From Pilot to Production: Building a Machine-Learning-Driven Ontology for WSDOT’s Engineering Content.

About the Client

WSDOT publishes dozens of engineering reference manuals that state engineers rely on to plan and execute transportation infrastructure projects, spanning topics from hydraulics and environmental compliance to roadside design and highway maintenance. This engagement continued findability work Iknow had begun for WSDOT through an earlier, NCHRP-sponsored pilot project focused on a smaller set of stormwater-related manuals.

Industry Context

This project represents a natural evolution from that earlier pilot, moving from a focused proof of concept into a fuller-scale application spanning 18 manuals and incorporating more sophisticated machine learning techniques — unsupervised clustering and topic modeling — rather than relying primarily on manual analysis. The distinction between a taxonomy and an ontology matters here. A taxonomy captures simple hierarchical “kind of” or “is a” parent-child relationships, while an ontology can capture much richer relationship types, such as material composition, part-whole structure, and process or lifecycle relationships. This gives engineering users a genuinely more powerful way to understand and navigate how concepts relate to each other, not just where they sit in a hierarchy.

Current Situation

As part of a WSDOT project on improving the management and findability of transportation engineering content, Iknow developed a transportation ontology focused on stormwater removal from roadways. The stormwater ontology allowed the team to prove the concept that an ontology could classify and document relationships among concepts while also being used to drive auto-classification of engineering content.

• A much larger and more complex content corpus than the earlier pilot addressed. 18 separate engineering manuals, each independently authored and maintained, needed to be analyzed together to reveal how much genuine topical overlap existed between them — information WSDOT did not have going in.
• A simple taxonomy could not capture the real complexity of engineering relationships. Concepts like culverts involve material composition, end treatments, part-whole structure, and system membership simultaneously — relationships a basic parent-child hierarchy cannot represent, but that engineers actually need to search and reason with.
• No existing quantitative evidence of how much manual content actually overlapped. WSDOT needed objective, defensible evidence, not just an assumption, that consolidating classified content across the 18 manuals would genuinely help users before investing further in the approach.
• Preparing genuinely messy source content for rigorous machine learning analysis. Engineering manuals existed as PDFs, requiring conversion to text and extensive preprocessing before any meaningful text mining or clustering could begin.

• Develop a transportation ontology focused on stormwater removal from roadways.
• Enable search and navigation across WSDOT’s engineering manuals covering hydrology, erosion, roadway design, highway runoff, environmental concerns, cost estimation, roadside policy, highway maintenance, and utilities accommodation.
• Prove the concept that an ontology could both classify and document relationships among concepts and drive auto-classification of engineering content.

How Iknow Structured the Work

Iknow followed an eight-step process combining rigorous machine-learning-based content analysis, subject matter expert validation, and manual ontology construction, followed by ingestion into an autoclassification system.

Key Activities & Decisions

• Content Analysis. Iknow converted the manuals from PDF to text and performed preprocessing — tokenization, punctuation and stopword removal, and stemming — then represented the unstructured text numerically using bag-of-words and tf-idf approaches, weighting term frequency against how commonly a term appeared across all documents. Using Python’s Scikit-learn library, Iknow tested two unsupervised machine learning models: K-means clustering, which assigns each document to a single cluster and can be evaluated using a silhouette metric to measure cluster fit quality, and Latent Dirichlet Allocation topic modeling, which assigns each document a probability distribution across multiple topics rather than a single hard cluster assignment.
• Analysis of Results. The clustering results demonstrated significant overlap in the content presented across the 18 manuals analyzed, visualized in a cross-manual topic distribution chart — directly supporting the case that a consolidated, interactively classified body of engineering content would help users discover related material across manuals they might not otherwise think to search.
• Subject Matter Expert Interviews. Iknow interviewed subject matter experts to confirm its understanding of terms and synonyms, gather feedback on the machine-learning-derived term set, and add, correct, or remove ambiguous terms based on that expert input.
• Term Selection and Ontology Authoring. Iknow selected the final terms used to describe content and the relationships among terms, using WebProtégé as the ontology authoring tool.
• Ontology Development. Iknow manually built the ontology around seven related classes representing the data required by engineering users — Drainage Assets, Best Management Practices, Master Deliverables, Practical Solutions Life Cycle, Drainage System, Materials, and Project Development Topic Areas — capturing relationship types well beyond simple hierarchy, including material composition, end-treatment, part-whole, system-element, lifecycle work-step, and deliverable-to-topic relationships.
• Ingestion and Classification. Iknow ingested the completed ontology into the autoclassification system and classified the full transportation engineering content corpus based on it.

Stakeholders & Collaboration

Iknow worked as subcontractor to prime contractor Spy Pond Partners, collaborating directly with WSDOT subject matter experts to validate terminology and confirm the ontology’s relationships reflected real engineering practice.

Proving Content Overlap With Evidence, Not Assumption

Consolidating classified content across 18 independently maintained manuals was only worth pursuing if real topical overlap existed between them, and WSDOT needed proof before committing further. Iknow addressed this by applying rigorous, quantitative machine learning analysis — both K-means clustering and LDA topic modeling via Scikit-learn — across the full corpus, producing an objective cross-manual topic distribution analysis that directly demonstrated significant content overlap, turning a plausible hypothesis into a defensible, evidence-based finding before any ontology construction began.

Capturing Relationships a Standard Taxonomy Could Not Represent

Engineering concepts like culverts carry multiple simultaneous attributes — material, structure, function, system membership — that a flat parent-child taxonomy cannot represent. Iknow addressed this by building a true ontology rather than a taxonomy, defining seven distinct, purpose-built relationship types so that concepts could be represented with the full richness of their real engineering attributes, not just a single position in a hierarchy.

Operational Outcomes

• Analyzed content across 18 WSDOT engineering manuals using text mining and cluster analysis, testing two distinct unsupervised machine learning models.
• Built a seven-class ontology — Drainage Assets, Best Management Practices, Master Deliverables, Practical Solutions Life Cycle, Drainage System, Materials, and Project Development Topic Areas — defining seven distinct relationship types beyond simple parent-child hierarchy.
• Ingested the ontology into WSDOT’s autoclassification system and classified the full engineering content corpus.

Strategic and Organizational Outcomes

The cluster analysis demonstrated significant topical overlap in content across the 18 manuals, validating the case for a consolidated, interactively classified body of engineering content. Users found the resulting tags and relationships genuinely beneficial to their work, specifically when searching for information — proving the underlying concept from Iknow’s earlier NCHRP-sponsored pilot at full production scale, across a much larger and more technically sophisticated implementation.

Timeline to Impact

Iknow’s engagement spanned roughly 20 months, reflecting the technical depth of the work — from rigorous machine-learning-based corpus analysis through manual ontology construction to full production ingestion and classification — and representing a multi-year evolution of WSDOT’s findability program that began with the earlier NCHRP pilot.

Core Skills

• Ontology Design & Development
• Machine Learning-Based Content Analysis
• Autoclassification Implementation
• Taxonomy & Knowledge Modeling

Methods & Frameworks

• Text mining and cluster analysis (K-means, Latent Dirichlet Allocation)
• tf-idf and bag-of-words text representation
• SME-validated term development
• Relationship-rich ontology modeling

Technologies & Tools

• Python (Scikit-learn, NumPy, SciPy)
• WebProtégé (ontology authoring)
• Autoclassification systems