Andy FitzgeraldInformation Architect

Book Note

Ontology Pipeline

A "semantic knowledge management system" structures concepts, their definitions, and their relationships so that people and machines can interpret shared knowledge consistently and act on it confidently. Such systems are a crucial component for enterprise knowledge management, and are an increasingly important complement to generative AI integrated platforms and products.

Jessica Talisman's Ontology Pipeline offers a detailed framework for scoping, designing, building, and maintaining these systems. Talisman charts a systematic, incremental path through the creation of controlled vocabularies, metadata schema, taxonomy, thesauri, ontologies, and knowledge graphs, noting at each step how to ensure quality, deliver incremental value, and shape the work to fit a range of organization shapes and sizes.

Highlights

  • A "controlled vocabulary is a context clue framework" that can "guide LLMs through a quagmire of text, code, and infrastructure workflows" (32).
  • Open standards are paramount: "Every term you borrow from an existing standard (SKOS, PROV-O, FOAF, DCAT, Dublin Core, etc.) is a bridge to every other system that uses that standard. [...] A well-designed domain ontology is mostly assembly" (196).
  • "The current AI landscape is dominated by neural networks, which excel at pattern recognition, but operate as black boxes. [...] While the neural network perceives, the ontology reasons. [...] ontologies shape the formal knowledge that AI systems require to move from pattern matching to genuine understanding" (180).
  • "If you cannot write a clear definition, you do not yet understand the concept well enough to model it" (200).
  • "Ontologies are build to be used, and the highest value use is as a live, queryable source of organizational truth that operational systems can depend on" (249).
  • "Knowledge graph programs fail when they attempt to deliver a production architecture in a single phase." The importance of phases, however, isn't only to break up the work into manageable chunks: it's to create the feedback loops that enable you to validate design decisions against read data and queries and to integrate that feedback "as early and as cheaply as possible" (309).
  • A knowledge graph architecture is not a product, it is a "practice, an ongoing discipline of aligning the semantic model with the domain it represents, the data it governs, and the systems it serves" (319).

Impressions

Ontology Pipeline is a practical handbook for designing and building knowledge management infrastructure. Talisman provides some history and background to the library science and semantic web concepts and approaches behind the steps, but this book is predominately focused on building systems to establish and manage shared meaning.

Though Ontology Pipeline moves through six ordered stages, beginning with establishing shared vocabularies and ending up at a richly connected network of real world facts and data, this is not a book about moonshots. Each stage includes insight into how to deliver incremental value, building on work done in the previous step. An organization that never moves past shared vocabularies and governed metadata standards (stages one and two) will likely be significantly better off for the effort — especially if where it started involved free-tagging into ad-hoc attribute fields.

For organizations that do eventually move into defining and integration an ontology to create a knowledge graph, the payoff is not limited to inferential reasoning alone. Throughout Ontology Pipeline Talisman articulates the complementary relationship between symbolic AI (driven by logic-based ontology) and generative AI. Put together strategically, these two approaches create what Talisman (among others) refers to as neurosymbolic AI: "The LLM's contribution is natural language interpretation and generation. The ontology's contribution is formal correctness" (245).

Follow along with RSS

Want to get updates when I publish content like this in the future? Subscribe via RSS to get updates directly in your feed reader.

Subscribe via RSS