Skip to main content

​
What is text-based emotion prediction?

This machine learning microservice analyzes text content to detect inappropriate language, classify emotional sentiment, and categorize messages across six toxicity dimensions. The model uses a combination of neural networks and traditional NLP techniques to provide comprehensive text analysis. The service identifies:
  • Inappropriate language detection - Binary classification of overall content appropriateness
  • Emotional sentiment - Positive or negative emotion classification
  • Six toxicity categories - toxic, severe_toxic, obscene, threat, insult, and identity_hate

​
Key features

Multi-model architecture

Combines PyTorch neural networks with scikit-learn classifiers for comprehensive text analysis

REST API endpoint

Simple HTTP GET request to analyze any text and receive detailed predictions

Real-time inference

Get instant predictions on text content through a lightweight Flask microservice

Six toxicity classifiers

Detect toxic, severe_toxic, obscene, threat, insult, and identity_hate content

​
Model architecture

The service uses a hybrid approach with two complementary models:

​
PyTorch neural network

A custom feedforward neural network (base_line class in microservice.py:51-70) processes word embeddings:
  • Input layer: 10-dimensional word embeddings from a trained embedding matrix
  • Hidden layers: Three fully connected layers (2048 → 1024 → 512 neurons)
  • Activation: ReLU activation between layers
  • Output layer: 6 neurons with sigmoid activation for multi-label classification
  • Architecture:
class base_line(nn.Module):
  def __init__(self,fin,out):
    super(base_line,self).__init__()
    self.out = out
    self.fin = fin
    self.fc1 = nn.Linear(self.fin,2048)
    self.fc2 = nn.Linear(2048,1024)
    self.fc3 = nn.Linear(1024,512)
    self.relu = nn.ReLU()
    self.fc4 = nn.Linear(512,self.out)
    self.sigmoid = nn.Sigmoid()
The model outputs six probability scores (threshold: 0.29) for each toxicity category.

​
Scikit-learn emotion classifier

A traditional machine learning classifier handles sentiment analysis:
  • Text vectorization: TF-IDF vectorizer converts cleaned text to numerical features
  • Classification: Binary positive/negative emotion prediction
  • Integration: Works alongside the neural network to provide overall sentiment
The neural network model weights are loaded from models/model_26_87.12.pth, achieving 87.12% accuracy on the validation set.

​
Toxicity categories

Each text input is evaluated across six dimensions:
CategoryDescription
toxicGeneral toxicity in language
severe_toxicExtremely toxic content
obsceneObscene or vulgar language
threatThreatening language
insultInsulting content
identity_hateHate speech targeting identity groups
A text is marked as inappropriate when all six categories exceed the 0.29 threshold (microservice.py:219).

​
Get started

Quickstart

Make your first prediction in under 5 minutes

Installation

Detailed setup instructions for local development

Build docs developers (and LLMs) love

Get started for freeTalk to us