Tutorial 4: Noise Injection

Configure the distractor system for robust model training.

Three-Layer Architecture

Layer	Weight	Intent	Slots
Surface Noise	50%	Preserved	Preserved
Incomplete Info	30%	Clear	Missing
Semantic Ambiguity	20%	Uncertain	May conflict

Configuration

distractor:
  enabled: true
  use_semantic: true
  activation_probability: 0.25   # 25% of queries get noise
  strategy_path: "input/distractor_strategy.yaml"

Strategy Configuration

input/distractor_strategy.yaml:

layer_weights:
  surface_noise: 0.5
  incomplete_info: 0.3
  semantic_ambiguity: 0.2

strategies:
  surface_noise:
    colloquial_speech:
      mandatory: true
    typo_misspelling:
      probability: 0.3

Programmatic Usage

from src.distractor import SemanticDistractorModel

distractor = SemanticDistractorModel(config)

result = distractor.apply_noise(
    text="Help me write a Python script",
    persona_features={'style': 'casual'}
)

print(f"Original: {result.original_text}")
print(f"Noisy: {result.noisy_text}")
print(f"Layer: {result.layer}")
print(f"Strategies: {result.applied_strategies}")

Examples

Surface Noise:

Original: "Help me write Python code"
Noisy:    "hey can u help me writ some python"

Incomplete Info:

Original: "Book a flight to Paris on Monday"
Noisy:    "I need to book a flight sometime next week"

Semantic Ambiguity:

Original: "Fix this bug"
Noisy:    "Fix this bug and also maybe refactor... actually just the bug"

Next Steps

Continue to Tutorial 5: Advanced Usage.