Research map / Architectures and efficiency / Attention and transformers
Lexical Simplification with Pretrained Encoders
Jipeng Qiang, Yun Li, Yi Zhu, Yunhao Yuan, Xindong Wu
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 |
|---|---|---|---|
| Generate candidates using masked BERT with the original sentence as context, rather than word-embedding similarity | BERT-LS candidate generation · LexMTurk | F1: 0.195 -> 0.259 | Improved |
| Generate candidates using masked BERT with the original sentence as context, rather than word-embedding similarity | BERT-LS candidate generation · BenchLS | F1: 0.236 -> 0.272 | Improved |
| Generate candidates using masked BERT with the original sentence as context, rather than word-embedding similarity | BERT-LS candidate generation · NNSeval | F1: 0.157 -> 0.218 | Improved |
| Use BERT-LS instead of the strongest full-pipeline baseline | Full lexical simplification pipeline · LexMTurk | Accuracy: 0.682 -> 0.776 | Improved |
| Use BERT-LS instead of the strongest full-pipeline baseline | Full lexical simplification pipeline · BenchLS | Accuracy: 0.434 -> 0.607 | Improved |
| Use BERT-LS instead of the strongest full-pipeline baseline | Full lexical simplification pipeline · NNSeval | Accuracy: 0.335 -> 0.423 | Improved |
| Use BERT-LS rather than the strongest full-pipeline baseline | Full lexical simplification pipeline · BenchLS and NNSeval | Precision: 0.642 -> 0.607; 0.544 -> 0.423 | Hurt |
| Increase the number of candidate substitutions (ablation) | BERT-LS candidate generation · LexMTurk, BenchLS, and NNSeval | Precision | Hurt |
| Increase the number of candidate substitutions (ablation) | BERT-LS candidate generation · LexMTurk, BenchLS, and NNSeval | Recall | Improved |
| Vary the number of candidate substitutions from 1 to 15 (ablation) | BERT-LS candidate generation · LexMTurk, BenchLS, and NNSeval | F1 | Mixed |
| Vary the number of candidate substitutions from 1 to 15 (ablation) | BERT-LS full pipeline · LexMTurk, BenchLS, and NNSeval | Accuracy | Mixed |
Closest papers by meaning
- From Tokens to Words: On the Inner Lexicon of LLMs (ICLR 2025 · similarity 0.86)
- Exploiting Vocabulary Frequency Imbalance in Language Model Pre-training (NeurIPS 2025 · similarity 0.85)
- Hierarchical Autoregressive Transformers: Combining Byte- and Word-Level Processing for Robust, Adaptable Language Models (ICLR 2025 · similarity 0.85)
- Over-Tokenized Transformer: Vocabulary is Generally Worth Scaling (ICML 2025 · similarity 0.85)
- An Analysis of Tokenization: Transformers under Markov Data (NeurIPS 2024 · similarity 0.84)
- Explaining and Mitigating Crosslingual Tokenizer Inequities (NeurIPS 2025 · similarity 0.84)
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.
