Research map / Large language models / Language, safety and interpretability
Towards a Deep and Unified Understanding of Deep Neural Models in NLP
Chaoyu Guan, Xiting Wang, Quanshi Zhang, Runjin Chen, Di He, Xing Xie
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 |
|---|---|---|---|
| increase L2 penalty alpha from 5e-5 to 1e-4 (ablation) | Transformer encoder · SST-2 saliency | words encoded in last hidden layer | Hurt |
| decrease L2 penalty alpha to 5 x 10-5 from 1 x 10-4 (ablation) | Transformer encoder · SST-2 saliency | words encoded in last hidden layer | Improved |
| use pre-trained BERT instead of Transformer | text classifier · SST-2 | accuracy: 0.8245 -> 0.9323 | Improved |
| use pre-trained BERT instead of Transformer | text classifier · CoLA | MCC: 0.1560 -> 0.6110 | Improved |
| use pre-trained BERT instead of LSTM | text classifier · QQP | accuracy: 0.8658 -> 0.9129 | Improved |
| use pre-trained BERT instead of CNN | text classifier · CoLA | MCC: 0.0985 -> 0.6110 | Improved |
| use LSTM instead of Transformer | text classifier · QQP | accuracy: 0.7637 -> 0.8658 | Improved |
| use LSTM instead of CNN | text classifier · SST-2 | accuracy: 0.8200 -> 0.8486 | Improved |
| compare Transformer with CNN, similar accuracy | text classifier · SST-2 | accuracy: 0.8245 vs 0.8200 | No clear effect |
| explain with information-based measure instead of gradient or LRP | saliency method · SST-2 layer analysis | coherency across layers | Improved |
| explain with information-based measure instead of perturbation | saliency method · SST-2 layer analysis | noise of saliency map | Improved |
| pre-training (BERT) versus none | models · SST-2 | stopword discrimination | Improved |
Closest papers by meaning
- Interpreting Deep Models for Text Analysis via Optimization and Regularization Methods (AAAI 2019 · similarity 0.89)
- Training Language Models to Explain Their Own Computations (NeurIPS 2026 · similarity 0.87)
- Exploring Activation Patterns of Parameters in Language Models (AAAI 2025 · similarity 0.86)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.86)
- Automated Natural Language Explanation of Deep Visual Neurons with Large Models (Student Abstract) (AAAI 2024 · similarity 0.86)
- Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals (ICLR 2024 · 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.
