Research map / Large language models / Language, safety and interpretability
Towards Semantics-Enhanced Pre-Training: Can Lexicon Definitions Help Learning Sentence Meanings?
Xuancheng Ren, Xu Sun, Houfeng Wang, Qun Liu
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 |
|---|---|---|---|
| Add SemPre word-definition and relational pre-training | RoBERTa-large · WG1 and WG2 | lowest rank: 332 -> 86; 262 -> 130 | Improved |
| Add SemPre word-definition and relational pre-training | RoBERTa-large · CA, WSC, SM, SMR, ARCT1, and ARCT2 | accuracy: 96.2 -> 96.7; 69.3 -> 73.5; 79.2 -> 80.4; 47.3 -> 48.4; 54.3 -> 58.3; 60.0 -> 60.9 | Improved |
| Add SemPre word-definition and relational pre-training | RoBERTa-large · SWAG and HellaSwag | accuracy: 76.1 -> 75.9; 48.9 -> 48.5 | Hurt |
| Add SemPre word-definition and relational pre-training | RoBERTa-large · WiC | accuracy: 74.6 -> 75.7; 70.3 -> 72.1 | Improved |
| Add SemPre word-definition and relational pre-training | RoBERTa-large · PIQA | accuracy: 81.3 -> 81.6 | Improved |
| Add word-definition training (ablation) | RoBERTa-large semantics-focused pre-training · CATs average and WiC validation | accuracy: 66.4 -> 67.6; 74.6 -> 74.8 | Improved |
| Use word pairs alone, without definitions (ablation) | RoBERTa-large semantics-focused pre-training · CATs average and WiC validation | accuracy: 66.4 -> 64.3; 74.6 -> 74.3 | Hurt |
| Remove the word-masking objective (ablation) | RoBERTa-large + SemPre · CATs average and WiC validation | accuracy: 67.8 -> 67.1; 75.7 -> 75.1 | Hurt |
| Remove the definition-masking objective (ablation) | RoBERTa-large + SemPre · CATs average and WiC validation | accuracy: 67.8 -> 67.4; 75.7 -> 74.9 | Hurt |
| Remove the relation-prediction objective (ablation) | RoBERTa-large + SemPre · CATs average and WiC validation | accuracy: 67.8 -> 67.7; 75.7 -> 74.6 | Hurt |
Closest papers by meaning
- The Inductive Bias of In-Context Learning: Rethinking Pretraining Example Design (ICLR 2022 · similarity 0.88)
- On the Transferability of Pre-trained Language Models: A Study from Artificial Datasets (AAAI 2022 · similarity 0.88)
- Pre-training Text-to-Text Transformers for Concept-centric Common Sense (ICLR 2021 · similarity 0.87)
- Exploring Visual Pretraining for Learning Language Intelligence (CVPR 2026 · similarity 0.85)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.85)
- ImpScore: A Learnable Metric For Quantifying The Implicitness Level of Sentences (ICLR 2025 · 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.
