Research map / Label-efficient and robust learning / Recognition and long-tail
SENCR: A Span Enhanced Two-Stage Network with Counterfactual Rethinking for Chinese NER
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 SENCR, a lexicon-free two-stage Chinese NER model (BERT-BiLSTM boundary detector plus TextCNN-style span type classifier) with a training-free counterfactual rethinking debiasing step on boundary predictions.
Problem
Chinese NER depends on costly, high-quality external lexicons, and models suffer boundary bias from high-frequency boundary characters that hurts generalization to new boundaries.
Approach
A lexicon-free two-stage network with boundary supervision and convolution-based n-gram span representations, trained jointly; at inference, counterfactual rethinking subtracts the logits from masked-context and fully masked inputs to remove boundary and dataset bias without changing parameters.
Key results
- SENCR reaches 73.42 F1 on Weibo, 97.13 on Resume, 96.17 on MSRA and 81.86 on CLUENER, outperforming lexicon-free baselines such as W2NER (72.32 Weibo) and DCSAN.
- The CR strategy improves F1 on new boundaries over SEN by 0.84, 1.40, 1.87 and 1.12 on Resume, CLUENER, Weibo and MSRA respectively.
- Removing multi-task joint training drops F1 by 2.36 on Resume and 3.33 on CLUENER, the largest ablation effect.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply counterfactual rethinking to SEN | SEN · Weibo | F1: 72.73 -> 73.42 | Helped |
| Apply counterfactual rethinking to SEN | SEN · Resume | F1: 96.86 -> 97.13 | Helped |
| Apply counterfactual rethinking to SEN | SEN · MSRA | F1: 96.12 -> 96.17 | Helped |
| Apply counterfactual rethinking to SEN | SEN · CLUENER | F1: 81.61 -> 81.86 | Helped |
| Add span enhancement with convolution-based n-gram features (ablation) | SEN · Resume | F1: 96.26 -> 96.86 | Helped |
| Add span enhancement with convolution-based n-gram features (ablation) | SEN · CLUENER | F1: 80.69 -> 81.61 | Helped |
| Add random negative-span sampling for NoneType augmentation (ablation) | SEN entity classifier · Resume | F1: 96.37 -> 96.86 | Helped |
| Add random negative-span sampling for NoneType augmentation (ablation) | SEN entity classifier · CLUENER | F1: 80.97 -> 81.61 | Helped |
| Use multi-task joint training instead of pipeline training (ablation) | SEN · Resume | F1: 94.50 -> 96.86 | Helped |
| Use multi-task joint training instead of pipeline training (ablation) | SEN · CLUENER | F1: 78.28 -> 81.61 | Helped |
Surprising findings
- Performance does not necessarily improve with more n-gram kernel sizes; longer n-grams than spans may add noise.
Limitations
- Gains over strong lexicon-based models on MSRA are limited (SENCR 96.17 vs W2NER 96.08 lexicon-free only marginally better).
- CR hyperparameters alpha1 and alpha2 must be grid-searched on a validation set per dataset.
- Gains on MSRA and CLUENER are small relative to baselines.
Experimental setup
Datasets: Weibo MSRA Resume CLUENER
Benchmarks: Weibo MSRA Resume CLUENER
Metrics: precision recall F1
Baselines: BERT+FLAT SoftLexicon LEBERT MECT DCSAN MCL BERT+CRF BERT+GAM W2NER
Builds on: TextCNN BERT FLAT W2NER
Closest papers by meaning
- Span Graph Transformer for Document-Level Named Entity Recognition (AAAI 2024 · similarity 0.87)
- Small Language Model Makes an Effective Long Text Extractor (AAAI 2025 · similarity 0.86)
- Robust Few-Shot Named Entity Recognition with Boundary Discrimination and Correlation Purification (AAAI 2024 · similarity 0.85)
- PaDeLLM-NER: Parallel Decoding in Large Language Models for Named Entity Recognition (NeurIPS 2024 · similarity 0.85)
- Three Heads Are Better than One: Improving Cross-Domain NER with Progressive Decomposed Network (AAAI 2024 · similarity 0.85)
- Discrepancy and Uncertainty Aware Denoising Knowledge Distillation for Zero-Shot Cross-Lingual Named Entity Recognition (AAAI 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.