# NLP is Not enough -- Contextualization of User Input in Chatbots

**Authors:** Nathan Dolbir, Triyasha Dastidar, Kaushik Roy  
**arXiv:** [2105.06511](https://arxiv.org/abs/2105.06511)  
**Date:** 2021-05-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2105.06511v1/

## Key points

- AI chatbots have made significant advancements in recent years and are being used in various industries, including healthcare.
- The application of AI chatbots in healthcare is particularly appealing due to the potential to reduce costs and alleviate strain on the healthcare system.
- Deep networks efficiently process user requests but struggle with capturing safe and medically accurate information due to variations in user text and speech.
- Knowledge in symbolic structures is better suited for accurate reasoning but cannot directly handle natural language processing.
- This paper aims to explore the combination of knowledge and neural representations in chatbots to improve safety, accuracy, and understanding.
- The study examines the effects of integrating deep learning with symbolic AI processing on chatbot interactions, focusing on safety, accuracy of medical information parsed, and user engagement.
- An experiment using a GPT-2 model trained on therapist conversation data was conducted to analyze its performance before proposing potential solutions to address its limitations.
- The key contributions of this research include scrutinizing a language model's ability to interact with mental health patients when fine-tuned on therapist conversations, identifying issues, and proposing solutions that leverage knowledge with deep learning methods for an optimal user experience.
- The findings suggest that incorporating domain-specific knowledge improves the chatbot's understanding of user inquiries and enhances its capability to generate accurate responses.

## Summary

AI chatbots have made significant advancements in recent years and are being used in various industries, with their application in healthcare particularly appealing due to the potential to reduce costs and alleviate strain on the healthcare system. While deep networks which rely on advanced Natural Language Processing techniques efficiently process user requests, they struggle with capturing safe and medically accurate information due to variations in user text and speech. On the other hand, knowledge in symbolic structures is better suited for accurate reasoning but cannot directly handle natural language processing. This paper aims to explore the combination of knowledge and neural representations in chatbots to improve safety, accuracy, and understanding. The study examines the effects of integrating deep learning with symbolic AI processing on chatbot interactions in terms of safety, accuracy of medical information parsed, and user engagement. The authors conduct an experiment using a GPT-2 model trained on therapist conversation data to analyze its performance before proposing potential solutions to address its limitations. The key contributions of this research include scrutinizing a language model's ability to interact with mental health patients when fine-tuned on therapist conversations; identifying issues; and proposing solutions that leverage knowledge with deep learning methods for an optimal user experience. The findings suggest that incorporating domain-specific knowledge improves the chatbot's understanding of user inquiries and enhances its capability to generate accurate responses.