Research map / Label-efficient and robust learning / Recognition and long-tail
COMM: Concentrated Margin Maximization for Robust Document-Level Relation Extraction
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 COMM, a two-stage DocRE framework that plugs existing relational encoders (ATLOP, DocuNet, KD-DocRE) into an instance-aware reasoning stage and trains them with a Concentrated Margin Maximization loss that reweights positive and easy/hard samples.
Problem
Document-level relation extraction suffers from label noise and extreme sparsity of positive entity-pair relations and skewed relation-type distribution, which biases optimization, especially when trained on low-quality data such as DocRED.
Approach
Adds a margin-based loss with a threshold class where positive logits are up-weighted by a concentration factor (1-q)^gamma, and negatives with confident scores (above 1-m) are filtered out via log-sigmoid rescaling, so the loss emphasizes rare positives and hard samples while de-emphasizing easy negatives.
Key results
- COMM improves Re-DocRED test F1 by 1.0 to 3.3 points over the baseline relational encoders (e.g., COMM-RoBERTa-ATLOP 80.8 vs 77.5).
- When trained on the noisier DocRED-Mixed training data, COMM gains 8.2 to 13.7 test F1 points over baselines (e.g., COMM-BERT-ATLOP 62.3 vs 50.1).
- With BERT-ATLOP, CMM loss reaches 76.12 F1 on Re-DocRED, above ATL (72.87), AFL (73.71) and HingeABL (75.15).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply COMM to BERT-ATLOP | BERT-ATLOP · DocRED-Mixed test | F1: +12.2 | Helped |
| Apply COMM to RoBERTa-ATLOP | RoBERTa-ATLOP · DocRED-Mixed test | F1: +12.9 | Helped |
| Apply COMM to BERT-DocuNet | BERT-DocuNet · DocRED-Mixed test | F1: +13.7 | Helped |
| Apply COMM to RoBERTa-DocuNet | RoBERTa-DocuNet · DocRED-Mixed test | F1: +13.4 | Helped |
| Apply COMM to BERT-KD-DocRE | BERT-KD-DocRE · DocRED-Mixed test | F1: +9.5 | Helped |
| Apply COMM to RoBERTa-KD-DocRE | RoBERTa-KD-DocRE · DocRED-Mixed test | F1: +8.2 | Helped |
| Apply COMM to ATLOP and other encoders | DocRE encoder models · Re-DocRED test | F1: +3.2; +3.3; +1.3; +1.0; +1.1; +1.0 | Helped |
| Replace HingeABL loss with CMM loss (ablation) | COMM-BERT-ATLOP_BASE · DocRED-Mixed | F1: 58.44 -> 62.31 | Helped |
| Replace HingeABL loss with CMM loss (ablation) | COMM-BERT-ATLOP_BASE · Re-DocRED | F1: 75.15 -> 76.12 | Helped |
Surprising findings
- Balanced Softmax and AML underperform the ATL baseline under the DocRED-Mixed setting (46.25 and 43.21 F1 vs 50.09), showing that common imbalance losses do not transfer directly to DocRE.
Limitations
- Gains on the higher-quality Re-DocRED setting are smaller (1.0 to 3.3 F1).
- Hyperparameters m and gamma are chosen by grid search, and the paper does not describe a sensitivity analysis beyond the grid.
Experimental setup
Datasets: DocRED Re-DocRED
Benchmarks: DocRED Re-DocRED
Metrics: F1 Ign F1
Baselines: ATLOP DocuNet KD-DocRE ATL Balanced Softmax AML AFL HingeABL
Builds on: ATLOP DocuNet KD-DocRE adaptive thresholding ATL
Closest papers by meaning
- Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction (AAAI 2024 · similarity 0.89)
- DocKS-RAG: Optimizing Document-Level Relation Extraction through LLM-Enhanced Hybrid Prompt Tuning (ICML 2025 · similarity 0.89)
- A Positive-Unlabeled Metric Learning Framework for Document-Level Relation Extraction with Incomplete Labeling (AAAI 2024 · similarity 0.88)
- A Hierarchical Network for Multimodal Document-Level Relation Extraction (AAAI 2024 · similarity 0.86)
- An Autoregressive Text-to-Graph Framework for Joint Entity and Relation Extraction (AAAI 2024 · similarity 0.86)
- Synergistic Anchored Contrastive Pre-training for Few-Shot Relation Extraction (AAAI 2024 · similarity 0.85)
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.