Engine Tuning

Back to NLP Topics

Configure NLP engine thresholds and advanced settings per language under Natural Language > NLU Config > Engine Tuning.

For per-engine configuration details:

• Fundamental Meaning Engine
• Machine Learning Engine
• Knowledge Graph Training
• Ranking and Resolver

---

Advanced NLP Configurations

Default settings are appropriate for most use cases. Change only if you understand the impact — incorrect values can degrade performance.

Go to Natural Language > NLU Config > Engine Tuning > Advanced NLP Configurations.

ML Engine Settings

Configuration	Description	Valid Values	Notes
Split Compound Words	Splits compound words into stems before processing.	Enable / Disable (default)	German language only.
None Intent	Creates a placeholder intent to reduce ML false positives.	Enable (default) / Disable
Epochs	Training iterations for neural network.	20–300 (increments of 10; default: 20)	MLP-BOW, MLP-WordEmbeddings, LSTM, CNN only.
Batch Size	Training samples per batch.	10–30 (increments of 5; default: 10)	MLP-BOW, MLP-WordEmbeddings, LSTM, CNN only.
Learning Rate	Controls weight adjustment vs. loss gradient.	1e-4 to 1e-3 (default: 1e-3)	MLP-BOW, MLP-WordEmbeddings, LSTM, CNN only.
Dropout	Regularization to prevent overfitting.	0–0.8 (increments of 0.1; default: 0)	MLP-BOW, MLP-WordEmbeddings, LSTM, CNN only.
Vectorizer	Feature extraction technique.	count (default) / tfidf	MLP-BOW only.
Maximum Sequence Length	Length of training samples / user input.	10–30 (increments of 5; default: 20)	MLP-WordEmbeddings, LSTM, CNN only.
Embeddings Type	Feature extraction method.	generated / random (default)	MLP-WordEmbeddings, LSTM, CNN only.
Embeddings Dimensions	Dimensions used in featurization.	100–400 (increments of 50; default: 300)	MLP-WordEmbeddings, LSTM, CNN only.
K Fold	K parameter for cross-validation.	2–10 (default: 2)
Fuzzy Match	Enables fuzzy matching for intent identification.	Enable (default) / Disable
Handle Negation	Handles negated words in intent identification.	Enable (default) / Disable
Ignore Multiple Occurrences	Disregards word frequency during vectorization.	Enable (default) / Disable	MLP-BOW only.
Entity Placeholders in User Utterances	Replaces entity values in utterances with placeholders.	Enable (default) / Disable	MLP-BOW only.
Sentence Split	Splits sentences before intent detection.	Enable (default) / Disable
Multiple Intent Models	Separate ML model per primary intent (including sub-intents).	Enable / Disable (default)
Use only Tagged Utterances for NER Training	NER model trains only on utterances with NER annotations.	Enable (default) / Disable
Neurons in Hidden Layer	Number of hidden-layer neurons.	0–1000	Standard Network Type only.
Softmax Temperature	Controls how confidently ML identifies the winning intent.	0–100	Non-Standard Network Types only.
Spell Correction in ML	Enables spell correction on ML dictionary during prediction.	Enable / Disable (default)	English VAs only.
Spell Correction Version	Version of spell correction to apply.	Version 1 / Version 2	English VAs only. Applied across all engines.
Spell Correction Status	Enable/disable spell correction globally.	Enable / Disable	English VAs only.

KG Engine Settings

Configuration	Description	Valid Values
Lemmatize KG Synonyms	Uses lemmatized forms of KG synonyms for intent detection.	Enable (default) / Disable
Cosine Similarity Dampening	Prevents short questions from being penalized in cosine scoring.	Enable (default) / Disable
FAQ Name as Intent Name	Uses primary question as intent name when FAQ is linked to a dialog.	Enable / Disable (default)
FAQs Order for Disambiguation	Order of FAQs shown during ambiguity resolution.	Order by Hierarchy / Default Order (default)
Auto qualify FAQs from Fully Matched Paths	Qualifies all FAQs from a path when the path is fully matched.	Enable / Disable

Ranking and Resolver Settings

Configuration	Description	Valid Values	Notes
Intent Elimination Rules	Applies prebuilt rules to eliminate low-confidence intent matches.	Enable (default) / Disable	English, Spanish, French, German only.