# e-SNLI: Natural Language Inference with Natural Language Explanations

**Authors:** Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, Phil Blunsom  
**arXiv:** [1812.01193](https://arxiv.org/abs/1812.01193)  
**Date:** 2018-12-04  
**Summary page:** https://www.summarizepaper.com/en/paper/1812.01193v2/

## Key points

- The paper addresses the need for machine learning models to provide interpretable and robust explanations for their decisions
- The authors extend the SNLI dataset by adding human-annotated natural language explanations of entailment relations
- They develop models that incorporate these explanations during training and output them during testing
- The authors introduce a new corpus called e-SNLI, which can be used for obtaining full sentence justifications, improving universal sentence representations, and transferring knowledge to out-of-domain NLI datasets
- By incorporating natural language explanations into the training process and making them available at test time, the authors demonstrate how their dataset can enhance models and establish trust in decision-making
- This work expands on the importance of interpretability in machine learning models and provides a valuable resource for researchers interested in using natural language explanations to improve model performance and reliability.

## Summary

The paper titled "e-SNLI: Natural Language Inference with Natural Language Explanations" by Oana-Maria Camburu, Tim Rocktäschel, Thomas Lukasiewicz, and Phil Blunsom addresses the need for machine learning models to provide interpretable and robust explanations for their decisions in order to gain widespread public adoption. The authors extend the Stanford Natural Language Inference (SNLI) dataset by adding a layer of human-annotated natural language explanations of the entailment relations. They also develop models that incorporate these explanations during training and output them during testing. The authors introduce a new corpus called e-SNLI, which consists of explanations that can be used for various purposes. These include obtaining full sentence justifications for a model's decisions, improving universal sentence representations, and transferring knowledge to out-of-domain NLI datasets. By incorporating natural language explanations into the training process and making them available at test time, the authors demonstrate how their dataset can be used to enhance models and establish trust in their decision-making. Overall, this work expands on the importance of interpretability in machine learning models and provides a valuable resource for researchers interested in using natural language explanations to improve model performance and reliability.