# Learning Improvised Chatbots from Adversarial Modifications of Natural Language Feedback

**Authors:** Makesh Narsimhan Sreedhar, Kun Ni, Siva Reddy  
**arXiv:** [2010.07261](https://arxiv.org/abs/2010.07261)  
**Date:** 2020-10-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2010.07261v1/

## Key points

- The paper discusses improving chatbots by utilizing data from user interactions
- The authors propose a self-feeding chatbot that solicits natural language feedback from users
- A generative adversarial model is introduced to convert noisy feedback into plausible responses
- The proposed model consists of a generator and a discriminator
- Performance improvement is demonstrated on the Personachat dataset, increasing from 69.94% to 75.96%
- Attention maps are analyzed for different modification types and statistics are provided based on random feedback texts
- Accuracy is assessed using FEED2RESP (the proposed model)
- The approach enhances chatbot performance by converting noisy user feedback into plausible responses
- Results demonstrate significant improvements in ranking correct responses on a large dataset
- The work contributes to advancing chatbot technology by leveraging user interactions to enhance capabilities.

## Summary

The paper discusses the potential for improving chatbots by utilizing the vast amount of data generated through their interactions with users. Specifically, the authors propose a self-feeding chatbot that continuously improves itself by soliciting natural language feedback from users when they are dissatisfied with its responses. To address the issue of extraneous sequences in user feedback hindering its usefulness as a training sample, the authors introduce a generative adversarial model that converts noisy feedback into plausible natural responses in a conversation. The proposed model consists of a generator and a discriminator. The generator's objective is to convert the feedback into an appropriate response while also fooling the discriminator, which is trained to distinguish between feedback and natural responses. By augmenting the original training data with these modified feedback responses, significant improvement in performance is demonstrated on the Personachat dataset - increasing from 69.94% to 75.96%. To evaluate their approach further, attention maps are analyzed for different modification types such as rewrite, remove and retain; statistics are provided based on 200 random feedback texts; and accuracy is assessed using FEED2RESP (the proposed model). In conclusion, this work presents a novel approach to enhancing chatbot performance by converting noisy user feedback into plausible natural responses through a generative adversarial model. The results demonstrate significant improvements in ranking correct responses on a large dataset and contribute to advancing chatbot technology by highlighting the potential for leveraging user interactions to continuously enhance their capabilities.