# PLATO-2: Towards Building an Open-Domain Chatbot via Curriculum Learning

**Authors:** Siqi Bao, Huang He, Fan Wang, Hua Wu, Haifeng Wang, Wenquan Wu, Zhen Guo, Zhibin Liu, Xinchao Xu  
**arXiv:** [2006.16779](https://arxiv.org/abs/2006.16779)  
**Date:** 2020-06-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2006.16779v1/

## Key points

- PLATO-2 is a cutting-edge open-domain chatbot developed using curriculum learning.
- Curriculum learning involves two stages to ensure high-quality performance.
- In the first stage, a coarse-grained generation model is trained for one-to-one mapping of input queries to appropriate responses.
- The second stage involves training a fine-grained generation model and an evaluation model.
- The fine-grained generation model enhances the chatbot's ability to provide diverse and contextually appropriate responses.
- The evaluation model estimates response coherence to ensure logical and meaningful interactions.
- PLATO-2 was trained on both Chinese and English data for effectiveness across different languages.
- Comprehensive evaluations confirm PLATO-2's superiority in terms of response quality and diversity.
- Siqi Bao, Huang He, Fan Wang, Hua Wu, Haifeng Wang, Wenquan Wu, Zhen Guo, Zhibin Liu, and Xinchao Xu contributed significantly to building PLATO-2 as an advanced open domain chatbot.
- Curriculum learning is highlighted as important in developing high-quality conversational AI systems like PLATO 2.

## Summary

PLATO-2 is a cutting-edge open-domain chatbot that has been developed using an effective training process called curriculum learning. This process involves two stages to ensure the chatbot's high-quality performance. In the first stage, a coarse-grained generation model is trained to learn how to generate responses under a simplified framework of one-to-one mapping, where the model learns to map input queries to appropriate responses. This stage focuses on establishing a basic understanding of response generation. In the second stage, further refinement is achieved through training a fine-grained generation model and an evaluation model. The fine-grained generation model aims to enhance the chatbot's ability to provide diverse and contextually appropriate responses. It learns to generate responses that are not only accurate but also varied and engaging. Simultaneously, the evaluation model is trained to estimate response coherence which helps in ensuring that the generated responses are coherent and logical within the conversation context. By incorporating this evaluation component, PLATO-2 can provide more consistent and meaningful interactions with users. To ensure its effectiveness across different languages, PLATO-2 was trained on both Chinese and English data. Comprehensive evaluations have confirmed its superiority over existing chatbot models as it achieves new state-of-the-art results in terms of response quality and diversity. The authors of this research paper - Siqi Bao, Huang He, Fan Wang, Hua Wu, Haifeng Wang, Wenquan Wu, Zhen Guo, Zhibin Liu and Xinchao Xu - have made significant contributions towards building PLATO-2 as an advanced open domain chatbot. Their work highlights the importance of curriculum learning in developing high quality conversational AI systems like PLATO 2.