# Custom NLP Rules

# Custom NLP Rules

Most production NLP setups require domain tuning. “Custom NLP rules” refers to the mechanisms you use to make extraction and linking align with your data and vocabulary.

# Why custom rules are needed

Off-the-shelf NLP is rarely enough for:

• product catalogs (aliases, abbreviations, versions)
• internal identifiers (ticket/asset formats)
• domain jargon (acronyms, shorthand)
• noisy text (chat logs, pasted stack traces)

# Common rule categories

• Aliases and normalization
  • map spelling variants and formatting differences to canonical forms
• Dictionaries / spotters
  • curate controlled vocabularies from your graph (or external lists)
• Patterns
  • capture IDs and semi-structured entities using shape/pattern rules
• Exclusions
  • prevent over-capture in common false-positive contexts

# Recommended process

1. Collect 50–200 representative text samples.
2. Run extraction experiments.
3. Add:
   • missing synonyms/aliases
   • pattern constraints
   • exclusions for false positives
4. Re-run experiments and measure improvement.
5. Enable linking only when precision is acceptable for your use case.

# Next steps

• Configure pipelines and extraction: Entity Extraction
• Turn extracted concepts into facets for retrieval: Search Optimization