Research map / Label-efficient and robust learning / Recognition and long-tail
Enhancing Low-Resource Relation Representations through Multi-View Decoupling
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 MVRE, a prompt-tuning relation extraction method that decouples each relation into multiple virtual-word views predicted by successive [MASK] tokens, with a Global-Local loss and Dynamic Initialization, improving low-resource RE.
Problem
Prompt-based relation extraction in low-resource (1-, 5-, 16-shot) settings yields poor relation representations because a complex relation is confined to a single virtual word vector.
Approach
Decouple each relation into m multi-view virtual words predicted at successive [MASK] positions, align them with a Global-Local contrastive-style loss, and initialize them with a blend of static label-word embeddings and dynamic MLM-predicted tokens.
Key results
- In 1-shot SemEval, MVRE reaches 54.6 micro F1 versus 33.3 for RetrievalRE, a reported gain of about 63.9%.
- On TACRED and TACREV at 16-shot, MVRE is slightly below RetrievalRE (34.8 vs 36.1 and 34.3 vs 35.3).
- Under the full-data standard setting on SemEval, MVRE reaches 90.2 F1, roughly on par with KnowPrompt and RetrievalRE (90.4).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MVRE rather than RetrievalRE | Relation extraction with RoBERTa-large · SemEval K=1 | micro F1 (%): 33.3{ ($\pm 1.6$)} -> 54.6{ ($\pm 2.8$)} | Helped |
| Use MVRE rather than RetrievalRE | Relation extraction with RoBERTa-large · TACRED K=1 | micro F1 (%): 19.5{ ($\pm 1.5$)} -> 21.2{ ($\pm 2.2$)} | Helped |
| Use MVRE rather than RetrievalRE | Relation extraction with RoBERTa-large · TACREV K=1 | micro F1 (%): 18.7{ ($\pm 1.8$)} -> 20.5{ ($\pm 1.9$)} | Helped |
| Use MVRE rather than RetrievalRE | Relation extraction with RoBERTa-large · TACRED and TACREV K=16 | micro F1 (%): 36.1{ ($\pm 1.2$)} -> 34.8{ ($\pm 0.8$)}; 35.3{ ($\pm 0.3$)} -> 34.3{ ($\pm 1.1$)} | Hurt |
| Add Global-Local Loss (ablation) | MVRE · SemEval K=5 | F1: 74.6 -> 77.6 | Helped |
| Add Global-Local Loss (ablation) | MVRE · SemEval K=16 | F1: 82.1 -> 82.5 | Helped |
| Add Global-Local Loss (ablation) | MVRE · SemEval K=1 | F1: 54.6 -> 54.6 | No clear effect |
| Add Dynamic Initialization (ablation) | MVRE with Static Initialization · SemEval K=1 | F1: 37.5 -> 54.6 | Helped |
| Add Dynamic Initialization (ablation) | MVRE with Static Initialization · SemEval K=5 | F1: 72.9 -> 77.6 | Helped |
| Add Dynamic Initialization (ablation) | MVRE with Static Initialization · SemEval K=16 | F1: 81.5 -> 82.5 | Helped |
| Add Static Initialization (ablation) | MVRE with Dynamic Initialization · SemEval K=1 | F1: 43.3 -> 54.6 | Helped |
| Add Static Initialization (ablation) | MVRE with Dynamic Initialization · SemEval K=5 | F1: 73.1 -> 77.6 | Helped |
Surprising findings
- Simply increasing the number of [MASK] tokens does not keep improving performance; performance peaks around m=3 to 5 and declines afterward, attributed to noise.
- Gains are much smaller on TACRED and TACREV than on SemEval, attributed to the high proportion of 'other' relations and similar relation labels.
Limitations
- Gains diminish or reverse at higher shot counts on TACRED and TACREV (e.g., 16-shot below RetrievalRE).
- Performance degrades with many [MASK] tokens due to introduced noise, so m must be tuned.
- Dynamic initialization requires manually written per-relation templates.
Experimental setup
Datasets: SemEval 2010 Task 8 TACRED TACREV
Benchmarks: SemEval 2010 Task 8 TACRED TACREV
Metrics: micro F1
Baselines: FINE-TUNING GDPN_ET PTR KnowPrompt RetrievalRE
Builds on: KnowPrompt RetrievalRE PTR
Closest papers by meaning
- DocKS-RAG: Optimizing Document-Level Relation Extraction through LLM-Enhanced Hybrid Prompt Tuning (ICML 2025 · similarity 0.87)
- Enhancing Relation Extraction via Supervised Rationale Verification and Feedback (AAAI 2025 · similarity 0.86)
- Few-Shot, No Problem: Descriptive Continual Relation Extraction (AAAI 2025 · similarity 0.86)
- Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective (AAAI 2025 · similarity 0.86)
- Synergistic Anchored Contrastive Pre-training for Few-Shot Relation Extraction (AAAI 2024 · similarity 0.85)
- Prototype-Guided Multimodal Relation Extraction based on Entity Attributes (AAAI 2025 · 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.