Research map / Continual learning / Online and replay-based
Few-Shot, No Problem: Descriptive Continual 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 DCRE, a retrieval-based few-shot continual relation extraction method that uses LLM-generated relation descriptions in a bi-encoder training scheme and reciprocal-rank-fusion inference (DRI).
Problem
Few-shot continual relation extraction suffers from catastrophic forgetting and overfitting because memory buffers hold only a few samples, and data scarcity hinders augmentation.
Approach
Use LLM-generated multiple detailed descriptions per relation as stable class anchors during training (hard margin and mutual information losses on description vectors), and predict via a reciprocal rank fusion of Euclidean prototype distance and description cosine similarity rather than nearest class mean.
Key results
- On FewRel 10-way 5-shot with BERT, DCRE reaches 68.24 final-task accuracy in the DRI ablation table, with a stated gap of up to 3.74% over CPL.
- On TACRED 5-way 5-shot with BERT, DCRE reaches Delta 22.99 versus 28.88 for CPL, and final accuracy 63.21 vs NCM 58.26.
- With LLM2Vec, DCRE achieves Delta 16.58 on FewRel and 9.96 on TACRED, versus 22.05 and 14.91 for CPL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DRI instead of NCM prediction | DCRE · FewRel | accuracy: 66.93 -> 68.24 (BERT); 79.26 -> 80.34 (LLM2Vec) | Helped |
| use DRI instead of NCM prediction | DCRE · TACRED | accuracy: 58.26 -> 63.21 (BERT); 75.00 -> 79.10 (LLM2Vec) | Helped |
| add sample contrastive loss (ablation) | DCRE · FewRel | accuracy: 67.58 -> 68.24 (BERT); 78.39 -> 80.34 (LLM2Vec) | Helped |
| add sample contrastive loss (ablation) | DCRE · TACRED | accuracy: 62.11 -> 63.21 (BERT); 77.01 -> 79.10 (LLM2Vec) | Helped |
| add mutual information loss (ablation) | DCRE · FewRel | accuracy: 65.10 -> 68.24 (BERT); 70.61 -> 80.34 (LLM2Vec) | Helped |
| add mutual information loss (ablation) | DCRE · TACRED | accuracy: 57.23 -> 63.21 (BERT); 74.17 -> 79.10 (LLM2Vec) | Helped |
| add hard margin loss (ablation) | DCRE · FewRel | accuracy: 66.20 -> 68.24 (BERT); 77.22 -> 80.34 (LLM2Vec) | Helped |
| add hard margin loss (ablation) | DCRE · TACRED | accuracy: 62.46 -> 63.21 (BERT); 74.75 -> 79.10 (LLM2Vec) | Helped |
| add hard soft margin triplet loss (ablation) | DCRE · FewRel | accuracy: 67.54 -> 68.24 (BERT); 77.48 -> 80.34 (LLM2Vec) | Helped |
| add hard soft margin triplet loss (ablation) | DCRE · TACRED | accuracy: 59.56 -> 63.21 (BERT); 73.77 -> 79.10 (LLM2Vec) | Helped |
| increase generated expert descriptions from K = 3 toward K = 7 (ablation) | DCRE · TACRED (5-way-5-shot) | accuracy | Helped |
Surprising findings
- ConPL results were reproduced by the authors rather than taken from the original paper.
Limitations
- Relies on memory buffer of few samples per relation
- Performance gains on LLM2Vec are smaller on FewRel than on TACRED in some comparisons
Experimental setup
Datasets: FewRel TACRED
Benchmarks: FewRel TACRED
Metrics: accuracy per task forgetting Delta
Baselines: RP-CRE CRL CRECL ERDA SCKD ConPL CPL CPL+MI NCM
Builds on: CPL supervised contrastive loss memory replay LLM2Vec nearest class mean
Closest papers by meaning
- Continual Relation Extraction via Sequential Multi-Task Learning (AAAI 2024 · similarity 0.89)
- Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective (AAAI 2025 · similarity 0.86)
- From RAG to Memory: Non-Parametric Continual Learning for Large Language Models (ICML 2025 · similarity 0.86)
- Enhancing Low-Resource Relation Representations through Multi-View Decoupling (AAAI 2024 · similarity 0.86)
- Knowledge Graph Enhanced Generative Multi-modal Models for Class-Incremental Learning (NeurIPS 2025 · similarity 0.84)
- MoEMeta: Mixture-of-Experts Meta Learning for Few-Shot Relational Learning (NeurIPS 2025 · 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.