Knowledge Graph Overview

Back to NLP Topics

A Knowledge Graph (KG) converts static FAQ text into a structured conversational experience. It supports two models: an ontology-based model using hierarchical terms, and an LLM-based Few-Shot model that requires no ontology.

---

Knowledge Graph Types

Ontology Knowledge Graph

Organizes FAQs using terms, synonyms, traits, and context. When a user asks a question, the engine matches utterance tokens against KG nodes (path qualification), then scores shortlisted questions using cosine similarity.

Enable: Go to Automation AI > Natural Language > NLU Config > Engine Tuning > Knowledge Graph and select Ontology Model.

How it works:

1. User utterance and KG nodes are tokenized; n-grams extracted (up to quad-gram).
2. Tokens are mapped to KG nodes to get indices.
3. Path qualification: paths are shortlisted based on term coverage and mandatory term presence.
4. Best match is selected by cosine scoring over shortlisted questions.

Training process:

1. All terms/nodes and their synonyms are indexed.
2. A flattened path is established for each KG intent using those indices.

Few-Shot Knowledge Graph

Uses Kore.ai’s LLM to identify FAQs by semantic similarity—no ontology needed. Add all FAQs to the root node.

Enable: Go to Automation AI > Natural Language > NLU Config > Engine Tuning > Knowledge Graph and select Few-Shot Model.

Prerequisites before enabling:

• Requires NLP V3 and Ranking & Resolver V2 (auto-updated when enabled).
• Embedding model options: BGE M3 and Pretrained MPNet.
• When switching from Ontology KG: Default terms are retained until updated, then become Organizer terms (can be set as Mandatory).
• Only Mandatory terms support path-level synonyms.

How it works:

The LLM computes semantic similarity between the user utterance and FAQs, returning a similarity score. The score determines match type (definite, probable, etc.) based on thresholds. Matched intents are sent to Ranking and Resolver to select the winner.

Embeddings Model Accuracy

Model accuracy across multiple languages using a single dataset, enabling a comparative evaluation of multilingual performance.

Model	Arabic	German	English	Spanish	French	Hindi	Japanese	Korean	Dutch	Polish	Avg. Accuracy
MPNet	34.98%	62.56%	82.76%	58.62%	64.53%	31.53%	49.75%	38.42%	58.13%	50.74%	53.20%
Pretrained MPNet	36.45%	65.02%	83.25%	61.08%	66.01%	31.53%	50.74%	38.92%	60.10%	55.17%	54.83%
LaBSE	71.43%	78.33%	80.79%	77.34%	79.80%	76.85%	80.79%	79.31%	76.85%	76.85%	77.83%
BGE-M3	76.35%	82.76%	83.74%	80.79%	80.79%	80.30%	84.73%	82.76%	82.76%	79.80%	81.48%

Model accuracy for English across different datasets from multiple domains, providing a concise summary for quick comparison.

Model	Avg. Accuracy
MPNet	79.47%
Pretrained MPNet	80.60%
LaBSE	78.25%
BGE-M3	83.49%

---

Selecting Your KG Type

Starting with v10.1, Few-Shot is the default for new KGs under NLP V3 in English.

Go to Automation > Knowledge AI > FAQs to switch types.

Before changing KG type, back up your existing graph by creating a new app version or exporting as JSON or CSV. Changes are captured in App Settings > Change Logs.

---

Feature Comparison

Feature	Few-Shot KG	Ontology KG
Ontology Structure	Optional	Mandatory
Default Terms	No (exception: existing terms when switching from Ontology)	Yes
Mandatory Terms	Yes	Yes
Organizer Terms	Yes	Yes
Path Qualification	No	Yes
Tags	Yes	Yes
Synonyms	Yes (Mandatory Terms and Tags only)	Yes
Path-Level Synonyms	Yes (Mandatory Terms only)	Yes
Knowledge Graph Synonyms	Yes (Mandatory Terms only)	Yes
Traits	Yes	Yes
Context	Yes	Yes
Stop Words	Yes	Yes
KG Import/Export	Yes	Yes
Auto-Generate KG	Yes	Yes
Bot Synonyms	Yes	Yes
Lemmatization using Parts of Speech	No	Yes
Path Coverage	No	Yes
Search in Answer	No	Yes
Qualify Contextual Paths	No	Yes
Auto-Correction	Yes	Yes
Min/Definitive Level for KG Intent	Yes	Yes
KG Suggestions Count	Yes	Yes
Proximity of Suggested Matches	Yes	Yes
Manage Long Responses	Yes	Yes
Intent Preconditions	Yes	Yes
Context Output	Yes	Yes
Supports All Platform Languages	Yes	Yes

---