> ## Documentation Index
> Fetch the complete documentation index at: https://koreai.mintlify.app/llms.txt
> Use this file to discover all available pages before exploring further.

# Natural Language Processing

<Badge icon="arrow-left" color="gray">[Back to NLP Topics](/ai-for-service/automation/natural-language/nlp-topics)</Badge>

The Platform uses a multi-engine NLP approach to identify user intent, extract entities, and execute the correct task. Understanding NLP helps you train your VA for accurate intent detection and natural conversations.

***

## NLP Engines

| Engine                          | Description                                                                                                                          |
| ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ |
| **Fundamental Meaning (FM)**    | Computational linguistics built on ChatScript. Analyzes word meaning, position, conjugation, capitalization, and sentence structure. |
| **Machine Learning (ML)**       | Trains on example utterances; learns and generalizes to recognize similar inputs.                                                    |
| **Knowledge Graph (KG)**        | Converts FAQ content into structured conversational responses.                                                                       |
| **Traits Engine**               | Multi-class classifier identifying characteristics in utterances to refine intent detection.                                         |
| **Small Talk Engine**           | Handles conversational pleasantries to make interactions feel natural.                                                               |
| **Ranking and Resolver (R\&R)** | Scores and ranks results from all engines to determine the winning intent.                                                           |

***

## Conversation Flow

1. **NLP Analysis** — User input passes through all NLP engines for intent detection and entity extraction.
2. **Task Execution** — The winning intent executes. The conversation engine maintains state (user details, previous intents, context).
   * **Preconditions** — If required conditions aren't met, the intent is rejected.
   * **Negative patterns** — Patterns that prevent incorrect intent matches.
   * **Event handling** — Welcome messages, sentiment, etc.
3. **Interruption Handling** — Manages mid-task intent switches and sentiment-triggered agent transfers.
4. **Response Generation** — A response is generated and rendered for the user's channel.

***

## NLP Preprocessing

Before intent detection, each utterance undergoes:

| Step                  | Description                                                                                  |
| --------------------- | -------------------------------------------------------------------------------------------- |
| **Tokenization**      | Splits utterance into sentences, then words. Uses TreeBank Tokenizer for English.            |
| **toLower()**         | Converts to lowercase. Not applied to German (word meaning changes by case). ML and KG only. |
| **Stop word removal** | Removes low-signal words. Language-specific; disabled by default. ML and KG only.            |
| **Stemming**          | Cuts words to their stem (e.g., "Working" → "work"). Output may not be a valid word.         |
| **Lemmatization**     | Converts to base form using a dictionary (e.g., "housing" → "house").                        |
| **N-grams**           | Combines co-occurring words (e.g., "New York City") for richer context.                      |

***

## Choosing an NLP Engine

| Scenario                                                            | Recommended Engine                                                                                               |
| ------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------- |
| Large corpus per intent                                             | **ML** — flexible, auto-learns from examples. Corpus of 200-300 for distinct intents; 1000+ for similar intents. |
| Query/FAQ-type intents or document-based answers                    | **KG** — semantic matching for knowledge content.                                                                |
| Idiomatic, command-like sentences, or tolerance for false positives | **FM** — deterministic, rule-based.                                                                              |

***

## NLP Configuration in the Platform

Go to **Automation > Natural Language**:

| Section        | Purpose                                                          |
| -------------- | ---------------------------------------------------------------- |
| **Training**   | Add ML utterances, synonyms, concepts, and patterns.             |
| **NLU Config** | Set confidence thresholds, engine tuning, and advanced settings. |

For engine-specific training and tuning, see:

* [Machine Learning Engine](/ai-for-service/automation/natural-language/training/machine-learning-engine#machine-learning-engine)
* [Fundamental Meaning Engine](/ai-for-service/automation/natural-language/training/fundamental-meaning#fundamental-meaning-engine)
* [Knowledge Graph Training](/ai-for-service/automation/knowledge-ai/knowledge-graph-training#knowledge-graph-training)
* [Traits](/ai-for-service/automation/natural-language/training/traits#traits)
* [Ranking and Resolver](/ai-for-service/automation/natural-language/training/ranking-and-resolver#ranking-and-resolver-engine)