Research map / Large language models / Language, safety and interpretability
e-SNLI: Natural Language Inference with Natural Language Explanations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 explanation-decoder training to InferSent | InferSent · SNLI test | accuracy: 84.01% (0.25) -> 83.96% (0.26) | No clear effect |
| Train sentence representations with human explanation supervision | InferSent · SentEval downstream tasks | task performance: Significantly outperforms on 4 tasks; significantly outperformed on 1 task | Mixed |
| Add a language generator that reconstructs premise and hypothesis instead of generating explanations (ablation) | InferSent · SentEval downstream tasks | task performance: Significantly worse on 6 tasks; significantly better on 2 tasks | Hurt |
| Generate an explanation first and predict the label from it | e-InferSent · SNLI test | accuracy: 83.96% (0.26) -> 81.59% (0.45) | Hurt |
| Add separate attention modules over premise and hypothesis tokens (ablation) | ExplainThenPredictSeq2Seq · SNLI test | correct explanations: 49.8% -> 64.27% | Helped |
| Add separate attention modules over premise and hypothesis tokens (ablation) | ExplainThenPredictSeq2Seq · SNLI test | accuracy: 81.59% (0.45) -> 81.71% (0.36) | No clear effect |
| Generate explanations from the hypothesis alone | PremiseAgnostic · SNLI test, first 100 examples | correct explanations: 6.83 correct explanations versus 66 correct labels | Hurt |
| Train an explanation-only label classifier | Explanation-only label model · SNLI test | accuracy: 96.83% | Helped |
| Train with explanation supervision and evaluate without fine-tuning | InferSent · SICK-E | accuracy: 53.27 (1.65) -> 53.54 (1.43) | No clear effect |
| Train with explanation supervision and evaluate without fine-tuning | InferSent · MultiNLI | accuracy: 57 (0.41) -> 57.16 (0.51) | No clear effect |
Closest papers by meaning
- Supervising Model Attention with Human Explanations for Robust Natural Language Inference (AAAI 2022 · similarity 0.89)
- Probing Natural Language Inference Models through Semantic Fragments (AAAI 2020 · similarity 0.89)
- Teach Me to Explain: A Review of Datasets for Explainable Natural Language Processing (NeurIPS 2021 · similarity 0.89)
- SeqGPT: An Out-of-the-Box Large Language Model for Open Domain Sequence Understanding (AAAI 2024 · similarity 0.86)
- Exploring the Limits of Sub-Billion Language Model Reasoners with Open Training Recipes (ICLR 2026 · 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.