# Low-Resource Adaptation of Open-Domain Generative Chatbots

**Authors:** Greyson Gerhard-Young, Raviteja Anantha, Srinivas Chappidi, Björn Hoffmeister  
**arXiv:** [2108.06329](https://arxiv.org/abs/2108.06329)  
**Date:** 2021-08-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.06329v1/

## Key points

- Increasing the size of chatbot models improves performance
- Limitations in latency and connectivity when deploying digital assistants on devices
- Need to reduce the size of chatbot models for compatibility with user devices
- Low-parameter models can retain conversational abilities while improving in a specific domain
- Proposed generic framework addresses various question types, tracks references, and eliminates inconsistent and toxic responses
- Framework seamlessly transitions between casual chatting and performing transactional tasks
- Evaluation of framework using automatic (Perplexity) and human (SSA - Sensibleness and Specificity Average) metrics
- Achieved comparable performance while reducing model parameters by 90%

## Summary

In recent work on open-domain chatbots, it has been shown that increasing the size of the model improves its performance. However, there are limitations in terms of latency and connectivity when it comes to deploying digital assistants on devices. This necessitates reducing the size of chatbot models so that they can fit on user devices like Siri, Alexa, or Google Assistant. In this study, the authors demonstrate that low-parameter models can retain their conversational abilities while also improving in a specific domain. They propose a generic framework that addresses various question types, tracks references in multi-turn conversations and eliminates inconsistent and potentially toxic responses. The framework seamlessly transitions between casual chatting and performing transactional tasks making interactions with digital assistants more human-like. The authors evaluate their framework on both internal and public benchmark datasets using automatic (Perplexity) and human (SSA - Sensibleness and Specificity Average) evaluation metrics. They achieve comparable performance while reducing model parameters by 90%.