Research map / Graph learning / Knowledge and temporal graphs
Improving Interpretability via Explicit Word Interaction Graph Layer
Arshdeep Sekhon, Hanjie Chen, Aman Shrivastava, Zhe Wang, Yangfeng Ji, Yanjun Qi
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 |
|---|---|---|---|
| Add WiGraph to LSTM | LSTM text classifiers · IMDB, SST-1, SST-2, AG News, TREC, Subj | accuracy: 88.39 -> 90.12; 43.84 -> 46.47; 83.74 -> 86.21; 91.03 -> 91.16; 90.40 -> 92.20; 90.20 -> 91.40 | Improved |
| Add WiGraph to BERT | BERT text classifiers · IMDB, SST-1, SST-2, AG News, TREC, Subj | accuracy: 91.88 -> 92.48; 51.63 -> 52.49; 92.15 -> 92.59; 92.05 -> 92.72; 97.40 -> 97.40; 96.40 -> 96.60 | Improved |
| Add WiGraph to a concept bottleneck model and apply test-time intervention | Inception V3 concept bottleneck · CUB | accuracy: 89.41%(+9.37%) -> 95.74%(+15.70%) | Improved |
| Use the learned top interactions instead of no pairwise interactions (ablation) | WiGraph · CUB not applicable; text datasets (IoS) | Interaction Occlusion Score | Improved |
| Use WiGraph-A instead of WiGraph-A-R (ablation) | WiGraph text classifiers · Prediction accuracy across 24 model-dataset combinations | accuracy: 89.12 vs 90.12; 44.32 vs 46.47; 84.08 vs 86.21; 91.16 vs 90.87; 91.65 vs 92.20; 90.60 vs 91.40 | Mixed |
| Use WiGraph-A instead of WiGraph-A-R (ablation) | distilBERT WiGraph · TREC and SST-2 | AOPC: 61.52 -> 68.74; 36.22 -> 44.33 | Improved |
| Use WiGraph-A instead of WiGraph-A-R (ablation) | distilBERT WiGraph · TREC and SST-2 | IoS: 96.20 -> 96.40; 86.90 -> 87.00 | Improved |
Closest papers by meaning
- Language Model Meets Prototypes: Towards Interpretable Text Classification Models through Prototypical Networks (AAAI 2025 · similarity 0.83)
- Towards Global-level Mechanistic Interpretability: A Perspective of Modular Circuits of Large Language Models (ICML 2025 · similarity 0.82)
- Analyze Feature Flow to Enhance Interpretation and Steering in Language Models (ICML 2025 · similarity 0.82)
- Neurons to Words: A Novel Method for Automated Neural Network Interpretability and Alignment (AAAI 2025 · similarity 0.82)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.81)
- Explaining Graph Neural Networks via Structure-aware Interaction Index (ICML 2024 · similarity 0.81)
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.
