# Recipes for building an open-domain chatbot

**Authors:** Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston  
**arXiv:** [2004.13637](https://arxiv.org/abs/2004.13637)  
**Date:** 2020-04-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2004.13637v1/

## Key points

- Challenges and advancements in developing chatbots using machine learning techniques
- Importance of factors beyond scaling neural models, such as engaging talking points, active listening, displaying knowledge, empathy, and appropriate personality expression
- Introduction of large-scale models named Blender with 90M, 2.7B, and 9.4B parameter variants
- Training strategies to learn essential conversational skills
- Human evaluations showing superior performance of Blender models in terms of engagingness and humanness during multi-turn dialogues
- Potential for further refinement to enhance performance and reliability of open-domain chatbots

## Summary

The paper titled "Recipes for building an open-domain chatbot" discusses the challenges and advancements in developing chatbots using machine learning techniques. The authors highlight that while previous research has focused on scaling neural models by increasing the number of parameters and training data size, there are other crucial factors for creating a high-performing chatbot. They emphasize that successful conversation requires various skills such as engaging talking points, active listening, displaying knowledge, empathy, and appropriate personality expression while maintaining a consistent persona. To demonstrate their findings, the authors present large-scale models with 90M, 2.7B, and 9.4B parameter variants named Blender. These models are trained using suitable data and generation strategies to learn these essential conversational skills. Human evaluations indicate that the best Blender models outperform existing approaches in terms of engagingness and humanness measurements during multi-turn dialogues. By addressing potential limitations identified through failure case analysis, future work can focus on refining these models even further to enhance the performance and reliability of open-domain chatbots. In summary, this paper provides valuable insights into building effective open-domain chatbots by considering not only model scalability but also incorporating important conversational skills. The Blender models developed through this research exhibit superior performance compared to existing approaches in terms of engagingness and humanness during multi-turn conversations.