 -> 58.91 | Helped |
| Use CRUP instead of Bert-Tagger | supervised NER · SO | F1: 49.59 -> 57.05 | Helped |
| Use BSDC prototype updating and contrastive learning rather than CE with heuristic prototype updates (ablation) | CRUP · FewNER-8 CI T8 | F1: 4.39 -> 49.6 | Helped |
| Add the designed regularization to BSDC-Con (ablation) | CRUP · FewNER-8 CI T8 | F1: 49.6 -> 51.2 | Helped |
| Replace EWC with the designed prototype regularization (ablation) | CRUP · FewNER-8 CI T8 | F1: 48.2 -> 51.2 | Helped |
| Replace SI with the designed prototype regularization (ablation) | CRUP · FewNER-8 CI T8 | F1: 48.7 -> 51.2 | Helped |
| Use Gaussian projection instead of point embeddings (ablation) | CRUP · FewNER-8 CI T8 | F1: 0.85 -> 51.2 | Helped |
Surprising findings
- Distillation-based methods (AddNER, ExtendNER, L&R) perform well on early tasks but degrade substantially on later tasks, and simple heuristics like Replace-CE and Mean-CE sometimes beat AddNER and ExtendNER.
- EWC and SI regularization improve little or even hurt performance relative to the prototype-based regularizer.
Limitations
- CRUP involves high complexity during optimization.
- Performance on early FewNER-66 tasks degrades relative to some baselines, and the regularizer is designed for the CI scenario only.
Experimental setup
Datasets: CoNLL-03 OntoNotes-5.0 FewNER-8 FewNER-66 StackOverflowNER
Benchmarks: CoNLL-03 OntoNotes-5.0 FewNER-8 FewNER-66 StackOverflowNER
Metrics: F1 precision recall average F1 over steps (AF1)
Baselines: Bert-Tagger AddNER ExtendNER L&R GDumb CoPE ODDL DSDM ChatGPT-3.5
Builds on: supervised contrastive learning semantic drift compensation (SDC) SCR CoPE Gaussian embeddings
Closest papers by meaning
- Prototypical Replay with Old-class Focusing Knowledge Distillation for Incremental Named Entity Recognition (AAAI 2025 · similarity 0.87)
- MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning (AAAI 2024 · similarity 0.86)
- ConsistNER: Towards Instructive NER Demonstrations for LLMs with the Consistency of Ontology and Context (AAAI 2024 · similarity 0.84)
- GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition (AAAI 2025 · similarity 0.84)
- A Reasoning Paradigm for Named Entity Recognition (AAAI 2026 · similarity 0.84)
- UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · 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.