Monday, November 24, 2014

Multi-Entity Models.... Baker, Coyle, Petiya

Multi-Entity Models of Resource Description in the Semantic Web: A comparison of FRBR, RDA, and BIBFRAME
by Tom Baker, Karen Coyle, Sean Petiya
Published in: Library Hi Tech, v. 32, n. 4, 2014 pp 562-582 DOI:10.1108/LHT-08-2014-0081
Open Access Preprint

The above article was just published in Library hi Tech. However, because the article is a bit dense, as journal articles tend to be, here is a short description of the topic covered, plus a chance to reply to the article.

We now have a number of multi-level views of bibliographic data. There is the traditional "unit card" view, reflected in MARC, that treats all bibliographic data as a single unit. There is the FRBR four-level model that describes a single "real" item, and three levels of abstraction: manifestation, expression, and work. This is also the view taken by RDA, although employing a different set of properties to define instances of the FRBR classes. Then there is the BIBFRAME model, which has two bibliographic levels, work and instance, with the physical item as an annotation on the instance.

In support of these views we have three RDF-based vocabularies:

FRBRer (using OWL)
RDA (using RDFS)

The vocabularies use a varying degree of specification. FRBRer is the most detailed and strict, using OWL to define cardinality, domains and ranges, and disjointness between classes and between properties. There are, however, no sub-classes or sub-properties. BIBFRAME properties all are defined in terms of domains (classes), and there are some sub-class and sub-property relationships. RDA has a single set of classes that are derived from the FRBR entities, and each property has the domain of a single class. RDA also has a parallel vocabulary that defines no class relationships; thus, no properties in that vocabulary result in a class entailment. [1]

As I talked about in the previous blog post on classes, the meaning of classes in RDF is often misunderstood, and that is just the beginning of the confusion that surrounds these new technologies. Recently, Bernard Vatant, who is a creator of the Linked Open Vocabularies site that does a statistical analysis of the existing linked open data vocabularies and how they relate to each other, said this on the LOV Google+ group:
" seems that many vocabularies in LOV are either built or used (or both) as constraint and validation vocabularies in closed worlds. Which means often in radical contradiction with their declared semantics."
What Vatant is saying here is that many vocabularies that he observes use RDF in the "wrong way." One of the common "wrong ways" is to interpret the axioms that you can define in RDFS or OWL the same way you would interpret them in, say, XSD, or in a relational database design. In fact, the action of the OWL rules (originally called "constraints," which seems to have contributed to the confusion, now called "axioms") can be entirely counter-intuitive to anyone whose view of data is not formed by something called "description logic (DL)."

A simple demonstration of this, which we use in the article, is the OWL axiom for "maximum cardinality." In a non-DL programming world, you often state that a certain element in your data is limited to the number of times it can be used, such as saying that in a MARC record you can have only one 100 (main author) field. The maximum cardinality of that field is therefore "1". In your non-DL environment, a data creation application will not let you create more than one 100 field; if an application receiving data encounters a record with more than one 100 field, it will signal an error.

The semantic web, in its DL mode, draws an entirely different conclusion. The semantic web has two key principles: open world, and non-unique name. Open world means that whatever the state of the data on the web today, it may be incomplete; there can be unknowns. Therefore, you may say that you MUST have a title for every book, but if a look at your data reveals a book without a title, then your book still has a title, it is just an unknown title. That's pretty startling, but what about that 100 field? You've said that there can only be one, so what happens if there are 2 or 3 or more of them for a book? That's no problem, says OWL: the rule is that there is only one, but the non-unique name rule says that for any "thing" there can be more than one name for it. So when an OWL program [2] encounters multiple author 100 fields, it concludes that these are all different names for the same one thing, as defined by the combination of the non-unique name assumption and the maximum cardinality rule: "There can only be one, so these three must really be different names for that one." It's a bit like Alice in Wonderland, but there's science behind it.

What you have in your database today is a closed world, where you define what is right and wrong; where you can enforce the rule that required elements absolutely HAVE TO be there; where the forbidden is not allowed to happen. The semantic web standards are designed for the open world of the web where no one has that kind of control. Think of it this way: what if you put a document onto the open web for anyone to read, but wanted to prevent anyone from linking to it? You can't. The links that others create are beyond your control. The semantic web was developed around the idea of a web (aka a giant graph) of data. You can put your data up there or not, but once it's there it is subject to the open functionality of the web. And the standards of RDFS and OWL, which are the current standards that one uses to define semantic web data, are designed specifically for that rather chaotic information ecosystem, where, as the third main principle of the semantic web states, "anyone can say anything about anything."

I have a lot of thoughts about this conflict between the open world of the semantic web and the needs for closed world controls over data; in particular whether it really makes sense to use the same technology for both, since there is such a strong incompatibility in underlying logic of these two premises. As Vatant implies, many people creating RDF data are doing so with their minds firmly set in closed world rules, such that the actual result of applying the axioms of OWL and RDF on this data on the open web will not yield the expected closed world results.

This is what Baker, Petiya and I address in our paper, as we create examples from FRBRer, RDA in RDF, and BIBFRAME. Some of the results there will probably surprise you. If you doubt our conclusions, visit the site that gives more information about the tests, the data and the test results.

[1] "Entailment" means that the property does not carry with it any "classness" that would thus indicate that the resource is an instance of that class.

[2] Programs that interpret the OWL axioms are called "reasoners". There are a number of different reasoner programs available that you can call from your software, such as Pellet, Hermit, and others built into software packages like TopBraid.

Tuesday, November 18, 2014

Classes in RDF

RDF allows one to define class relationships for things and concepts. The RDFS1.1 primer describes classes succinctly as:
Resources may be divided into groups called classes. The members of a class are known as instances of the class. Classes are themselves resources. They are often identified by IRIs and may be described using RDF properties. The rdf:type property may be used to state that a resource is an instance of a class.
This seems simple, but it is in fact one of the primary areas of confusion about RDF.

If you are not a programmer, you probably think of classes in terms of taxonomies -- genus, species, sub-species, etc. If you are a librarian you might think of classes in terms of classification, like Library of Congress or the Dewey Decimal System. In these, the class defines certain characteristics of the members of the class. Thus, with two classes, Pets and Veterinary science, you can have:
- dogs
- cats

Veterinary science
- dogs
- cats
In each of those, dogs and cats have different meaning because the class provides a context: either as pets, or information about them as treated in veterinary science.

For those familiar with XML, it has similar functionality because it makes use of nesting of data elements. In XML you can create something like this:
and it is clear which price goes with which type of drink, and that the bits directly under the <drink> level are all drinks, because that's what <drink> tells you.

Now you have to forget all of this in order to understand RDF, because RDF classes do not work like this at all. In RDF, the "classness" is not expressed hierarchically, with a class defining the elements that are subordinate to it. Instead it works in the opposite way: the descriptive elements in RDF (called "properties") are the ones that define the class of the thing being described. Properties carry the class information through a characteristic called the "domain" of the property. The domain of the property is a class, and when you use that property to describe something, you are saying that the "something" is an instance of that class. It's like building the taxonomy from the bottom up.

This only makes sense through examples. Here are a few:
1. "has child" is of domain "Parent".

If I say "X - has child - 'Fred'" then I have also said that X is a Parent because every thing that has a child is a Parent.

2. "has Worktitle" is of domain "Work"

If I say "Y - has Worktitle - 'Der Zauberberg'" then I have also said that Y is a Work because every thing that has a Worktitle is a Work.

In essence, X or Y is an identifier for something that is of unknown characteristics until it is described. What you say about X or Y is what defines it, and the classes put it in context. This may seem odd, but if you think of it in terms of descriptive metadata, your metadata describes the "thing in hand"; the "thing in hand" doesn't describe your metadata. 

Like in real life, any "thing" can have more than one context and therefore more than one class. X, the Parent, can also be an Employee (in the context of her work), a Driver (to the Department of Motor Vehicles), a Patient (to her doctor's office). The same identified entity can be an instance of any number of classes.
"has child" has domain "Parent"
"has licence" has domain "Driver"
"has doctor" has domain "Patient"

X - has child - "Fred"  = X is a Parent 
X - has license - "234566"  = X is a Driver
X - has doctor - URI:765876 = X is a Patient
Classes are defined in your RDF vocabulary, as as the domains of properties. The above statements require an application to look at the definition of the property in the vocabulary to determine whether it has a domain, and then to treat the subject, X, as an instance of the class described as the domain of the property. There is another way to provide the class as context in RDF - you can declare it explicitly in your instance data, rather than, or in addition to, having the class characteristics inherent in your descriptive properties when you create your metadata. The term used for this, based on the RDF standard, is "type," in that you are assigning a type to the "thing." For example, you could say:
X - is type - Parent
X - has child - "Fred"
This can be the same class as you would discern from the properties, or it could be an additional class. It is often used to simplify the programming needs of those working in RDF because it means the program does not have to query the vocabulary to determine the class of X. You see this, for example, in BIBFRAME data. The second line in this example gives two classes for this entity:
a bf:Instance, bf:Monograph .

One thing that classes do not do, however, is to prevent your "thing" from being assigned the "wrong class." You can, however, define your vocabulary to make "wrong classes" apparent. To do this you define certain classes as disjoint, for example a class of "dead" would logically be disjoint from a class of "alive." Disjoint means that the same thing cannot be of both classes, either through the direct declaration of "type" or through the assignment of properties. Let's do an example:
"residence" has domain "Alive"
"cemetery plot location" has domain "Dead"
"Alive" is disjoint "Dead" (you can't be both alive and dead)

X - is type - "Alive"                                         (X is of class "Alive")
X - cemetery plot location - URI:9494747      (X is of class "Dead")
Nothing stops you from creating this contradiction, but some applications that try to use the data will be stumped because you've created something that, in RDF-speak, is logically inconsistent. What happens next is determined by how your application has been programmed to deal with such things. In some cases, the inconsistency will mean that you cannot fulfill the task the application was attempting. If you reach a decision point where "if Alive do A, if Dead do B" then your application may be stumped and unable to go on.

All of this is to be kept in mind for the next blog post, which talks about the effect of class definitions on bibliographic data in RDF.

Note: Multiple domains are treated in RDF as an AND (an intersection). Using a library-ish example, let's assume you want to define a note field that you can use for any of your bibliographic entities. For this example, we'll define entities Work, Person, Manifestation for ex:note. You define your note property something like:

    a rdf:Property ;
    rdfs:label "Note"@en ;
    rdfs:domain ex:Work ;
    rdfs:domain ex:Person ;
    rdfs:domain ex:Manifestation ;
    rdfs:range rdfs:literal .

Any subject on which you use ex:note would be inferred to be, at the same time, Work and Person and Manifestation - which is manifestly illogical. There is not a way to express the rule: "This property CAN be used with these classes" in RDF. For that, you will need something that does not yet exist in RDF, but is being worked on in the W3C community, which is a set of rules that would allow you to validate property usage. You might also want see what has done for domain and range.