AI Paper Atlas

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

ICML 2019 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations · 71 citations (OpenAlex)

Read the paper

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 changedWhereMeasuredEffect
increase L2 penalty alpha from 5e-5 to 1e-4 (ablation)Transformer encoder · SST-2 saliencywords encoded in last hidden layerHurt
decrease L2 penalty alpha to 5 x 10-5 from 1 x 10-4 (ablation)Transformer encoder · SST-2 saliencywords encoded in last hidden layerImproved
use pre-trained BERT instead of Transformertext classifier · SST-2accuracy: 0.8245 -> 0.9323Improved
use pre-trained BERT instead of Transformertext classifier · CoLAMCC: 0.1560 -> 0.6110Improved
use pre-trained BERT instead of LSTMtext classifier · QQPaccuracy: 0.8658 -> 0.9129Improved
use pre-trained BERT instead of CNNtext classifier · CoLAMCC: 0.0985 -> 0.6110Improved
use LSTM instead of Transformertext classifier · QQPaccuracy: 0.7637 -> 0.8658Improved
use LSTM instead of CNNtext classifier · SST-2accuracy: 0.8200 -> 0.8486Improved
compare Transformer with CNN, similar accuracytext classifier · SST-2accuracy: 0.8245 vs 0.8200No clear effect
explain with information-based measure instead of gradient or LRPsaliency method · SST-2 layer analysiscoherency across layersImproved
explain with information-based measure instead of perturbationsaliency method · SST-2 layer analysisnoise of saliency mapImproved
pre-training (BERT) versus nonemodels · SST-2stopword discriminationImproved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy