Research map / Label-efficient and robust learning / Domain adaptation and generalization
Three Heads Are Better than One: Improving Cross-Domain NER with Progressive Decomposed Network
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 Progressive Decomposed Networks that split NER into mention detection and entity typing, transferring knowledge from multiple source-domain NER models explicitly (copied potential entities) and implicitly (knowledge progressive networks) to low-resource cross-domain targets.
Problem
Cross-domain NER with sparse target-domain labels where label sets mismatch across domains, and prior pretrain-finetune or joint-training paradigms fail to exploit multiple source domains effectively.
Approach
Decouple NER into a domain-unified mention detection (B/I/O) task and a domain-specific entity typing task sharing the encoder; inject potential entities predicted by source models into the target input and fuse source-model hidden states via learned Knowledge Progressive Networks; analyze source contribution with perturbation sensitivity and Fisher information.
Key results
- Averaged over 8 target datasets, the method improves F1 by about 12.50% over models without source-domain data.
- On the AI target domain, F1 rises from 50.46 (BERT-CRF, no source) to 70.05, a 19.59-point gain.
- Versus prior SOTA DH-GAT, averages +0.66 F1 on English targets and +1.39 F1 on Chinese targets, with Student's t-test p<0.05.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Progressive Decomposed Network with progressive transfer from multiple source domains | cross-domain NER model · Politics | F1: 76.88 -> 77.82 | Helped |
| Add potential-entity prompts to the target-domain input (ablation) | Progressive Decomposed Network · AI | F1: 69.25 -> 70.05 | Helped |
| Add Knowledge Progressive Networks to transfer source-domain knowledge (ablation) | Progressive Decomposed Network · AI | F1: 68.08 -> 70.05 | Helped |
| Use progressive transfer instead of joint training (ablation) | Progressive Decomposed Network · AI | F1: 68.53 -> 70.05 | Helped |
| Use a decomposed mention-detection and entity-typing network instead of merging the tasks (ablation) | Progressive Decomposed Network · AI | F1: 68.12 -> 70.05 | Helped |
| Use a decomposed network instead of a Structured Semantic Network for entity typing (ablation) | Progressive Decomposed Network · AI | F1: 68.11 -> 70.05 | Helped |
| Order source datasets as WNUT, OntoNotes, then CoNLL (ablation) | Progressive Decomposed Network · Politics | F1: 77.07 -> 77.82 | Helped |
Surprising findings
- Source-model sensitivity analysis shows CoNLL 2003 and MSRA transfer the most knowledge, while WNUT and OntoNotes transfer least, matching which sources help most in ablations.
- Replacing the decomposed network with a structured semantic network for typing drops performance by 2.14% on average.
Limitations
- Chinese target datasets use only 0.2k training samples, and the approach depends on available general-domain source NER data.
- Fisher and perturbation analyses are approximate and perturbation analysis is costly and affected by randomness.
Experimental setup
Datasets: CoNLL 2003 WNUT 17 OntoNotes 5.0 MSRA Politics Natural Science Music Literature AI Resume Address Biomedical
Benchmarks: Liu et al. 2021 English cross-domain NER targets (Politics, Science, Music, Literature, AI) Chinese cross-domain NER targets (Resume, Address, Biomedical)
Metrics: F1 (entity mention and type matching) F1 (entity mention matching)
Baselines: BiLSTM-CRF Coach LM-NER MultiCell-LM BERT-JF BERT-PF Style-NER LST-NER LANER DoSEA MTD MTD-MoCL
Builds on: BERT Liu et al. 2021 cross-domain NER benchmark Asghar et al. 2020 progressive neural networks MTD DH-GAT
Closest papers by meaning
- CrossNER: Evaluating Cross-Domain Named Entity Recognition (AAAI 2021 · similarity 0.93)
- A Simple Yet Effective Subsequence-Enhanced Approach for Cross-Domain NER (AAAI 2023 · similarity 0.92)
- UniversalNER: Targeted Distillation from Large Language Models for Open Named Entity Recognition (ICLR 2024 · similarity 0.87)
- MCL-NER: Cross-Lingual Named Entity Recognition via Multi-View Contrastive Learning (AAAI 2024 · similarity 0.86)
- A Multi-Agent LLM Framework for Multi-Domain Low-Resource In-Context NER via Knowledge Retrieval, Disambiguation and Reflective Analysis (AAAI 2026 · similarity 0.86)
- ConsistNER: Towards Instructive NER Demonstrations for LLMs with the Consistency of Ontology and Context (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.