NLP Settings and Guidelines

Back to NLP Topics Best practices for intent naming, ML training, patterns, synonyms, and concepts. Follow these guidelines before building or training your assistant.

---

​

Intent Naming

Do:

• Use 2-4 words: action verb + object + optional modifier. Example: Send Detailed Report, Get 3 Day Forecast.
• Use the same verb across similar tasks: Show Issue / Show Report, not Show Issue / Get Report.
• Use numerical form for digits: Get 3 Day Forecast.

Don’t:

• Use single-word actions.
• Use determiners (the, a, my, that) or pronouns.
• Use platform terms: task, alert, action, cancel, discard, amend, webhook.
• Use entity values in intent names (for example, avoid “Get Weather Today” where “Today” is an entity).
• Use special characters: () & / \ $ [ ] + *
• Use punctuation: - , . ! ? ' "
• Use the app name or task type as part of the intent name.
• Use the same word as both verb and noun across tasks (for example, Update Issue / Get Updates).

Dialog Tasks: Map to verb + object — “the goal is to Do Something.” Alerts: Map to object only (no verbs) — “alert me on: Status Updates.”

---

​

ML Training

General:

• Use batch test suites to compare ML models — there’s no fixed best configuration; iterate and compare.
• Prepare utterances that are as diverse as possible.
• Avoid noise (pleasantries); if unavoidable, spread it evenly across intents.
• For large datasets, stop words and synonyms are auto-recognized by ML — explicit configuration isn’t required.

Using the Confusion Matrix:

• Reassign utterances in False quadrants to the correct intent.
• Improve cohesion by rephrasing or adding synonyms.
• Avoid phrases that are similar across different intents.

Using K-fold Validation:

• Ideal for large datasets (minimum 250 utterances total, 5 per intent).
• Track Precision, Recall, and F1. Prioritize a higher Recall score.

---

​

Patterns

Patterns match idiomatic or non-obvious expressions that synonyms can’t cover. Pattern matches take priority over synonym matches. General rules:

• Use at minimum 3 words.
• Use canonical forms: infinitive verbs, singular nouns.
• Use lowercase; use US spelling.
• Avoid determiners, pronouns, digits.
• Don’t use elision (what's), special characters, or punctuation.

​

Pattern Operators

​

Entity Patterns

• Include the positional wildcard * to mark where the entity appears.
• Use surrounding words to delimit the entity search range.
• Use < and > to mark sentence boundaries.
• Don’t use field names or their synonyms in entity patterns.
• Only the first instance with a valid value is extracted.

Example — Extracting ToFlight and FromFlight:

• ToFlight pattern: to *1
• FromFlight pattern: from * to and to from *1

​

Negative Patterns

Negative patterns prevent false intent matches. Example: was trying to * — prevents triggering Book a Flight when the user says “I was trying to book a flight when I faced an issue.” To enable: Natural Language > NLU Config > Ranking and Resolver Engine > Negative Patterns.

---

​

Synonyms

Use synonyms when the same concept can be expressed with different words. Synonyms are unidirectional (tool=bar ≠ bar=tool). Do:

• Use canonical forms (infinitive verbs, singular nouns).
• Use lowercase and US spelling.
• Keep phrases under 5 words.
• Use intent-driven synonyms: get is a synonym for show if the action means “find and display.”

Don’t:

• Add synonyms to determiners, pronouns, or digits.
• Use a synonym that could match conflicting tasks.
• Use special characters or punctuation.
• Assign synonyms to multiple words in a single entry.
• Abbreviate to fewer than 2 letters.

​

Synonym Match Types

---

​

Concepts

Concepts group related terms under a single label, usable in patterns. Naming rules:

• Must start with ~.
• Allowed characters: a-z, A-Z, 0-9, _.
• At least one letter after ~.
• Can’t start or end with _.
• Case-insensitive.

Valid: ~my_concept, ~Sample, ~test123 Invalid: ~_concept, ~concept_, ~a-concept, ~123test Use concepts in patterns: buy ~fruit matches buy orange, buy banana, etc. System concepts: Pre-built by the platform (for example, ~emohello, ~yes, ~no). Add the ~disable_help concept to disable the default help response.

---

​

Standard Responses

Standard responses are template messages for common conversational situations: ambiguity resolution, authorization requests, confirmations, interruptions. Customize them at Natural Language > Standard Responses. Responses support:

• Plain text or JavaScript for dynamic content.
• Contextual tags replaced at runtime (for example, <list-of-tasks>).
• Channel-specific formatting.

---

​

Knowledge Graph — Key Considerations

The KG model adds domain intelligence on top of ML and FM. Assign FAQs to nodes based on key domain terms to improve disambiguation. Traits in KG: Use when FAQs are mutually exclusive in nature. For example, “booking process” vs. “booking issues” — creating separate traits ensures the engine doesn’t suggest booking process FAQs for issue-related queries. Use traits judiciously — overuse causes false negatives. Ensure:

• Good trait coverage across utterances.
• Traits are mutually exclusive where intended.
• All relevant FAQs are tagged to the correct trait.