Skip to main content

​
Get your prediction in 3 steps

This guide walks you through setting up the microservice and making your first emotion prediction request.
1

Install dependencies

Create a virtual environment and install the required packages:
python3 -m venv text_prediction
source text_prediction/bin/activate
pip3 install -r requirements.txt
The installation includes PyTorch, Flask, NLTK, scikit-learn, and other ML libraries. This may take a few minutes.
2

Start the microservice

Run the Flask server to start accepting prediction requests:
python3 microservice.py
You should see output confirming the server is running:
Microserver running in port 3200
 * Serving Flask app 'microservice'
 * Running on http://127.0.0.1:3200
The service loads three pre-trained models on startup:
  • Word embedding dictionary (models/word_dict.json)
  • PyTorch neural network (models/model_26_87.12.pth)
  • Emotion classifier (models/emotion_classifier.model)
  • TF-IDF vectorizer (models/vectorizer2.pickle)
3

Make your first prediction

Send a GET request to the /textbased_emotion endpoint with your text:
curl "http://127.0.0.1:3200/textbased_emotion?text=This%20is%20a%20wonderful%20day"
The endpoint returns an HTML page with prediction results showing:
  • Whether the text contains inappropriate language
  • Emotion classification (Positive/Negative)
  • Scores for all six toxicity categories (toxic, severe_toxic, obscene, threat, insult, identity_hate)
  • Additional extracted entities (countries, people names, dates)

​
Understanding the response

The microservice processes your text through multiple stages:
  1. Text preprocessing (microservice.py:94-120)
    • Converts to lowercase
    • Expands contractions (“can’t” → “cannot”)
    • Removes hyperlinks and special characters
    • Applies lemmatization
  2. Neural network inference (microservice.py:203-204)
    • Converts cleaned text to word embeddings
    • Passes through the 4-layer neural network
    • Generates six probability scores
  3. Toxicity classification (microservice.py:206-219)
    • Applies 0.29 threshold to each score
    • Returns “Yes”/“No” for each category
    • Flags as inappropriate if all six categories are positive
  4. Sentiment analysis (microservice.py:221-229)
    • Vectorizes text using TF-IDF
    • Classifies overall emotion as Positive or Negative
    • Adjusts based on inappropriate content detection
The threshold value of 0.29 (microservice.py:206) balances precision and recall based on validation set performance.

​
Try different examples

Test the model with various text inputs: 
# Positive sentiment
curl "http://127.0.0.1:3200/textbased_emotion?text=I%20love%20this%20amazing%20product"

# Negative sentiment
curl "http://127.0.0.1:3200/textbased_emotion?text=This%20is%20terrible%20and%20disappointing"

# Text with entities
curl "http://127.0.0.1:3200/textbased_emotion?text=Meeting%20in%20Paris%20on%20March%2020,%202024"

​
Next steps

Installation guide

Learn about detailed setup options and model file requirements

Model architecture

Dive deeper into the neural network design and training

Build docs developers (and LLMs) love

Get started for freeTalk to us