Stereotype and Category Classification

Overview

This project builds upon HEART’s research to classify stereotypes in textual data. The goal is to create lightweight, efficient models capable of categorizing text into stereotype themes while considering sustainability and inference time.

Key Features

1. Stereotype Classification:

   • Classifies text into specific stereotype categories (e.g., racial, gender-based stereotypes).
   • Uses a combination of TF-IDF, cosine similarity, and embedding techniques.

2. Category Classification:

   • Distinguishes between stereotype-related and unrelated text.
   • Enhances decision-making through a hierarchical model pipeline.

3. Generative Model Exploration:

   • Investigates the use of large-scale generative models (e.g., Llama LoRA fine-tuning) for stereotype detection.
   • Approach would entail coarse to fine modelling from high level theme/ stereotype_type tokens, to then the specific tokens in the sentence that leads to the classification descision (during training), with the final stereotype, anti-stereotype, or unrelated token.
   • This approach is computationally intensive and was not implemented due to its impracticality for real-time use.

Methodology

1. Generative Model Approach (Not Implemented)

• Concept: Fine-tune a generative language model (e.g., Llama 3 8B) to generate stereotype reasoning and classification.
• Challenges: High computational cost, slower inference times, and limited practicality for real-world applications.

2. Implemented Approach

• Model 1: Stereotype Type Classification

  • Features:
    • TF-IDF vectorization of input text.
    • Embeddings: Calculate mean embeddings for each stereotype class and compute cosine similarity with input text embeddings.
  • Classifier: LightGBM model trained on feature vectors combining TF-IDF and cosine similarity.

• Model 2: Category Classification

  • Features:
    • Outputs from Model 1 (class probabilities).
    • TF-IDF vectorization and cosine similarity features.
  • Classifier: LightGBM model trained on combined feature vectors from Model 1 and additional text representations.

Results

• Model 1: Stereotype Type Classification
  • F1 Score: 0.913
  • Carbon Emissions: 3.7 × 10^-5
• Model 2: Category Classification
  • F1 Score: 0.653

Business Value

• Moderation of malicious content on platforms like Discord or social media.
• Identification and mitigation of harmful stereotypes in online communication.

---

Limitations and Future Directions

Current Limitations:

• Only applicable to predefined stereotypes.
• Limited to short sentences; struggles with complex or long texts.

Potential Improvements:

• Incorporating deep learning for more complex text processing.
• Enhancing fairness and addressing inherent biases in the model.

Installation and Setup

1. Clone the repository:

   git clone https://github.com/alexkstern/steryclass_priv.git
   

2. pip install -r requirements.txt