TikTok Bias Detection Pipeline

A cohesive pipeline for detecting political and religious bias in TikTok videos by analyzing video content, text descriptions, and audio.

Overview

This project provides a comprehensive pipeline for analyzing TikTok videos for bias. It combines multiple analysis components:

1. Text Analysis: Analyzes video descriptions for political and religious bias using BERT-based models.
2. Video Analysis: Analyzes video content for visual bias indicators using video feature extraction.
3. Audio Analysis: Analyzes speech and audio content for bias using speech recognition and sentiment analysis.
4. Combined Scoring: Combines all analysis components with customizable weights to provide final bias scores.

The pipeline produces both individual component scores and a combined bias score for each video, categorizing them from "Minimal Bias" to "Extreme Bias".

Installation

Requirements

• Python 3.7 or higher
• PyTorch
• Transformers
• OpenCV
• Librosa
• MoviePy
• Pandas
• NumPy
• scikit-learn

Install from Source

# Clone the repository
git clone https://github.com/yourusername/tiktok-bias-detector.git
cd tiktok-bias-detector

# Install dependencies
pip install -r requirements.txt

# Install the package
pip install -e .

Usage

Python API Example

Here's a comprehensive example of using the TikTok Bias Detection API:

from pipeline import TikTokBiasDetectionPipeline

# Initialize the pipeline
pipeline = TikTokBiasDetectionPipeline()

# Update weights if needed (optional)
pipeline.update_weights(
    text_weight=0.4,
    video_weight=0.4,
    audio_weight=0.2,
    political_weight=0.3,
    religious_weight=0.3
)

# Process a batch of videos
df = pipeline.process_batch(
    input_csv="path/to/your/data_table.csv", 
    data_dir="path/to/video/directory",  
    output_csv="path/to/results.csv"
)

# Print a summary of the results
if df is not None:
    print("\nAnalysis Results Summary:")
    print(f"Total videos analyzed: {len(df)}")
    
    # Get average bias scores
    political_avg = df['political_bias'].mean()
    religious_avg = df['religious_bias'].mean()
    overall_avg = df['overall_bias_score'].mean()
    
    print(f"Average political bias: {political_avg:.2f}")
    print(f"Average religious bias: {religious_avg:.2f}")
    print(f"Average overall bias: {overall_avg:.2f}")
    
    # Count bias categories
    categories = df['overall_bias_category'].value_counts()
    print("\nBias category distribution:")
    for category, count in categories.items():
        print(f"  {category}: {count}")

Command Line Interface

Batch Processing

# Process all videos listed in a CSV file
python main.py --input data/data_table.csv --output data/results.csv

# Customize weights for different analysis components
python main.py --input data/data_table.csv --text-weight 0.4 --video-weight 0.4 --audio-weight 0.2

Single Video Analysis

# Analyze a single video
python main.py --video path/to/video.mp4 --description "Video description text"

Input Format

The input CSV file should have the following columns:

• Video_ID: Unique identifier for each video
• Description: Text description of the video
• Url, Likes, Comments, etc.: Additional metadata (optional)

Videos should be stored in the data/ directory with filenames matching the Video_ID (e.g., 12345.mp4).

Output Format

The output CSV includes all original columns plus:

• text_political_bias, text_religious_bias, text_combined_bias: Text analysis scores
• video_political_bias, video_religious_bias, video_combined_bias: Video analysis scores
• audio_political_bias, audio_religious_bias, audio_combined_bias: Audio analysis scores (if enabled)
• political_bias, religious_bias, overall_bias_score: Combined final scores
• political_bias_category, religious_bias_category, overall_bias_category: Categorization of bias levels

Configuration

Weights can be customized through command-line arguments, the Python API, or via a JSON configuration file:

{
  "text_weight": 0.4,
  "video_weight": 0.4,
  "audio_weight": 0.2,
  "political_weight": 0.3,
  "religious_weight": 0.3
}