Empath AI

A multimodal AI chatbot for emotion detection. EmpathAI is an AI-powered chatbot designed to detect and interpret human emotions through audio and textual inputs. By combining natural language processing (NLP) andspeech signal analysis, the system aims to provide emotionally intelligent interactions.

---

About the project

The project uses:

• Speech analysis to detect pitch, prosody, and emotion.
• Text processing to understand sentiment, intent, and ambiguity.
• Future extensions include facial expression analysis via video input, enabling full multimodal emotion detection.
• This project is a FastAPI microservice for text sentiment analysis using classic machine learning (TF-IDF + Logistic Regression + SVM).
• It can be trained on datasets like GoEmotions.

---

🚀 Current Features (as per the latest update)

• Train your own sentiment classifier from a CSV dataset (text,label format).
• Predict fine-grained emotions like joyful, angry, hopeful, disappointed.
• Predict top 'k' emotions corresponding to a particular text.
• Lightweight and resource-friendly (no heavy LLMs required).
• Deployable as a Docker container.

---

File structure (think of another name)

Project Setup

1. Clone Repository

git clone https://github.com/rshaurya/Empath-AI.git
cd Empath-AI

2. Create a Virtual Environment

Make sure you have Python 3.11 installed. To verify installation:

python3.11 --version

If a version appears, then Python3.11 is installed.

Create a virtual environment:

python3.11 -m venv .venv

Activate the environment(Linux/Mac)

source .venv/bin/activate

Activate the environement (Windows PowerShell)

.venv\Scripts\activate

3. Install requirements

Navigate to the Empath-AI directory and run the following command to install the requirements:

pip install -r requirements.txt

By now, you must have created a virutal environment, activated it and installed all the dependencies. Follow further steps to run the application.

---

Running the Application

1. Merge CSV files

There are 3 files under train/. Run the following command to merge the 3 .csv files into one:

cat train/goemotions_*.csv > train/goemotions_full.csv

This will create a new file inside train/

2. Data Preprocessing

The train/goemotions_full.csv has a lot of columns like text, id, author_id, date etc. out of which only text and label are necessary for training our model.

• Run the bin/re-format-goemotions.py file using any code editor or run the following command:

python3.11 bin/re-format-goemotions.py

This will create a new file, train/goemotions_text_label.csv having only 2 columns i.e. text and label.

3. Run server

To run and test the application run the following command from the project root:

uvicorn app.main:app --reload  --port 8000

Now open your browser at: http://localhost:8000/docs
This gives you an interactive Swagger UI where you can try training and prediction endpoints. This is where you can test your API endpoints, just click on a particular end point and click the "Try it out" button. Swagger tells you a lot about your API. It tells you:

1. The type of method (GET, POST etc.).
2. The parameters required (if any) with its type.
3. the response you can expect from the API.
4. And the response type in detail.

4. Train the Model

• In the Swagger UI, select the /train endpoint (POST) and click on Try it out.
• Keep the default parameter use_logreg as false.
• Under the request_body upload the goemotions_text_label.csv file from your system.
• Then click execute. And let it do its magic. The model is being trained!
• This step will create the model model.joblib in the app/ directory.

5. Predict text

• Now, select the predict/ endpoint (POST) and click on Try it out.
• Under the request_body edit the values in json:

{
  "text": "string",
  "top_k": 1
}

• Here, text contains the sentence for which we have to detect the emotion. And top_k means the top 1 or 2 or 3 emotions corresponding to the particular text. It controls how many top labels are returned

Response:

{
  "text": "I am so happy today!",
  "prediction": "joyful"
}

---

Yay! The text sentiment analysis is completed! More features coming soon...