Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Liberating Seen Classes: Boosting Few-Shot and Zero-Shot Text Classification via Anchor Generation and Classification Reframing
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.
Generates pseudo-samples with GPT2-XL as per-class anchors and reframes multi-class text classification as BERT-based query-anchor binary similarity, enabling few-shot and zero-shot text classification without seen-class training.
Problem
Few-shot and zero-shot text classification rely on transferring knowledge from seen classes, which is difficult and inefficient when labeled novel samples are scarce and classes are dissimilar.
Approach
Use a pre-trained LM to generate pseudo samples from label-description templates, keep the samples closest to each class prototype as anchors, and classify queries by binary similarity to anchors, never training on seen classes.
Key results
- On 1-shot 20News, Ours-top reaches 0.7310 accuracy vs 0.7229 for ContrastNet (best baseline).
- On CLINC zero-shot, Ours-top reaches 0.9787 accuracy vs 0.9357 for CTIR-ZSDNN.
- On SNIPS zero-shot, Ours-top gets 0.9625 accuracy vs 0.9613 for CTIR-BERT, a limited margin attributed to only two unseen classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use anchor generation and classification reframing instead of ContrastNet | Ours-top · HuffPost 1-shot | accuracy: 0.5045 -> 0.6914 | Helped |
| Use anchor generation and classification reframing instead of ContrastNet | Ours-top · Reuters 5-shot | F1: 0.9417 -> 0.9612 | Helped |
| Use anchor generation and classification reframing instead of ContrastNet | Ours-top · Amazon 1-shot | accuracy: 0.7606 -> 0.7888 | Helped |
| Use anchor generation and classification reframing instead of ContrastNet | Ours-top · 20News 5-shot | accuracy: 0.8418 -> 0.8288 | Hurt |
| Use anchor generation and classification reframing instead of CTIR-ZSDNN | Ours-top · CLINC zero-shot | accuracy: 0.9357 -> 0.9787 | Helped |
| Use anchor generation and classification reframing instead of CTIR-BERT | Ours-top · SNIPS zero-shot | accuracy: 0.9613 -> 0.9625 | No clear effect |
| Increase the number of category anchors (ablation) | Proposed method · News and review datasets, 5-way 5-shot | accuracy | Helped |
Surprising findings
- The 0-shot variant of the method beats several 1-shot baselines on HuffPost and Reuters, e.g. HuffPost 0.6837 vs 0.5045 for ContrastNet (1-shot).
Limitations
- Improvement is limited on SNIPS, which has only two unseen classes, so pairwise reframing offers little benefit.
- Results depend on generation quality from the PLM and on hand-designed label-specific templates.
- Only single-label classification is handled; multi-label extension is left to future work.
Experimental setup
Datasets: 20News Amazon HuffPost Reuters SNIPS CLINC
Benchmarks: 20News Amazon HuffPost Reuters SNIPS CLINC
Metrics: accuracy weighted F1
Baselines: NN Prototypical Network MAML Induction Network TPN DS-FSL MLADA ContrastNet CDSSM ZSDNN CapsNet CTIR
Builds on: ContrastNet DNNC label smoothing
Closest papers by meaning
- BTZSC: A Benchmark for Zero-Shot Text Classification Across Cross-Encoders, Embedding Models, and Rerankers (ICLR 2026 · similarity 0.88)
- Boosting Few-Shot Text Classification via Distribution Estimation (AAAI 2023 · similarity 0.87)
- I2MVFormer: Large Language Model Generated Multi-View Document Supervision for Zero-Shot Image Classification (CVPR 2023 · similarity 0.87)
- Multi-Cue Zero-Shot Learning With Strong Supervision (CVPR 2016 · similarity 0.87)
- Envisioning Class Entity Reasoning by Large Language Models for Few-shot Learning (AAAI 2025 · similarity 0.87)
- Improving Generalized Zero-Shot Learning by Exploring the Diverse Semantics from External Class Names (CVPR 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.