Research map / Large language models / Language, safety and interpretability
What if We Simply Swap the Two Text Fragments? A Straightforward yet Effective Way to Test the Robustness of Methods to Confounding Signals in Nature Language Inference Tasks
Haohan Wang, Da Sun, Eric P. Xing
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 |
|---|---|---|---|
| Swap premise and hypothesis at test time | CBOW, InferSent, DGA, ESIM, KIM, and ADV · NLI Test, entailment pairs | accuracy: 0.856 -> 0.080 (CBOW); 0.880 -> 0.087 (InferSent); 0.854 -> 0.422 (DGA); 0.884 -> 0.324 (ESIM); 0.895 -> 0.095 (KIM); 0. | Hurt |
| Swap premise and hypothesis at test time | CBOW, InferSent, DGA, ESIM, KIM, and ADV · NLI Test, contradiction pairs | accuracy: 0.740 -> 0.580 (CBOW); 0.859 -> 0.682 (InferSent); 0.711 -> 0.650 (DGA); 0.861 -> 0.701 (ESIM); 0.845 -> 0.796 (KIM); 0. | Hurt |
| Swap premise and hypothesis at test time | CBOW, InferSent, DGA, ESIM, KIM, and ADV · NLI Test, neutral pairs | accuracy: 0.659 -> 0.589 (CBOW); 0.795 -> 0.712 (InferSent); 0.700 -> 0.619 (DGA); 0.801 -> 0.720 (ESIM); 0.781 -> 0.675 (KIM); 0. | Hurt |
| Train with some neutral and contradiction pairs swapped (ablation) | InferSent · Stress Test Antonymy | accuracy: 22.87 -> 27.61 (matched); 16.92 -> 21.57 (mismatched) | Improved |
| Train with some neutral and contradiction pairs swapped (ablation) | ESIM · Stress Test Word Overlap | accuracy: 57.41 -> 37.73 (matched); 58.27 -> 37.52 (mismatched) | Hurt |
| Train with some neutral and contradiction pairs swapped (ablation) | ESIM · Stress Test Spelling Error | accuracy: 57.09 -> 37.40 (matched); 56.48 -> 37.54 (mismatched) | Hurt |
| Train with some neutral and contradiction pairs swapped (ablation) | ADV · Stress Test Antonymy | accuracy: 33.63 -> 20.69 (matched); 30.22 -> 20.88 (mismatched) | Hurt |
| Train with some neutral and contradiction pairs swapped (ablation) | DGA · Stress Test | deviation: Devi 0.073 for Antonymy; 0.009 for Length Mismatch; 0.025 for Negation; 0.005 for Word Overlap; 0.011 for Spelling Error | Improved |
| Train InferSent with swapped sentence pairs (ablation) | InferSent sentence embeddings · NLP transfer tasks | task metric: MR 76.92 -> 77.3; CR 78.15 -> 79.84; MPQA 87.64 -> 87.72; SUBJ 90.79 -> 90.99; TREC 83.6 -> 85; SICK-E 82.3 -> 83.24; SI | Mixed |
Closest papers by meaning
- Probing Natural Language Inference Models through Semantic Fragments (AAAI 2020 · similarity 0.90)
- Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals (ICLR 2024 · similarity 0.86)
- MBR and QE Finetuning: Training-time Distillation of the Best and Most Expensive Decoding Methods (ICLR 2024 · similarity 0.85)
- Auto-GDA: Automatic Domain Adaptation for Efficient Grounding Verification in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.84)
- SeqGPT: An Out-of-the-Box Large Language Model for Open Domain Sequence Understanding (AAAI 2024 · similarity 0.84)
- Generative Judge for Evaluating Alignment (ICLR 2024 · similarity 0.84)
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.
