Research map / Label-efficient and robust learning / Recognition and long-tail
Improving Language Modelling with Noise Contrastive Estimation
Farhana Ferdousi Liza, Marek Grzes
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 NCE instead of softmax for a large LSTM trained for 55 epochs | Large PTB LSTM · PTB test | perplexity: 78.826 -> 69.995 | Improved |
| Use NCE instead of softmax for a large LSTM trained for 20 epochs | Large PTB LSTM · PTB test | perplexity: 76.935 -> 74.129 | Improved |
| Use a long search period, τ=25, rather than τ=6 (ablation) | NCE LSTM · PTB test, medium model | perplexity: 83.858 -> 75.959 | Improved |
| Use a long search period, τ=25, rather than τ=6 (ablation) | Softmax LSTM · PTB test, medium model | perplexity: 78.567 -> 79.906 | Hurt |
| Add smaller-than-standard uniform weight initialization (ablation) | NCE LSTM · PTB test, large model | perplexity: 71.184 -> 70.444 -> 69.995 | Improved |
| Add empirically tuned smaller uniform weight initialization (ablation) | NCE LSTM · PTB test, medium model | perplexity: 75.960 -> 75.286 | Improved |
| Reduce the standard uniform initialization range by a factor of 4 (ablation) | NCE LSTM · PTB test, small model | perplexity: 104.449 -> 102.245 | Improved |
| Use empirically tuned smaller uniform initialization rather than the range reduced by 4 (ablation) | NCE LSTM · PTB test, small model | perplexity: 102.245 -> 102.237 | Improved |
| Increase dropout from 60% to 70% (ablation) | Softmax large LSTM · PTB validation | perplexity | Mixed |
| Increase the softmax search-period learning rate to raise early perplexity (ablation) | Softmax large LSTM · PTB validation | perplexity | No clear effect |
| Use 600 noise samples rather than other tested sample counts (ablation) | NCE LSTM · PTB | perplexity | Mixed |
| Use lower-variance initialization for NCE (ablation) | NCE LSTM · PTB test, large model | perplexity: 71.184 -> 69.995 | Improved |
Closest papers by meaning
- Denoising Pre-training for Machine Translation Quality Estimation with Curriculum Learning (AAAI 2023 · similarity 0.87)
- Conditional Noise-Contrastive Estimation of Unnormalised Models (ICML 2018 · similarity 0.86)
- Signal and Noise: A Framework for Reducing Uncertainty in Language Model Evaluation (NeurIPS 2025 · similarity 0.84)
- Noise Contrastive Alignment of Language Models with Explicit Rewards (NeurIPS 2024 · similarity 0.84)
- Enhancing Large Language Model Performance with Gradient-Based Parameter Selection (AAAI 2025 · similarity 0.83)
- Large Language Models are Efficient Learners of Noise-Robust Speech Recognition (ICLR 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.
