Research map / Label-efficient and robust learning / Recognition and long-tail
MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive 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 MCL-NER, which reframes cross-lingual NER as token-pair relation classification and trains with multi-view contrastive losses (source vs. code-switched sentence semantics plus token-to-token relations) and self-training on unlabeled target data.
Problem
Cross-lingual NER suffers large performance drops for low-resource and non-English languages; prior methods align semantic spaces but ignore token-to-token relations within entities across languages.
Approach
Recast NER as token-pair relation classification (start-end, neighbor, and extra-entity relations) and align both sentence-level semantics and token-pair relation representations between source, code-switched, and translated sentences via contrastive learning, combined with self-training with MSE consistency.
Key results
- On XTREME-40 with XLM-R base, average F1 of 69.7 versus 68.2 for CROP (+1.5).
- On XTREME-40 with mBERT base, average F1 of 70.4 versus 68.5 for CROP.
- On CoNLL (de, es, nl), average F1 of 80.5 versus 79.2 for CROP and 78.2 for TSL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use MCL-NER instead of CROP | mBERT cross-lingual NER · XTREME-40 Avg_all | F1: 68.5 -> 70.4 | Helped |
| use MCL-NER instead of CROP | XLM-R cross-lingual NER · XTREME-40 Avg_all | F1: 68.2 -> 69.7 | Helped |
| use MCL-NER instead of CROP | mBERT cross-lingual NER · CoNLL Avg | F1: 79.2 -> 80.5 | Helped |
| add semantic contrastive learning (ablation) | MCL-NER · XTREME-Avg | F1 | Helped |
| add token-to-token relation contrastive learning (ablation) | MCL-NER · XTREME-Avg | F1 | Helped |
| add semantic and token-to-token relation contrastive learning (ablation) | MCL-NER · XTREME-Avg | F1 | Helped |
Surprising findings
- Training with only 10% of the source data reportedly yields large gains, according to the authors.
Limitations
- Gains on some languages are small or negative (e.g., ka, ml, ur with mBERT in comparisons; ar on XLM-R is lower than CROP on some columns).
- Self-training pseudo-label noise requires thresholding and filtering heuristics that are not fully specified.
Experimental setup
Datasets: XTREME-40 (WikiAnn) CoNLL-02 CoNLL-03 code-switched sentences built from translations
Benchmarks: XTREME-40 CoNLL-02 CoNLL-03
Metrics: F1
Baselines: mBERT Translate Train UniTrans CROP TSL XLM-R MulDA Meta-learning-based
Builds on: CROP UniTrans TSL mBERT XLM-R
Closest papers by meaning
- ConsistNER: Towards Instructive NER Demonstrations for LLMs with the Consistency of Ontology and Context (AAAI 2024 · similarity 0.88)
- GuideNER: Annotation Guidelines Are Better than Examples for In-Context Named Entity Recognition (AAAI 2025 · similarity 0.88)
- UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · similarity 0.88)
- HLMEA: Unsupervised Entity Alignment Based on Hybrid Language Models (AAAI 2025 · similarity 0.87)
- Multi-Grained Query-Guided Set Prediction Network for Grounded Multimodal Named Entity Recognition (AAAI 2025 · similarity 0.87)
- Three Heads Are Better than One: Improving Cross-Domain NER with Progressive Decomposed Network (AAAI 2024 · similarity 0.86)
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.