# WT5?! Training Text-to-Text Models to Explain their Predictions

**Authors:** Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, Karishma Malkan  
**arXiv:** [2004.14546](https://arxiv.org/abs/2004.14546)  
**Date:** 2020-04-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2004.14546v1/

## Key points

- Neural networks have made significant advancements in natural language processing (NLP) tasks, achieving human-level performance.
- Understanding why these networks make specific predictions remains a challenge.
- The paper titled "WT5?! Training Text-to-Text Models to Explain their Predictions" proposes a novel approach to address this issue.
- The authors leverage the text-to-text framework introduced by Raffel et al. in 2019 to train language models that generate predictions and provide natural text explanations alongside them.
- This method does not require modifications to the loss function or training and decoding procedures.
- The model is trained to output an explanation after generating the prediction.
- This approach achieves state-of-the-art performance on explainability benchmarks.
- It enables learning from a limited set of labeled explanations and facilitates the transfer of rationalization abilities across different datasets.
- The authors have made their code available for training these models to promote reproducibility and further research in this area.
- Overall, this paper presents an innovative solution to enhance the interpretability of neural network predictions in NLP tasks.

## Summary

In recent years, neural networks have made significant advancements in natural language processing (NLP) tasks, achieving human-level performance. However, understanding why these networks make specific predictions remains a challenge. In this paper titled "WT5?! Training Text-to-Text Models to Explain their Predictions," authors Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel and Karishma Malkan propose a novel approach to address this issue. The authors leverage the text-to-text framework introduced by Raffel et al. in 2019 to train language models that not only generate predictions but also provide natural text explanations alongside them. Importantly, this method does not require any modifications to the loss function or training and decoding procedures. Instead, the model is trained to output an explanation after generating the prediction. The results of their study demonstrate that this approach achieves state-of-the-art performance on explainability benchmarks. Moreover, it enables learning from a limited set of labeled explanations and facilitates the transfer of rationalization abilities across different datasets. To promote reproducibility and further research in this area, the authors have made their code available for training these models. Overall, this paper presents an innovative solution to enhance the interpretability of neural network predictions in NLP tasks by training models to provide natural text explanations alongside their outputs.