# PADA: A Prompt-based Autoregressive Approach for Adaptation to Unseen Domains

**Authors:** Eyal Ben-David, Nadav Oved, Roi Reichart  
**arXiv:** [2102.12206](https://arxiv.org/abs/2102.12206)  
**Date:** 2021-02-24  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2102.12206v1/

## Key points

- Natural Language Processing (NLP) algorithms have improved in recent years
- NLP algorithms still struggle with out-of-distribution examples
- The paper addresses a challenging and previously underexplored version of the domain adaptation problem
- PADA is presented as a solution, which is a Prompt-based Autoregressive Domain Adaptation algorithm based on the T5 model
- The prompt acts as a unique signature that maps the test example to the semantic space spanned by the source domains
- Experiments were conducted with two tasks: Rumour Detection and Multi-Genre Natural Language Inference (MNLI)
- Results show that PADA strongly outperforms state-of-the-art approaches and additional strong baselines in both Rumour Detection and MNLI tasks
- PADA is effective in adapting to unseen domains with no prior knowledge about them.

## Summary

Natural Language Processing (NLP) algorithms have made significant progress in recent years. However, they still struggle when applied to out-of-distribution examples. This paper addresses a challenging and previously underexplored version of the domain adaptation problem. The setup involves training an algorithm on several source domains and then applying it to examples from an unseen domain that is unknown at training time. To tackle this problem, the authors present PADA: A Prompt-based Autoregressive Domain Adaptation algorithm based on the T5 model. The prompt acts as a unique signature that maps the test example to the semantic space spanned by the source domains. Experiments are conducted with two tasks: Rumour Detection and Multi-Genre Natural Language Inference (MNLI). Results show that PADA strongly outperforms state-of-the-art approaches and additional strong baselines in both Rumour Detection and MNLI tasks. It is effective in adapting to unseen domains with no prior knowledge about them.