# Investigation of Sentiment Controllable Chatbot

**Authors:** Hung-yi Lee, Cheng-Hao Ho, Chien-Fu Lin, Chiung-Chih Chang, Chih-Wei Lee, Yau-Shian Wang, Tsung-Yuan Hsu, Kuan-Yu Chen  
**arXiv:** [2007.07196](https://arxiv.org/abs/2007.07196)  
**Date:** 2020-07-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2007.07196v1/

## Key points

- The paper explores the limitations of conventional seq2seq chatbot models
- Four different models are investigated: persona-based model, reinforcement learning, plug and play model, and CycleGAN
- These models aim to scale or adjust the sentiment of chatbot responses
- Machine-evaluated metrics are introduced to assess the reasonableness of chatbot responses
- Reinforcement learning and CycleGAN show promise in controlling sentiment in chatbot responses
- Potential for improving chatbot interactions by incorporating sentiment considerations
- Research contributes to advancing sentiment controllability in chatbots through different models and novel metrics

## Summary

The paper titled "Investigation of Sentiment Controllable Chatbot" explores the limitations of conventional seq2seq chatbot models, which focus solely on finding sentences with the highest probabilities based on input sequences without considering the sentiment of the output sentences. To address this limitation, the authors investigate four different models: a persona-based model, reinforcement learning, a plug and play model, and CycleGAN. These models are all built upon the seq2seq model and aim to scale or adjust the sentiment of chatbot responses. In addition to developing these models, the authors also introduce machine-evaluated metrics to assess the reasonableness of chatbot responses given specific inputs. These metrics, along with human evaluation, are used to analyze and compare the performance of the four models across various aspects. The findings reveal that both reinforcement learning and CycleGAN show significant promise in terms of their ability to control sentiment in chatbot responses. This suggests that these two approaches may be particularly attractive for improving chatbot interactions by incorporating sentiment considerations. Overall, this research contributes to advancing sentiment controllability in chatbots by exploring different models and evaluating their performance using novel metrics. The results highlight potential avenues for enhancing user experiences with chatbots through more nuanced and contextually appropriate responses.