Skip to main content

​
Overview

The emotion prediction system is built as a Flask microservice that combines multiple machine learning models to analyze text for toxicity classification and emotion detection. The architecture consists of three main components:
  1. Flask Application Layer - HTTP API server
  2. PyTorch Neural Network - Deep learning model for toxicity classification
  3. Scikit-learn Classifier - Traditional ML model for emotion prediction

​
Application Stack

The microservice runs on: 
HOST = "127.0.0.1"
PORT = 3200

​
Core Dependencies

  • Flask 2.2.2 - Web framework for HTTP endpoints
  • PyTorch 1.13.0 - Deep learning framework
  • scikit-learn 1.0.1 - Machine learning library
  • NLTK 3.7 - Natural language processing toolkit
The application is configured with JSON_AS_ASCII = False to properly handle Unicode characters in text responses.

​
Model Components

​
1. PyTorch Deep Neural Network

The base_line model is a multi-layer perceptron that classifies text into 6 toxicity categories:
  • Toxic
  • Severe Toxic
  • Obscene
  • Threat
  • Insult
  • Identity Hate
model = base_line(10, 6)  # 10-dimensional input, 6 outputs
model.load_state_dict(torch.load('models/model_26_87.12.pth'))
model.eval()

​
2. Word Embeddings

A 10-dimensional embedding layer converts words to numerical vectors: 
embedding = nn.Embedding(len(word_dict), 10)
The embedding dimension of 10 is relatively small, making the model lightweight and fast for production use.

​
3. Emotion Classifier

A pre-trained scikit-learn model loaded from pickle: 
emotion_classifier = pickle.load(open('models/emotion_classifier.model', 'rb'))

​
4. Text Vectorizer

Converts preprocessed text into feature vectors for the emotion classifier: 
vectorizer = pickle.load(open('models/vectorizer2.pickle', 'rb'))

​
Processing Pipeline

The system processes incoming text through the following stages:

​
Request Flow

  1. Input: GET request to /textbased_emotion?text=<input_text>
  2. Preprocessing: Text cleaning and lemmatization
  3. Parallel Processing:
    • PyTorch model analyzes toxicity (6 binary classifications)
    • Scikit-learn model predicts emotion
  4. Output: HTML template with all predictions
The threshold for toxicity classification is set at 0.29. Scores above this value indicate the presence of that toxicity type.

​
Model Files

The application requires the following model artifacts:
FilePurposeFormat
model_26_87.12.pthNeural network weightsPyTorch state dict
emotion_classifier.modelEmotion classifierPickle
vectorizer2.pickleText vectorizerPickle
word_dict.jsonWord-to-index mappingJSON
The model filename model_26_87.12.pth suggests this is epoch 26 with 87.12% accuracy.

​
Additional Features

Beyond emotion prediction, the system extracts:
  • Dates - Various date formats using regex patterns
  • Countries - Using pycountry library
  • People Names - Using NLTK POS tagging (NNP tags)
  • Time Ranges - Hour ranges in format “HH:MM - HH:MM”

​
API Endpoint

@app.route('/textbased_emotion', methods=['GET'])
def predict_textbased_emotion():
    text = request.args.get('text')
    # Processing pipeline...
    return render_template('index.html', ...)
The endpoint returns an HTML response with all extracted information and predictions.

Build docs developers (and LLMs) love

Get started for freeTalk to us