← Back to Analyzer

🧠 How Our Model Works

Understanding the Multi-Task Learning architecture behind the AI Text Analyzer

🏗️ Multi-Task Learning Architecture

Our model uses a Multi-Task Learning (MTL) approach with shared layers to simultaneously predict three different aspects of text:

😊 Emotion Detection

6 emotions: sadness, joy, love, anger, fear, surprise

💬 Hate Speech

3 classes: offensive speech, Neither, Hate Speech

⚠️ Violence Type

5 types: sexual, physical, emotional, harmful traditional practice, economic

🔄 Architecture Flow:

1Input Layer: Text tokenized into sequences (max 50 tokens)
2Embedding Layer: Shared 128-dimensional word embeddings
3LSTM Layer: Shared 64-unit LSTM for sequence processing
4Pooling & Dropout: Global average pooling + 50% dropout
5Output Layers: 3 separate dense layers with softmax activation

📊 Training & Dataset

📚 Datasets Used:

  • Emotions: 12,000 balanced samples (2,000 per emotion)
  • Hate Speech: ~19,000 samples with balanced classes
  • Violence: ~20,000 samples across 5 violence types

⚙️ Training Configuration:

  • Optimizer: Adam
  • Loss: Sparse Categorical Crossentropy
  • Epochs: 10
  • Batch Size: 4

✨ Why Multi-Task Learning?

🎯 Better Generalization

Shared layers learn common linguistic patterns across tasks, improving overall performance.

⚡ Efficient Training

Single model handles multiple tasks, reducing computational resources.

🔄 Knowledge Transfer

Learning from one task helps improve performance on related tasks.

📈 Comprehensive Analysis

Get insights into multiple aspects of text simultaneously.