# Learning From Free-Text Human Feedback -- Collect New Datasets Or Extend Existing Ones?

**Authors:** Dominic Petrak, Nafise Sadat Moosavi, Ye Tian, Nikolai Rozanov, Iryna Gurevych  
**arXiv:** [2310.15758](https://arxiv.org/abs/2310.15758)  
**Date:** 2023-10-24  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.15758v1/

## Key points

- Researchers Dominic Petrak, Nafise Sadat Moosavi, Ye Tian, Nikolai Rozanov, and Iryna Gurevych focus on learning from free-text human feedback for dialog systems.
- They address the scarcity of annotated data in conversational AI by using synthetic dialog generation to augment existing datasets with necessary annotations.
- The authors investigate dialog datasets such as MultiWoZ, SGD, BABI, PersonaChat, Wizards-of-Wikipedia, and the human-bot split of the Self-Feeding Chatbot to assess their approach's feasibility.
- Through observations and analysis of free-text human feedback in dialogs, they develop new taxonomies for annotating this type of data and examine its impact on response generation for language generation models like GPT-2, LLAMA, and Flan-T5.
- The research identifies error types, user response types, and their relationships within these datasets.
- Accepted for presentation at EMNLP 2023, this research has the potential to enhance conversational AI technology significantly by improving existing datasets with free-text human feedback annotations.

## Summary

Dominic Petrak, Nafise Sadat Moosavi, Ye Tian, Nikolai Rozanov, and Iryna Gurevych explore the importance of learning from free-text human feedback for dialog systems. Their research focuses on addressing the scarcity of annotated data in conversational AI and proposes a solution using synthetic dialog generation to augment existing datasets with necessary annotations. The authors investigate commonly used dialog datasets such as MultiWoZ, SGD, BABI, PersonaChat, Wizards-of-Wikipedia, and the human-bot split of the Self-Feeding Chatbot to assess the feasibility of their approach. Through their observations and analysis of free-text human feedback in dialogs, they develop new taxonomies for annotating this type of data and examine its impact on response generation for state-of-the-art language generation models like GPT-2, LLAMA, and Flan-T5. This work provides valuable insights into the composition of these datasets by identifying error types, user response types, and their relationships. Accepted for presentation at EMNLP 2023, this research has the potential to significantly contribute to advancements in conversational AI technology through enhancing existing datasets with free-text human feedback annotations.