Research map / Large language models / Language, safety and interpretability
Rare Words: A Major Problem for Contextualized Embeddings and How to Fix it by Attentive Mimicking
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 Attentive Mimicking to BERT | BERT-base and BERT-large · WNLaMPro rare antonyms, hypernyms, cohyponyms+, and corruptions | MRR: BERT-base: 0.149 -> 0.449 (ant), 0.276 -> 0.300 (hyp), 0.147 -> 0.213 (coh+), 0.020 -> 0.254 (cor); BERT-large: 0.234 -> | Helped |
| Add Attentive Mimicking to BERT-base | BERT-base · WNLaMPro medium antonyms, hypernyms, and cohyponyms+ | MRR: ant 0.089 -> 0.511; hyp 0.327 -> 0.343; coh+ 0.177 -> 0.213 | Helped |
| Add Attentive Mimicking to BERT-large | BERT-large · WNLaMPro medium-frequency words | MRR: ant 0.218 -> 0.558; hyp 0.350 -> 0.323; coh+ 0.210 -> 0.216 | Mixed |
| Add Attentive Mimicking to BERT | BERT-base and BERT-large · WNLaMPro frequent hypernyms and cohyponyms+ | MRR: BERT-base hyp 0.416 -> 0.377, coh+ 0.294 -> 0.262; BERT-large hyp 0.462 -> 0.401, coh+ 0.337 -> 0.292 | Hurt |
| Add AM+context configuration instead of default AM (ablation) | AM on BERT-base · WNLaMPro development set | MRR: 5 epochs: 0.258 -> 0.262; 10 epochs: 0.253 -> 0.276 | Helped |
| Train AM for 10 rather than 5 epochs (ablation) | AM+context on BERT-base · WNLaMPro development set | MRR: 0.262 -> 0.276 | Helped |
| Remove OTA from AM training (ablation) | AM on BERT-base · WNLaMPro development set | MRR: 0.258 -> 0.219 (5 epochs); 0.253 -> 0.220 (10 epochs) | Hurt |
| Remove the form component from AM (ablation) | AM on BERT-base · WNLaMPro development set | MRR: 0.258 -> 0.138 (5 epochs); 0.253 -> 0.133 (10 epochs) | Hurt |
| Remove the context component from AM (ablation) | AM on BERT-base · WNLaMPro development set | MRR: 0.258 -> 0.227 (5 epochs); 0.253 -> 0.225 (10 epochs) | Hurt |
| Use random rather than static contexts for OTA (ablation) | OTA · 1,000 one-token BERT-vocabulary words | cosine distance | Helped |
Closest papers by meaning
- Learning Semantic Representations for Novel Words: Leveraging Both Form and Context (AAAI 2019 · similarity 0.88)
- SimReg: Achieving Higher Performance in the Pretraining via Embedding Similarity Regularization (NeurIPS 2026 · similarity 0.86)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.85)
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.85)
- Model-Aware Tokenizer Transfer (NeurIPS 2026 · similarity 0.85)
- Zero-Shot Tokenizer Transfer (NeurIPS 2024 · 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.