Research map / Label-efficient and robust learning / Recognition and long-tail
Description Based Text Classification with Reinforcement Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add class descriptions using the template strategy | BERT text classification · Single-label benchmarks: AGNews, 20news, DBPedia, Yahoo! Answers, YelpP, IMDB | test error rate: -0.7 on AGNews, -1.1 on 20news, -0.07 on DBPedia, -0.6 on Yahoo, -0.2 on YelpP and -1.0 on IMDB | Helped |
| Replace fixed templates with RL-generated descriptions | Description-based text classification · Single-label benchmarks | test error rate | Helped |
| Use extractive rather than abstractive description generation | Description-based text classification · AGNews and IMDB versus other single-label benchmarks | test error rate | Mixed |
| Add class descriptions | BERT text classification · Reuters and AAPD | test error rate: 11.0 -> 10.0 on Reuters; 26.6 -> 25.7 on AAPD | Helped |
| Add aspect-specific class descriptions | BERT text classification · BeerAdvocate and TripAdvisor | test error rate: 27.8 -> 15.6 on BeerAdvocate; 35.6 -> 17.0 on TripAdvisor | Helped |
| Replace label-index descriptions with keyword descriptions (ablation) | Template Description · 20news | error rate: 16.8 -> 16.4 | Helped |
| Replace keyword descriptions with keyword-expansion descriptions (ablation) | Template Description · 20news | error rate: 16.4 -> 16.2 | Helped |
| Use Wikipedia definitions instead of keyword expansions (ablation) | Template Description · 20news | error rate: 16.2 -> 15.8 | Helped |
| Initialize the extractive RL policy with ROUGE-L spans instead of dummy tokens (ablation) | Extractive description model · Yahoo! Answers and AAPD | error rate: 22.2 -> 25.3 on Yahoo Answer; 26.0 -> 27.2 on AAPD | Hurt |
| Initialize the extractive RL policy randomly instead of with dummy tokens (ablation) | Extractive description model · Yahoo! Answers and AAPD | error rate: 22.2 -> 28.0 on Yahoo Answer; 26.0 -> 30.1 on AAPD | Hurt |
| Initialize the abstractive RL policy randomly instead of with templates (ablation) | Abstractive description model · Yahoo! Answers and AAPD | error rate: 22.0 -> 87.9 on Yahoo Answer; 25.7 -> 78.4 on AAPD | Hurt |
Closest papers by meaning
- Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach (AAAI 2024 · similarity 0.85)
- Incremental Sequence Classification with Temporal Consistency (NeurIPS 2025 · similarity 0.84)
- Gen-Z: Generative Zero-Shot Text Classification with Contextualized Label Descriptions (ICLR 2024 · similarity 0.84)
- Boosting Short Text Classification with Multi-Source Information Exploration and Dual-Level Contrastive Learning (AAAI 2025 · similarity 0.84)
- Improved Graph Contrastive Learning for Short Text Classification (AAAI 2024 · similarity 0.83)
- Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision (NeurIPS 2025 · similarity 0.83)
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.