Research map / Large language models / Language, safety and interpretability
Analyzing Compositionality-Sensitivity of NLI Models
Yixin Nie, Yicheng Wang, Mohit Bansal
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 compositionality-sensitivity examples selected by a confidently wrong bag-of-words model | NLI evaluation · SNLI and MNLI CS0.7 | accuracy | Mixed |
| Use syntactic TreeLSTM instead of sequential BiLSTM | ESIM / S-TLSTM · SNLI CS0.7 | accuracy: 55.28 -> 57.51 | Improved |
| Use syntactic TreeLSTM instead of sequential BiLSTM | ESIM / S-TLSTM · MNLI Matched CS0.7 | accuracy: 48.31 -> 48.90 | Improved |
| Use syntactic TreeLSTM instead of sequential BiLSTM | ESIM / S-TLSTM · MNLI Mismatched CS0.7 | accuracy: 48.32 -> 50.61 | Improved |
| Replace RNN layers with fully-connected layers (ablation) | RSE · SNLI CS0.7 | accuracy: 52.73 -> 43.60 | Hurt |
| Replace RNN layers with fully-connected layers (ablation) | ESIM · SNLI CS0.7 | accuracy: 55.28 -> 40.49 | Hurt |
| Replace RNN layers with fully-connected layers (ablation) | DR-BiLSTM · SNLI CS0.7 | accuracy: 55.28 -> 41.38 | Hurt |
| Train with shuffled input words (ablation) | ESIM · SNLI CS0.7 | accuracy: 55.28 -> 45.49 | Hurt |
| Train with shuffled input words (ablation) | DR-BiLSTM · SNLI CS0.7 | accuracy: 55.28 -> 44.77 | Hurt |
| Train with subject-object-swap adversarial examples (ablation) | ESIM · subject-object-swap and adjective-modifier-swap adversaries | correct-label prediction rate: 2.1 -> 99.1; 9.6 -> 26.9 | Mixed |
| Train with adjective-modifier-swap adversarial examples (ablation) | ESIM · adjective-modifier-swap and subject-object-swap adversaries | correct-label prediction rate: 4.8 -> 99.6; 2.1 -> 1.0 | Mixed |
Closest papers by meaning
- Measuring Compositionality in Representation Learning (ICLR 2019 · similarity 0.87)
- Compositionality with Variation Reliably Emerges in Neural Networks (ICLR 2023 · similarity 0.87)
- Natural Language Inference Improves Compositionality in Vision-Language Models (ICLR 2025 · similarity 0.86)
- SUGARCREPE++ Dataset: Vision-Language Model Sensitivity to Semantic and Lexical Alterations (NeurIPS 2024 · similarity 0.85)
- ConMe: Rethinking Evaluation of Compositional Reasoning for Modern VLMs (NeurIPS 2024 · similarity 0.85)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.85)
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.
