Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain Agnostic Real-Valued Specificity Prediction
Wei-Jen Ko, Greg Durrett, Junyi Jessy Li
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| use self-ensembling with consistency loss and distribution regularization | sentence specificity prediction · Twitter, Yelp, and Movie | MAE: 0.237 -> 0.113; 0.325 -> 0.105; 0.226 -> 0.114 | Improved |
| add mean-std distribution regularization to consistency training (ablation) | SE+A · Twitter | MAE: 0.151$\pm$0.005 -> 0.113$\pm$0.001 | Improved |
| add mean-std distribution regularization to consistency training (ablation) | SE+A · Yelp | MAE: 0.137$\pm$0.005 -> 0.107$\pm$0.003 | Improved |
| add mean-std distribution regularization to consistency training (ablation) | SE+A · Movie | MAE: 0.148$\pm$0.001 -> 0.114$\pm$0.001 | Improved |
| add consistency loss without distribution regularization (ablation) | SE · Twitter | MAE: 0.148$\pm$0.006 -> 0.151$\pm$0.005 | Hurt |
| add mean-std distribution regularization without consistency loss (ablation) | SE · Twitter | MAE: 0.148$\pm$0.006 -> 0.125$\pm$0.001 | Improved |
| remove data augmentation (ablation) | SE+AD (mean-std) · Twitter | Spearman: 0.676$\pm$0.004 -> 0.666 | Hurt |
| remove data augmentation (ablation) | SE+AD (mean-std) · Yelp | Spearman: 0.750$\pm$0.016 -> 0.728 | Hurt |
| remove data augmentation (ablation) | SE+AD (mean-std) · Movie | Spearman: 0.702$\pm$0.003 -> 0.669 | Hurt |
| use KL divergence rather than mean and standard deviation to regularize the posterior distribution (ablation) | SE+AD · Twitter, Yelp, and Movie | MAE: Twitter 0.113$\pm$0.001 vs 0.115$\pm$0.002; Yelp 0.107$\pm$0.003 vs 0.105$\pm$0.002; Movie 0.114$\pm$0.001 vs 0.114$\pm$ | No clear effect |
| filter general responses by predicted specificity instead of filtering short responses | OpenSubtitles dialogue generation · OpenSubtitles dialogue | informativeness preference: 29.0 vs 38.9 | Improved |
| filter general responses by predicted specificity instead of filtering short responses | OpenSubtitles dialogue generation · OpenSubtitles dialogue | perplexity: 51.97 -> 46.92 | Improved |
Closest papers by meaning
- Task-Adaptive Pretrained Language Models via Clustered-Importance Sampling (ICLR 2025 · similarity 0.85)
- Once Read is Enough: Domain-specific Pretraining-free Language Models with Cluster-guided Sparse Experts for Long-tail Domain Knowledge (NeurIPS 2024 · similarity 0.84)
- BLADE: Enhancing Black-Box Large Language Models with Small Domain-Specific Models (AAAI 2025 · similarity 0.83)
- Narrowing the Gap between Supervised and Unsupervised Sentence Representation Learning with Large Language Model (AAAI 2024 · similarity 0.83)
- DataMan: Data Manager for Pre-training Large Language Models (ICLR 2025 · similarity 0.83)
- Improving the Robustness of Knowledge-Grounded Dialogue via Contrastive Learning (AAAI 2024 · 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.
