Enterprise Ontology KG Step 3: Multi-source Data Governance & Entity Alignment

Actual Operation Process

This step accounts for 30% of the project weight, second only to modeling. Data governance quality determines the lower bound of graph query and reasoning accuracy; even rigorous modeling cannot compensate for poor data quality.

Step 1: Data Source Analysis

Profile each data source to understand data volume, quality baseline, update frequency, and primary key systems.

Step 2: Data Cleaning

Perform deduplication, completion, format standardization, and anomaly handling.

Step 3: Schema Mapping

Document the mapping relationships from source fields to ontology attributes.

Step 4: Information Extraction

Structured data is mapped directly; unstructured documents (Word, Excel, PDF, PPT, system databases, regulations) are processed using LLM / NLP to extract entities and relationships.

Step 5: Entity Disambiguation and Alignment

Align cross-system IDs and normalize aliases.

Step 6: Triple Generation

Generate RDF triples conforming to the ontology schema.

Step 7: Manual Calibration

Human experts verify and correct extracted data.

Step 8: Data Quality Report Output

Produce a report covering alignment accuracy, coverage, and conflict rate.

Technical Selection and Key Points

ETL

can use DataX, Kettle, or custom Python pipelines. Unstructured extraction uses LLM APIs with few-shot examples and structured output constraints. Vector similarity employs Chinese embedding models such as BGE or M3E.

Before extraction, an evaluation benchmark set must be established: manually annotate 200–500 samples as ground truth. After each prompt adjustment, run evaluation to compare accuracy. Without a benchmark, extraction quality (70 vs 90) cannot be judged, prompt iteration effects cannot be measured, and go/no-go decisions lack evidence.

Entity Alignment: The Real Hard Battle

Adopt a hierarchical funnel strategy:

Rule layer first : ID mapping tables, coding rules — highest determinism.

Similarity algorithm layer : Edit distance, embedding vector similarity.

LLM-assisted judgment layer : Cases where semantics are highly ambiguous and rules/similarity are inconclusive.

Manual fallback layer : Handle remaining tail cases that previous layers cannot resolve.

Principle: use rules where possible, models where rules fail, humans only for the tail. This layered filtering keeps human cost controllable.

Three Baseline Requirements

Scriptable, rerunnable logic : All cleaning and transformation logic must be scripted and rerunnable; manual file edits are prohibited — non-reproducible processing equals no processing.

Provenance metadata on every triple : Each triple must carry source system, source document, extraction timestamp, and confidence. This is mandatory. Every alert and reasoning conclusion must answer "what is the basis, where did the data come from"; without provenance, graph outputs cannot earn business trust.

Calibration feedback loop : Business corrections to erroneous alignments must be codified into alignment rules or training samples, forming an "extract–verify–feedback" closed loop.

Deliverables

Upon completion, four artifacts are delivered:

Standard triple dataset with provenance metadata and confidence scores.

Schema mapping table .

Entity alignment rule library : alias cross-reference and ID mapping tables, constituting a long-term reusable data asset.

Data quality report : alignment accuracy, coverage, conflict rate.

The entity alignment rule library, together with the business glossary from Step 2, forms the reusable asset base for all future enterprise data governance work.