# CheerBots: Chatbots toward Empathy and Emotionusing Reinforcement Learning

**Authors:** Jiun-Hao Jhan, Chao-Peng Liu, Shyh-Kang Jeng, Hung-Yi Lee  
**arXiv:** [2110.03949](https://arxiv.org/abs/2110.03949)  
**Date:** 2021-10-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.03949v1/

## Key points

- The study focuses on the development of empathetic chatbots called CheerBots
- CheerBots prioritize understanding and responding to users' emotions
- The framework involves multiple dialogue turns, training the chatbots to understand implied feelings and respond empathetically
- Chatbots can be retrieval-based or generative-based and have been fine-tuned using deep reinforcement learning
- A simulating agent called the Conceptual Human Model aids in training by considering changes in the user's emotional states over time
- CheerBots outperform other baseline chatbots and achieve reciprocal altruism, as evidenced by automatic metrics and human rating results
- RL-based models generate sentences more similar to ground truth sentences collected in an empathetic manner, indicating improved coherence in conversation responses
- Deep reinforcement learning helps speakers transition to positive emotional states, with higher rewards observed in conversations with three turns compared to one turn
- Human rating results support the effectiveness of CheerBots in generating empathetic responses
- The study provides a comprehensive framework for developing empathetic chatbots and highlights their potential for enhancing user experiences through reciprocal altruism

## Summary

The study focuses on the development of empathetic chatbots called CheerBots, which prioritize understanding and responding to users' emotions. These chatbots aim to provide a more interactive and supportive experience by incorporating altruistic behaviors in human interactions. The framework presented in the study involves multiple dialogue turns, where CheerBots are trained to understand users' implied feelings and respond empathetically. The chatbots can be retrieval-based or generative-based and have been fine-tuned using deep reinforcement learning. To ensure empathetic responses, the researchers developed a simulating agent called the Conceptual Human Model. This model aids CheerBots in training by considering changes in the user's emotional states over time, with the goal of arousing sympathy. The study demonstrates that CheerBots outperform other baseline chatbots and achieve reciprocal altruism, as evidenced by automatic metrics and human rating results. In addition to the summary provided, further details from the study include an evaluation of different models based on empathy coherence, fluency, and empathy scores. The results show that RL-based models generate sentences more similar to ground truth sentences collected in an empathetic manner. This indicates improved coherence in conversation responses. The performance of deep reinforcement learning (RL) is also evaluated using rewards computed from emotional states between conversation turns. The results demonstrate that our models help speakers transition to positive emotional states, with higher rewards observed in conversations with three turns compared to one turn. Human rating results further support the effectiveness of CheerBots, indicating their ability to generate empathetic responses. Overall, this study provides a comprehensive framework for developing empathetic chatbots and highlights their potential for enhancing user experiences through reciprocal altruism. The code and pre-trained models used in this research will be made available for further exploration and implementation.