knowledge-graph-construction（知识图谱构建）

Table of Contents

• Workflow
• Architecture Selection Guide
• Schema Patterns
• Output Template

Knowledge Graph Construction

Workflow

Copy this checklist and work through each step:

KG Construction Progress:
- [ ] Step 1: Identify data sources and domain scope
- [ ] Step 2: Select graph data model
- [ ] Step 3: Design schema and ontology
- [ ] Step 4: Configure extraction pipeline
- [ ] Step 5: Define layered architecture
- [ ] Step 6: Validate and quality-check the graph

Step 1: Identify data sources and domain scope

Catalog the input data: document types (papers, clinical notes, web pages, logs), volume, update frequency, and language. Define the domain boundary -- what entity types and relation types matter for the target use case. Determine whether the KG will serve RAG retrieval, reasoning/inference, analytics, or a combination. This scoping step prevents over-extraction and keeps the schema focused.

Step 2: Select graph data model

Choose the underlying data model using the Architecture Selection Guide. Key trade-offs: LPG for flexibility and rapid prototyping, RDF/OWL for standards-based interoperability and inference, Hypergraphs for complex N-ary relations, Temporal Graphs for time-evolving knowledge. Consider query language, tooling maturity, and vector integration needs. For detailed model comparisons, see Data Models Reference.

Step 3: Design schema and ontology

Define node types (entity classes), edge types (relation classes), and property schemas. Apply patterns from Schema Patterns: entity-relation for simple domains, event reification for N-ary relations, layered tiers for multi-source integration. Decide on controlled vocabularies, cardinality constraints, and whether to adopt or extend an existing ontology (e.g., Schema.org, UMLS, SNOMED). For methodology details, see Methodology Reference.

Step 4: Configure extraction pipeline

Build the pipeline that populates the graph. Core components: LLM-assisted entity extraction with multi-round verification, relation extraction via prompt-based or dependency-parsing methods, entity normalization (synonym merging, ontology linking), and schema enforcement through post-processing validation. Use few-shot examples in prompts to improve extraction consistency. Include a second-pass LLM verification to catch missed entities. For full pipeline design, see Methodology Reference.

Step 5: Define layered architecture

Structure the KG into tiers for maintainability and trust. A common pattern: Layer 1 (instance data) holds user-specific or case-specific entities and relations; Layer 2 (domain knowledge) holds curated facts from literature or domain experts; Layer 3 (canonical ontology) holds the formal schema and upper ontology. Add provenance and evidence layering so every fact traces back to its source document, extraction method, and confidence score. Temporal subgraphs capture time-indexed state for domains where knowledge evolves.

Step 6: Validate and quality-check the graph

Run validation at multiple levels: schema conformance (do all nodes and edges match declared types?), coverage (are expected entity types populated?), consistency (no contradictory edges), and completeness (sample-based human review). Use a second LLM as a validator to fact-check extracted triples against source documents. Compute graph statistics (node degree distribution, connected components, orphan nodes) to identify extraction gaps. Quality criteria are defined in Quality Rubric.

Architecture Selection Guide

By Use Case

By Domain

Schema Patterns

Entity-Relation Pattern

The simplest pattern. Nodes represent entities, edges represent binary relations. Properties on nodes hold attributes; properties on edges hold relation metadata (confidence, source, timestamp).

(:Person {name, role}) -[:WORKS_AT {since}]-> (:Organization {name, type})
(:Drug {name, class})  -[:TREATS {efficacy}]-> (:Disease {name, icd_code})

Best for: domains with primarily binary relationships and moderate complexity.

Event Reification Pattern

Model N-ary relations and complex events as first-class nodes. An event node connects to all participants via typed role edges. This avoids information loss from forcing N-ary relations into binary edges.

(:ClinicalTrial {id, phase, start_date})
  -[:HAS_DRUG]->     (:Drug {name})
  -[:HAS_CONDITION]-> (:Disease {name})
  -[:HAS_OUTCOME]->   (:Outcome {measure, value})
  -[:CONDUCTED_BY]->  (:Organization {name})

Best for: events with multiple participants, clinical data, news events, financial transactions.

Layered Tier Pattern

Separate the graph into trust-differentiated layers that can be queried independently or together.

Layer 3 (Canonical Ontology): Formal class hierarchy, relation definitions, constraints
Layer 2 (Domain Knowledge):   Curated facts from literature, expert-validated
Layer 1 (Instance Data):      Extracted from user documents, case-specific, lower confidence

Cross-layer edges link instances to domain concepts and domain concepts to ontology classes. Provenance metadata on every edge records: source document, extraction method, confidence score, and timestamp.

Best for: multi-source integration, RAG with trust scoring, enterprise knowledge management.

Output Template

KNOWLEDGE GRAPH CONSTRUCTION SPECIFICATION
============================================

Domain: [Target domain and scope]
Use Case: [RAG / Reasoning / Analytics / Hybrid]
Data Sources: [List of input data types and volumes]

Data Model: [LPG / RDF / Hypergraph / Temporal]
Query Language: [Cypher / SPARQL / Gremlin / Custom]
Storage Backend: [Neo4j / Amazon Neptune / Virtuoso / etc.]

Schema Definition:
  Node Types:
  1. [EntityType] - [description]
     Properties: [list with types]
  2. [EntityType] - [description]
     Properties: [list with types]
  3. [Continue for each node type...]

  Edge Types:
  1. [RelationType] (source -> target) - [description]
     Properties: [list with types]
  2. [Continue for each edge type...]

  Constraints:
  - [Cardinality, uniqueness, required properties]

Extraction Pipeline:
  1. Entity Extraction
     - Method: [LLM-assisted / NER / Hybrid]
     - Prompt template: [summary or reference]
     - Verification: [Multi-round / Second-LLM / Manual sample]
  2. Relation Extraction
     - Method: [Prompt-based / Dependency parsing / Hybrid]
     - Few-shot examples: [count and source]
  3. Normalization
     - Deduplication: [method]
     - Ontology linking: [target ontology]
     - Synonym resolution: [approach]

Layered Architecture:
  Layer 1 (Instance): [description of instance-level data]
  Layer 2 (Domain):   [description of curated domain knowledge]
  Layer 3 (Ontology): [description of formal schema]
  Provenance: [How source/confidence/timestamp are tracked]

Validation Plan:
  - Schema conformance: [automated checks]
  - Coverage: [expected entity/relation counts]
  - Consistency: [contradiction detection method]
  - Human review: [sampling strategy]

Estimated Scale: [node count, edge count, properties per node]
Key Dependencies: [libraries, APIs, ontologies]

NEXT STEPS:
- Implement extraction pipeline on sample data
- Populate graph and run validation suite
- Iterate schema based on extraction results
- Integrate with downstream application (RAG, reasoning, etc.)