ChaCha (CHAtbot for CHildren's emotion Awareness): LLM-Driven Chatbot for Enhancing Emotional Awareness in Children

This repository is a source code of the chatbot implementation presented in the ACM CHI 2024 paper, titled "ChaCha: Leveraging Large Language Models to Prompt Children to Share Their Emotions about Personal Events."

Woosuk Seo, Chanmo Yang, and Young-Ho Kim. 2024.
ChaCha: Leveraging Large Language Models to Prompt Children to Share Their Emotions about Personal Events.
In Proceedings of ACM CHI Conference on Human Factors in Computing Systems (CHI'24). To appear.

Project Website

https://naver-ai.github.io/chacha/

System Requirements

1. Python 3.11.2 or higher
2. Poetry - Python project dependency manager
3. NodeJS and NPM - tested on 18.17.0
4. Paid OpenAI API key (ChaCha uses GPT-3.5 and GPT-4 models internally).

How To Run

Installation

1. In the root directory, install dependencies using poetry.

> poetry install

2. Install frontend Node dependencies

> cd frontend
> npm install
> cd ..

4. Run the setup script and follow the steps. It would help if you prepared the OpenAI API Key ready.

> poetry run python setup.py

Testing Chatbot on Command Line

3. Run chat.py on the command line:

> poetry run python chat.py

Testing Chatbot on Web

Running in development mode

1. Run backend server

   > poetry run python main.py

   The default port is 8000. You can set --port to designate manually.

   > poetry run python main.py --port 3000

2. Run frontend server

   The frontend is implemented with React in TypeScript. The development server is run on Parcel.

   > cd frontend
   > npm install <-- Run this to install dependencies 
   > npm run dev

   Access http://localhost:8888 on web browser.

You can perform the above steps using a shell script:

> sh ./run-web-dev.sh

Running in production mode

The backend server can serve the frontend web under the hood via the same port. To activate this, build the frontend code once:

> cd frontend
> npm run build

Then run the backend server:

> cd ..
> poetry run python main.py --production --port 80

Access http://localhost on web browser.

Analysis of Chat Logs

Chat Session Reviewing on Web

A session chat can be reviewed by visiting [domain]/share/{session_id}. There, you can also download the chat logs in CSV.

Log Files

To keep the framework lightweight, ChaCha leverages a file storage instead of a database. The session information and chat messages are stored in ./data/sessions/{session_name} in real time.

In the session directory, info.json maintains the metadata and the global configuration of the current session. For example:

{
  "id": "User_001",
  "turns": 1,
  "response_generator": {
    "state_history": [
      [
        "explore",
        null
      ]
    ],
    "verbose": false,
    "payload_memory": {},
    "user_name": "John",
    "user_age": 12,
    "locale": "kr"
  }
}

In the same location, dialogue.jsonl keeps the list of chat messages in a format of JsonLines, where each message is formatted as a single-lined json object.

Authors of the Code

• Young-Ho Kim (NAVER AI Lab) - Maintainer ([email protected])
• Woosuk Seo (Intern at NAVER AI Lab, PhD candidate at University of Michigan)

Acknowledgments

• This work was supported by NAVER AI Lab through a research internship.
• The conversational flow design of ChaCha is grounded in Woosuk Seo’s dissertation research, which was supported by National Science Foundation CAREER Grant #1942547 (PI: Sun Young Park), advised by Sun Young Park and Mark S. Ackerman, who are instrumental in shaping this work.