Research map / Large language models / Language, safety and interpretability
Tracing Text Provenance via Context-Aware Lexical Substitution
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 |
|---|---|---|---|
| rank BERT-generated lexical substitutes using inference-based semantic relatedness | BERT lexical substitution · SWORDS | F^10: 31.6 -> 36.7 (Lenient); 16.8 -> 18.3 (Strict) | Helped |
| rank BERT-generated lexical substitutes using inference-based semantic relatedness | BERT lexical substitution · SWORDS | F_c^10: 53.3 -> 56.1 (Lenient); 26.1 -> 28.7 (Strict) | Helped |
| use context-aware lexical substitution with synchronicity and substitutability tests | Text watermarking · Watermarking datasets | SR: Proposed SR: 0.9844, 0.9852, 0.9854, 0.9864, 0.9850, 0.9763; baseline results vary by dataset | Helped |
| use context-aware lexical substitution with synchronicity and substitutability tests | Text watermarking · Human evaluation | meaning-preservation rating: 2.8\pm1.3, 2.4\pm1.0, 2.0\pm1.2 -> 4.5 \pm0.6 | Helped |
| remove the synchronicity test (ablation) | Sequence incremental watermarking · Watermark extraction | successful watermark extraction | Hurt |
| remove the substitutability test (ablation) | Sequence incremental watermarking · Watermark extraction | correct word location | Hurt |
| increase minimum substitution distance f from 1 to 2 (ablation) | Sequence incremental watermarking · Watermarking | payload (bpw): 0.091 -> 0.044 | Hurt |
| increase minimum substitution distance f from 1 to 3 (ablation) | Sequence incremental watermarking · Watermarking | payload (bpw): 0.091 -> 0.031 | Hurt |
| increase minimum substitution distance f from 1 to 3 (ablation) | Sequence incremental watermarking · Watermarking | SR: 0.983 -> 0.985 | No clear effect |
| increase minimum substitution distance f from 1 to 3 (ablation) | Sequence incremental watermarking · Watermarking | SS: 0.988 -> 0.995 | No clear effect |
Closest papers by meaning
- Adaptive Text Watermark for Large Language Models (ICML 2024 · similarity 0.87)
- On the Reliability of Watermarks for Large Language Models (ICLR 2024 · similarity 0.86)
- StealthInk: A Multi-bit and Stealthy Watermark for Large Language Models (ICML 2025 · similarity 0.86)
- On the Learnability of Watermarks for Language Models (ICLR 2024 · similarity 0.86)
- Learning to Watermark: A Selective Watermarking Framework for Large Language Models via Multi-Objective Optimization (NeurIPS 2025 · similarity 0.86)
- Can Watermarking Large Language Models Prevent Copyrighted Text Generation and Hide Training Data? (AAAI 2025 · 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.