Guest Blog

While on her semantic modelling journey and as a Data Engineer herself, Ilaria Maresi encountered a range of challenges. There was not one definite source where she could quickly look things up, many of the resources were extremely technical and geared towards a more experienced audience while others were too wishy-washy. Therefore, she decided to compose this 50-page document where she explains semantic modelling and her most important lessons-learned – all in an engaging and down-to-earth writing style.

She starts off with the basics: what is a semantic model and why should you consider building one? Obviously, this is best explained by using a famous rock band as an example. In this way, you learn to draw the basic elements of a semantic model and some fun facts about Led Zeppelin at the same time!

For your model to actually work, it is essential that machines can also understand these fun facts. This might sound challenging if you are not a computer scientist but this guide will walk you through it  step-by-step – it even has pictures of baby animals! You will learn how to structure your model in Resource Description Framework (RDF) and give it meaning with the vocabulary extension that wins the prize for cutest acronym: Web Ontology Language (OWL).

All other important aspects of semantic modelling will be discussed. For example, how to make sure we all talk about the same Led Zeppelin by using Uniform Resource Identifiers (URIs). Moreover, you are not the first one thinking and learning about knowledge representation: many domain experts have spent serious time and effort in defining the major concepts of their field, called ontologies. To prevent you from re-inventing the wheel, we list the most important resources and explain their origin.

Are you a Data Engineer that has just started with semantic modelling? Want to refresh your memory? Maybe you have no experience with semantic modelling yet but feel it might come in handy? Well, this guide is for you!

Click here to access a data engineer’s guide to semantic modelling

Written by Tess Korthout

Phil Blackwood, Ph.D.

It’s no secret that most companies have silos of data and continue to create new silos.  Data that has the same meaning is often represented hundreds or thousands of different ways as new data models are introduced with every new software application, resulting in a high cost of integration.

By contrast, the data-centric approach starts with the common meaning of the data to address the root cause of data silos:

An enterprise is data-centric to the extent that all application functionality is based on a single, simple, extensible, federate-able data model.

An early step along the way to becoming data-centric is to establish a semantic model of the common concepts used across your business.  This might sound like a huge undertaking, and perhaps it will be if you start from scratch.  A better option is to adopt an existing core semantic model that has been designed for businesses and has a track record of success, such as gist.

Gist is an open source semantic model created by Semantic Arts.  It is the result of more than a decade of refinement based on data-centric projects done with major corporations in a variety of lines of business.  Semantic Arts describes gist as “… designed to have the maximum coverage of typical business ontology concepts with the fewest number of primitives and the least amount of ambiguity.”  The Wikipedia entry for upper ontologies compares gist to other ontologies, and gives a sense of why it is a match for corporate data management.

 

This blog post introduces gist by examining how some of the major Classes and Properties can be used.  We will not go into much detail; just enough to convey the general idea.

Everyone in your company would probably agree that running the business involves products, services, agreements, and events like payments and deliveries.  In turn, agreements and events involve “who, what, where, when, and why”, all of which are included in the gist model.  Gist includes about 150 Classes (types of things), and different parts of the business can be often be modeled by adding sub-classes.  Here are a few of the major Classes in gist:

Gist also includes about 100 standard ways things can be related to each other (Object Properties), such as:

• owns
• produces
• governs
• requires, prevents, or allows
• based on
• categorized by
• part of
• triggered by
• occurs at (some place)
• start time, end time
• has physical location
• has party (e.g. party to an agreement)

For example, the data representing a contract between a person and your company might include things like:

In gist, a Contract is a legally binding Agreement, and an Agreement is a Commitment involving two or more parties.  It’s clear and simple.  It’s also expressed in a way that is machine-readable to support automated inferences, Machine Learning, and Artificial Intelligence.

The items and relationships of the contract can be loaded into a knowledge graph, where each “thing” is a node and each relationship is an edge.  Existing data can be mapped to this standard representation to make it possible to view all of your contracts through a single lens of terminology.  The knowledge graph for an individual contract as sketched out above would look like:

Note that this example is just a starting point.  In practice, every node in the diagram would have additional properties (arrows out) providing more detail.  For example, the ID would link to a text string and to the party that allocated the ID (e.g. the state government that allocated a driver’s license ID).  The CatalogItem would be a detailed Product or Service Specification.

In the knowledge graph, there would be a single Person entry representing a given individual, and if two entries were later discovered to represent the same person, they could be linked with a sameAs relationship.

Relationships in gist (Properties) are first class citizens that have a meaning independent of the things they link, making them highly re-usable.  For example, identifiedBy is not limited to contracts, but can be used anywhere something has an ID.  Note that the Properties in gist are used to define relationships between instances rather than Classes; there are also a few standard relationships between Classes such as subClassOf and equivalentTo.

The categorizedBy relationship is a powerful one, because it allows the meaning of an item to be specified by linking to a taxonomy rather than by creating new Classes.  This pattern contributes to extensibility; adding new characteristics becomes comparable to adding valid values to a relational semantic model instead of adding new attributes.

Unlike traditional data models, the gist semantic model can be loaded into a knowledge graph and then the data is loaded into the same knowledge graph as an extension to the model.  There is no separation between the conceptual, logical, and physical models.  Similar queries can be used to discover the model or to view the data.

Gist uses the W3C OWL standard (Web Ontology Language), and you will need to understand OWL to get the most value out of gist.  To get started with OWL for corporate data management, check out the book Demystifying OWL for the Enterprise, by Michael Uschold.  There’s also a brief introduction to OWL and the way it uses set theory here.

The technology stack that supports OWL is well-established and has minimal vendor lock-in because of the simple standard data representation.  A semantic model created in one knowledge graph (triple store) can generally be ported to another tool without too much trouble.

To explore gist in more detail, you can download an ontology editor such as Protégé and then select File > Open From URL and enter: https://ontologies.semanticarts.com/o/gistCore9.4.0  Once you have the gist model loaded, select Entities and then review the descriptions of Classes, Object Properties (relationships between things), and Data Properties (which point to string or numeric values with no additional properties).  If you want to investigate gist in an orderly sequence, I’d suggest viewing items in groups of “who, what, when, where, and how.”

Take a look at gist.  It’s worth your time, because having a standard set f common terms like gist is a significant step toward reversing the trend toward more and more expensive data silos.

Click here to learn more about gist.

The Ontologist and Mathematician should be able to find common ground because Cantor introduced set theory into the foundation of mathematics, and W3C OWL uses set theory as a foundation for ontology language.  Let’s listen in as they mash up Cantor and OWL …

Ontologist: What would you like to talk about?

Mathematician: Anything.

Ontologist: Pick a thing. Any. Thing. You. Like.

Mathematician: [looks across the street]

Ontologist: Sure, why not?  Wells Fargo it is.  If we wanted to create an ontology for banking, we might need to have a concept of a company being a bank to differentiate it from other types of companies.  We would also want to generalize a bit and include the concept of Organization.

Mathematician: That’s simple in the world of sets.

Ontologist: In my world, every item in your diagram is related to every other item.  For example, Wells Fargo is not only a Bank, but it is also an Organization.  Relationships to “Thing” are applied automatically by my ontology editor.  When we build our ontology, we would first enter the relationships in the diagram below (read it from the bottom to the top):

Then we would run a reasoner to infer other relationships.  The result would look like this:

Mathematician: My picture has “Banks” and yours has “Bank”.  You took off the “s”.

Ontologist: Well, yes, I changed all the set names to make them singular because that’s the convention for Class names.  Sorry.  But now that you mention it … whenever I create a new Class I use a singular name just like everyone else does, but I also check to see if the plural is the good name for the set of things in the Class.  If the plural doesn’t sound like a set, I rethink it.  Try that with “Tom’s Stamp Collection” and see what you get.

Mathematician: I’d say you would have to rethink that Class name if you wanted the members of the Class to be stamps.  Otherwise, people using your model might not understand your intent.  Is a Class more like a set, or more like a template?

Ontologist: Definitely not a template, unlike object-oriented programming.  More like a set where the membership can change over time.

Mathematician: OK.  S or no S, I think we are mostly talking about the same thing.  In fact, your picture showing the Classes separated out instead of nested reminds me of what Georg Cantor said: “A set is a Many that allows itself to be thought of as a One.”

Ontologist: Yes.  You can think of a Class as a set of real world instances of a concept that is used to describe a subject like Banking.  Typically, we can re-use more general Classes and only need to create a subclass to differentiate its members from the other members of the existing Class (like Bank is a special kind of Company).  We create or re-use a Class when we want to give the Things in it meaning and context by relating them to other things.

Mathematician: Like this?

Ontologist: Exactly.  Now we know more about Joan, and we know more about Wells Fargo.  We call that a triple.

Mathematician: A triple.  How clever.

Ontologist: Actually, that’s the way we store all our data.  The triples form a knowledge graph.

Mathematician: Oh, now that’s interesting …  nice idea. Simple and elegant.  I think I like it.

Ontologist: Good.  Now back to your triple with Joan and Wells Fargo.  How would you generalize it in the world of sets?

Mathematician: Simple.  I call this next diagram a mapping, with Domain defined as the things I’m mapping from and Range defined as the things I’m mapping to.

Ontologist: I call worksFor an Object Property.  For today only, I’m going to shorten that to just “Property”.  But.  Wait, wait, wait.  Domain and Range?

In my world, I need to be careful about what I include in the Domain and Range, because any time I use worksFor, my reasoner will conclude that the thing on the left is in the Domain and the thing on the right is in the Range.

Ontologist continues: Imagine if I set the Domain to Person and the Range to Company, and then assert that Sparkplug the horse worksFor Tom the farmer.  The reasoner will tell me Sparkplug is a Person and Tom is a Company.  That’s why Domain and Range always raise a big CAUTION sign for me.  I always ask myself if there is anything else that might possibly be in the Domain or Range, ever, especially if the Property gets re-used by some else.  I need to define the Domain and Range broadly enough for future uses so I won’t end up trying to find the Social Security number of a horse.

Mathematician: Bummer.  Good luck with that.

Ontologist: Oh, thank you.  Now back your “mapping”.  I suppose you think of it as a set of arrows and you can have subsets of them.

Mathematician: Yes, pretty much.  If I wanted to be more precise, I would say a mapping is a set of ordered pairs.  I’m going to use an arrow to show what order the things are in; and voila, here is your set diagram for the concept:

You will notice that there are two different relationships:

The pair (Joan, Wells Fargo) is in both sets, so it is in both mappings.  Does that make sense to you?

Ontologist: Yes, I think it makes sense.  In my world, if I cared about both of these types of relationships, I would make isAManagerAt a subProperty of worksFor, and enter the assertion that Joan is a manager at Wells Fargo.  My reasoner would add the inferred relationship that Joan worksFor Wells Fargo.

Mathematician: Of course!  I think I’ve got the basic idea now.  Let me show you what else I can do with sets.  I’ll even throw in some your terminology.

Ontologist: Oh, by all means. [O is silently thinking, “I bet this is all in OWL, but hey, the OWL specs don’t have pictures of sets.”]

Mathematician: [takes a deep breath so he can go on and on … ]

Let’s start with two sets:

The intersection is a subset of each set, and each of the sets is a subset of the union.  If we want to use the intersection as a Class, we should be able to infer:

And if we want to use the union as a Class, then each original Class is a Sub Class of the union:

If two Classes A and B have no members in common (disjoint), then every Sub Class of A is disjoint from every sub class of B:

A mapping where there is at most one arrow out from each starting point is called a function.

A mapping where there is at most one arrow into each ending point is called inverse-functional.

You get the inverse of a mapping by reversing the direction of all the arrows in it.  As the name implies, if a mapping is inverse-functional, it means the inverse is a function.

Sometimes the inverse mapping ends up looking just like the original (called symmetric), and sometimes it is “totally different” (disjoint or asymmetric).

Sometimes a mapping is transitive, like our diagram of inferences with subClassOf, where a subclass of a subclass is a subclass.  I don’t have a nice simple set diagram for that, but our Class diagram is an easy way to visualize it.  Take two hops using the same relationship and you get another instance of the relationship:

Sets can be defined by combining other sets and mappings, such as the set of all people who work for some bank (any bank).

Ontologist: Not bad.  Here’s what I would add:

Sometimes I define a set by a phrase like you mentioned (worksFor some Bank), and in OWL I can plug that phrase into any expression where a Class name would make sense.  If I want to turn the set into a named Class, I can say the Class is equivalent to the phrase that defines it.  Like this:

BankEmployee is equivalentTo (worksFor some Bank).

The reasoner can often use the phrase to infer things into the Class BankEmployee, or use membership in the Class to infer the conditions in the phrase are true.  A lot of meaning can be added to data this way.  Just as in a dictionary, we define things in terms of other things.

When two Classes are disjoint, it means they have very distinct and separate meanings.  It’s a really good thing, especially at more general levels.  When we record disjointness in the ontology, the reasoner can use it to detect errors.

Whenever I create a Property, I always check to see if it is a function.  If so, I record the fact that it is a function in the ontology because it sharpens the meaning.

We never really talked about Data Properties.  Maybe next time.  They’re for simple attributes like “the building is 5 stories tall”.

A lot of times, a high level Property can be used instead of creating a new subProperty.  Whenever I consider creating a new subProperty, I ask myself if my triples will be just as meaningful if I use the original Property.  A lot of times, the answer is yes and I can keep my model simple by not creating a new Property.

An ontology is defined in terms of sets of things in the real world, but our data base usually does not have a complete set of records for everything defined in the ontology.  So, we should not try to infer too much from the data that is present.  That kind of logic is built in to reasoners.

On the flip side, the data can include multiple instances for the same thing, especially when we are linking multiple data sets together.  We can use the sameAs Property to link records that refer to the same real-world thing, or even to link together independently-created graphs.

The OWL ontology language is explained well at: https://www.w3.org/TR/owl-primer/

However, even if we understand the theory, there are many choices to be made when creating an ontology.  If you are creating an ontology for a business, a great book that covers the practical aspects is Demystifying OWL for the Enterprise by Michael Uschold.

Mathematician: I want the last word.

Ontologist: OK.

Mathematician:

Ontologist: I agree, but that wasn’t a word.  🙂

Mathematician: OK.  I think I’m starting to see what you are doing with ontologies.  Here’s what it looks like to me: since it is based on set logic and triples, the OWL ontology language has a rock-solid foundation.

Written By: Phil Blackwood, Ph.D.

‘The Graceful Adaption of St Frances Xavier Cabrini Hospital since 1948.

---

This post has been updated to reflect the current corona virus crisis. Joe Pine and Kim Korn authors of ‘Infinite Possibility: Creating Customer Value on the Digital Frontier’ say the coronacrisis will change healthcare for the better. Kim points out that although it is 20 years late. It is good to see healthcare dipping a toe in the 21^st Century. However, I think, we have a long way to go.

Dave McComb in his book, ‘The Data-Centric Revolution: Restoring Sanity to Enterprise Information Systems’, suggests that buildings change more gracefully than software does. He uses Stewart Brand’s model which shows how buildings can change after they are built.

Graceful Change

I experienced the graceful change during the 19 years that I worked at Cabrini Hospital Malvern. The buildings changed whilst still serving its customers. To outsiders, it may have appeared seamless. To insiders charged with changing the environment, it took constant planning and endless negotiation with internal stakeholders and the surrounding community.

Geoff Fazakerley, the director of buildings, engineering, and diagnostic services orchestrated most of the activities following the guiding principles of the late Sister Irma Lunghi MSC, “In all that we must first keep our eyes on our patients’ needs.”

Geoff, took endless rounds with his team to assess where space could be allocated for temporary offices and medical suites. He then negotiated with those that were impacted to move to temporary locations. On several instances space for non-patient services had to be found outside the hospital so that patients would not be inconvenienced. The building as it now stands bears testament to how Geoff’s team and the architects managed the graceful change.

Why software change is not as graceful

Most enterprise software changes are difficult. In ‘Software Wasteland: How the Application-Centric Mindset is Hobbling our Enterprises’, McComb explains that by focusing on the application as the foundation or starting point instead of taking a data-centric approach this hobbles agile adoption, extension and innovation. When a program is loaded first, it creates a new data structure. With this approach, each business problem requires yet another application system. Each application creates and manages yet another data model. Over time this approach leads to many more applications and many additional complex data models. Rather than improving access to information this approach steadily erodes it.

McComb says that every year the amount of data available to enterprises doubles, while the ability to effectively use it decreases. Executives lament their legacy systems, but their projects to replace them rarely succeed. This is not a technology problem, it is more about mindset.

From personal experience, we know that some buildings adapt well and some don’t. Those of us that work in large organisations also know that this is the same with enterprise software. However, one difference between buildings and software is important. Buildings are physical. They are, to use a phrase, ‘set in stone’. They are situated on a physical site. You either have sufficient space or you need to acquire more. You tend to know these things even before adaption and extension begin. With software it is different, there are infinite possibilities.

Software is different

The boundaries cannot be seen. Software is made of bits, not atoms. Hence, we experience software differently. As Joe Pine and Kim Korn explain in their book, ‘Infinite Possibility: Creating Customer Value on the Digital Frontier’, software exists beyond the physical limitation of time, space, and matter. With no boundaries, the software provides the opportunity for infinite possibilities. But as James Gilmore says in the foreword of the book, most enterprises treat digital technology as an incremental adjunct to their existing processes. As a result, the experience is far from enriching. Instead of making the real world experience better, software often worsens it. In hospitals, software forces clinicians to take their eyes off the patient to input data into screens and to focus on the content of their screens.

More generally, there appears to be a gap between what the end-users of enterprise software expect, the champions of the software, and the expectations of software vendors themselves. The blog post by Virtual Stacks makes our thinking about software sound like a war of expectations.

The war of expectations

People that sell software, the people that buy software, and the people that eventually use the software view things very differently:

1. Executives in the C-Suite implement an ERP System to gain operational excellence and cost savings. They often put someone in charge of the project that doesn’t know enough about ERP systems and how to manage the changes that the software demands in work practice.
2. Buyers of an ERP system expect that it will fulfil their needs straight out-of-the-box. Sellers expect some local re-configuration.
3. Reconfiguring enterprise software calls for a flexible budget. The budget must provide for consultants that may have to be called in. It needs to provide for additional training and more likely than not major change management initiatives.
4. End-users have to be provided with training even before the software is launched. This is especially necessary when they have to learn new skills that are not related to their primary tasks. In hospitals, clinicians find that their workload blows out. They see themselves as working for the software rather than the other way around.

The organisational mindset

Shaun Snapp in his blogpost for Brightwork Research points out that what keeps the application-centric paradigm alive is how IT vendor and services are configured and incentivised. There is a tremendous amount of money to be made when building, implementing and integrating applications in organisations. Or as McComb says, ‘the real problem surrounds business culture’.

Can enterprise software adapt well?

The short answer is yes. However, McComb’s key point is not that some software adapts well and some don’t, it is that legacy software doesn’t. Or as the quotation in Snapp’s post suggests:

‘The zero-legacy startups have a 100:1 cost and flexibility advantage over their established rivals. Witness the speed and agility of Pinterest, Instagram, Facebook and Google. What do they have that their more established competitors don’t? It’s more instructive to ask what they don’t have: They don’t have their information fractured into thousands of silos that must be continually integrated” at great cost.’

Snapp goes on to say that ‘in large enterprises, simple changes that any competent developer could make in a week typically take months to implement. Often the change requests get relegated to the “shadow backlog” where they are ignored until the requesting department does the one thing that is guaranteed to make the situation worse: launch another application project.’

Adapting well

Werner Vogels, VP &CTO of Amazon provides a good example of successful change. Use this link to read the full post.

Perhaps as Joe Pine and Kim Korn say the coronacrisis will change healthcare for the better. I have written about using a data-centric model to query hospital ERP systems to track and trace materials. My eBook, ‘Hidden Hospital Hazards: Saving Lives and Improving Margins’ can be purchased from Amazon, or you may obtain a free PDF version here.

WRITTEN BY
Len Kennedy
Author of ‘Hidden Hospital Hazards’