Machine Learning Engine

Back to NLP Topics The ML engine maps user utterances to intents using example-based training. It supports supervised and unsupervised learning, with a range of network types for different use cases.

How the ML Engine Works

Intent detection pipeline:

User utterance is preprocessed (tokenization, stop word removal, lemmatization, n-grams).
The ML model classifies the utterance as a probable or definitive match against trained intents.
Results are passed to Ranking and Resolver.

Entity extraction: After intent is identified, the engine extracts entity values from the utterance using NER (Named Entity Recognition) trained annotations. Supervised vs. unsupervised learning:

Unsupervised (auto-expand): The assistant learns from all successful interactions automatically.
Supervised (recommended): Developers monitor interaction logs, tune NLP settings for failures, and retrain.

Add ML Utterances

Go to Natural Language > Training > Intents.
Select an intent and add sample utterances.
Click Train to update the model.

Utterance guidelines:

Utterances don’t need to be full sentences; phrases work too.
For distinct intents (for example, “Find Flight” vs. “Book Hotel”): 200-300 utterances each.
For similar intents (for example, “Change Seat” vs. “Change Flight”): 1000+ utterances each.
For Deep Neural Networks: larger datasets (data-hungry models) produce better results.

ML Network Types

Configure under Natural Language > NLU Config > Engine Tuning > ML.

Network Type	Description
Standard (default)	Bag-of-words classification. Suitable for most use cases.
Zero-shot Model	No training required; uses semantic similarity. Requires NLP V3.
Few-shot Model (Kore.ai Hosted Embeddings)	LLM-based semantic similarity. Reduces training data requirements. Requires NLP V3.
MLP-BOW (Deprecated)	Multi-layer perceptron with bag-of-words features.
MLP-WordEmbeddings (Deprecated)	MLP with pre-trained word embeddings.
LSTM (Deprecated)	Recurrent neural network. Captures sequential context.
CNN (Deprecated)	Convolutional neural network. Captures local n-gram patterns.
Transformer (Deprecated)	Transformer-based model. High accuracy for English.
KAEN (Deprecated)	Kore.ai custom model. Improved accuracy for English.

Thresholds and Configuration

Go to Natural Language > NLU Config > Engine Tuning > Machine Learning.

Setting	Description
ML Threshold	Minimum confidence score for a probable match (default: 0.3).
Auto Training	Automatically retrain the model when utterances are added or modified.
Network Type	Select the ML model architecture.
Stop Words	Enable/disable stop word removal during training.
Synonyms in ML	Include bot synonyms in ML training.

Synonyms and Concepts for ML

Bot Synonyms: Go to Training > Bot Synonyms. Add synonyms for words in task and entity names. Synonyms are app-wide.

To add: click New Synonym, enter the keyword and its synonyms (press Enter after each).
Parallel editing is supported — multiple users can add/delete synonyms concurrently without data conflicts.

Entity Synonyms: Go to Training > Entities. Add synonyms per entity node. Concepts: Custom concept groups prefixed with ~. Define in Training > Concepts. Used in patterns and entity detection.

App Substitutions

Define domain-specific word replacements using the ~bot_substitutions concept:

Chevy|Chevrolet    → replaces "Chevy" with "Chevrolet"
?_km|kilometers    → replaces "5 km" with "5 kilometers"

Syntax: original_word|replacement (case-insensitive match, case-sensitive replacement).

App Word Aliases

Define alternate spellings or STT transcription corrections using Bot Synonyms with a + prefix. Example: +Alpha Romeo as an alias for Alfa Romeo — corrects speech-to-text transcription errors.

Aliases don’t change the original utterance.
They normalize values returned for Custom Concept entities.

Homophones

Handle words that sound the same but have different spellings. Define using a concept with | separator:

write|right|rite

The engine considers all homophones during POS tagging and replaces the word if a homophone is a better grammatical fit. To define: Training > Concepts > New Concept — enter homophones separated by |.

Emoji Support

Import default emoji mappings into system concepts to handle emoji inputs:

Go to Natural Language > Training > Intents > ⋯ > Import Emojis into Concepts.
Emojis are added to system concepts (for example, ~emohappy, ~emohello).
Review and remove unwanted emojis after import.

Patterns for Intents and Entities

See NLP Guidelines — Patterns for operators, syntax, and examples. Add intent pattern:

Training > Intents → click Patterns / +Pattern for an intent.
Enter patterns; press Enter between each.

Add entity pattern:

Training > Entities → click the edit icon for an entity.
Select the Patterns tab → enter patterns.

Add intent rules:

Training > Intents → click Rules / +Rule for an intent.
Enter Trait or context tag conditions using AND/OR operators.

Negative patterns:

Enable at NLU Config > Ranking and Resolver Engine > Negative Patterns.
Per intent: Training > Intents → select intent → Negative Patterns tab → add patterns.

NLU Config — FM Engine Tuning

Setting	Description
Intent Detection using Task Name Words	When disabled, intent is detected only if input starts/ends with the exact task name. Useful for tasks with shared words (for example, multiple “order” tasks). Enabled by default.
FM Threshold	Max % above the top-scoring intent allowed for other intents to be considered (0-20%; default 2%). Supported for English, Spanish, French, German.

​How the ML Engine Works

​Add ML Utterances

​ML Network Types

​Thresholds and Configuration

​Synonyms and Concepts for ML

​App Substitutions

​App Word Aliases

​Homophones

​Emoji Support

​Patterns for Intents and Entities

​NLU Config — FM Engine Tuning