Skip to main content

​
Overview

One of the most critical aspects of building a fair fake news detector is preventing source-based bias. Without proper bias mitigation, the model could learn to associate certain news agencies or locations with “real” or “fake” labels, rather than analyzing the actual content. This page explains how the detector implements anti-bias preprocessing to ensure classification is based purely on content quality and writing patterns.

​
The Metadata Problem

Many news articles begin with metadata that identifies the source: 
WASHINGTON (REUTERS) - The Federal Reserve announced...
NEW YORK (AP) - Stock markets rallied today...
BRUSSELS (AFP) - European Union leaders agreed...
If this metadata is left in the training data, the model will learn shortcuts:
  • Articles mentioning “REUTERS”, “AP”, or “AFP” → classified as “real”
  • Articles without agency tags → potentially classified as “fake”
This is not true learning - it’s pattern matching based on source credibility, not content analysis.

​
Implementation: Regex-Based Metadata Removal

The limpiar_texto function in /home/daytona/workspace/source/fake_news_ia.py:54-69 implements metadata removal as the first step of text preprocessing: 
def limpiar_texto(texto):
    # 1. ELIMINAR METADATA/FUENTE: Quita patrones como 'LUGAR (AGENCIA) - '
    texto = re.sub(r'([A-Z\s]+)\s*\((REUTERS|AP|AFP)\)\s*\-\s*', '', str(texto), flags=re.IGNORECASE)
    
    # 2. Convertir a minúsculas
    texto = str(texto).lower()
    
    # 3. Eliminar puntuación, números y caracteres especiales
    texto = re.sub(r'[^a-z\s]', '', texto) 
    
    # 4. Tokenización con split() 
    tokens = texto.split() 

    # 5. Filtrar stopwords y tokens de una sola letra
    tokens = [t for t in tokens if t not in stop_words and len(t) > 1]
    return " ".join(tokens)

​
The Critical Regex Pattern

Let’s break down the bias-removal regex (fake_news_ia.py:56): 
r'([A-Z\s]+)\s*\((REUTERS|AP|AFP)\)\s*\-\s*'
Pattern Components:
ComponentMatchesExample
([A-Z\s]+)One or more capital letters or spacesWASHINGTON, NEW YORK
\s*Optional whitespaceHandles spacing variations
`((REUTERSAPAFP))`News agency in parentheses(REUTERS), (AP), (AFP)
\s*\-\s*Hyphen with optional surrounding spaces-
Input: 
WASHINGTON (REUTERS) - The Federal Reserve announced interest rates...
Output: 
The Federal Reserve announced interest rates...
The flags=re.IGNORECASE parameter ensures that variations like (Reuters) or (reuters) are also caught and removed.

​
Why This Step Comes First

Notice that metadata removal happens before any other preprocessing:
  1. Remove metadata ← Prevents source bias
  2. Convert to lowercase
  3. Remove punctuation/numbers
  4. Tokenize
  5. Remove stopwords
If we removed metadata after lowercasing, the regex wouldn’t match (REUTERS) anymore. The order matters!

​
Impact on Model Fairness

By removing source metadata, the model must classify based on:
  • Writing quality: Grammar, coherence, structure
  • Content patterns: Sensationalism, conspiracy language, vague claims
  • Semantic features: Word choice, context, logical flow
This produces a content-based classifier rather than a source-based filter.
If you modify the preprocessing pipeline, always ensure metadata removal happens first. Skipping this step will significantly reduce model fairness and generalization.

​
Production Consistency

The same limpiar_texto function is used in both training (fake_news_ia.py:54) and production (app.py:24). This ensures:
  • Consistent preprocessing between training and inference
  • No source-based shortcuts in real-world predictions
  • Fair classification regardless of article origin

​
Extending the Anti-Bias Filter

You can extend the regex to catch additional metadata patterns: 
# Also remove dateline patterns like "LONDON, March 3 -"
texto = re.sub(r'([A-Z\s]+),\s*\w+\s*\d+\s*\-\s*', '', texto)

# Remove "By [Author Name]" bylines
texto = re.sub(r'By\s+[A-Z][a-z]+\s+[A-Z][a-z]+', '', texto, flags=re.IGNORECASE)
When adding new bias mitigation rules, test them on your training data first to ensure they don’t accidentally remove legitimate content.

​
Validation

To verify bias mitigation is working: 
# Test the function
test_text = "WASHINGTON (REUTERS) - The president announced today..."
cleaned = limpiar_texto(test_text)
print(cleaned)
# Should NOT contain "WASHINGTON", "REUTERS", or the hyphen

​
Key Takeaways

  • Source metadata creates unfair shortcuts for ML models
  • Regex-based removal ensures content-focused classification
  • Metadata removal must happen first in the preprocessing pipeline
  • The same preprocessing must be used in training and production
  • Fair models generalize better to new, unseen sources
Next: Learn how to save and load trained models with Model Persistence.

Build docs developers (and LLMs) love

Get started for freeTalk to us