An Exercise in Irrelevance

Abstract

The Tawny-OWL library provides a fully-programmatic environment for ontology building; it enables the use of a rich set of tools for ontology development, by recasting development as a form of programming. It is built in Clojure - a modern Lisp dialect, and is backed by the OWL API. Used simply, it has a similar syntax to OWL Manchester syntax, but it provides arbitrary extensibility and abstraction. It builds on existing facilities for Clojure, which provides a rich and modern programming tool chain, for versioning, distributed development, build, testing and continuous integration. In this paper, we describe the library, this environment and the its potential implications for the ontology development process.

• Phillip Lord

• http://arxiv.org/abs/1303.0213

Plain English Summary

In this paper, I describe some new software, called Tawny-OWL, that addresses the issue of building ontologies. An ontology is a formal hierarchy, which can be used to describe different parts of the world, including biology which is my main interest.

Building ontologies in any form is hard, but many ontologies are repetitive, having many similar terms. Current ontology building tools tend to require a significant amount of manual intervention. Rather than look to creating new tools, Tawny-OWL is a library written in full programming language, which helps to redefine the problem of ontology building to one of programming. Instead of building new ontology tools, the hope is that Tawny-OWL will enable ontology builders to just use existing tools that are designed for general purpose programming. As there are many more people involved in general programming, many tools already exist and are very advanced.

This is the first paper on the topic, although it has been discussed before here.

This paper was written for the OWLED workshop in 2013.

Reviews

Reviews are posted here with the kind permission of the reviewers. Reviewers are identified or remain anonymous at their option. Copyright of the review remains with the reviewer and is not subject to the overall blog license.

Review 1

The given paper is a solid presentation of a system for supporting the development of ontologies – and therefore not really a scientific/research paper.

It describes Tawny OWL in a sufficiently comprehensive and detailed fashion to understand both the rationale behind as well as the functioning of that system. The text itself is well written and also well structured. Further, the combination of the descriptive text in conjunction with the given (code) examples make the different functionality highlights of Tawny OWL very easy to grasp and appraise.

As another big plus of this paper, I see the availability of all source code which supports the fact that the system is indeed actually available – instead of being just another description of a “hidden” research system.

The possibility to integrate Tawny OWL in a common (programming) environment, the abstraction level support, the modularity and the testing “framework” along with its straightforward syntax make it indeed very appealing and sophisticated.

But the just said comes with a little warning: My above judgment (especially the last comment) are highly biased by the fact that I am also a software developer. And thus I do not know how much the above would apply to non-programmers as well.

And along with the above warning, I actually see a (more global) problem with the proposed approach to ontology development: The mentioned “waterfall methodologies” are still most often used for creating ontologies (at least in the field of biomedical ontologies) and thus I wonder how much programmatic approaches, as implemented by Tawny OWL, will be adapted in the future. Or in which way they might get somehow integrated in those methodologies.

When I started work on Clojure-owl the original intention was to provide myself with a more programmatic environment for writing ontologies, where I could work with a full programming language at to define the classes I wanted (http://www.russet.org.uk/blog/2214). After some initial work with functions taking strings, I have moved to an approach where classes (and other ontological entities), are each assigned to a Lisp symbol (http://www.russet.org.uk/blog/2254). I’m using “symbol” rather than “atom” because its a bit more accurate, especially as Clojure uses “atom” with a different meaning.

This means that I now have something which allows me to write ontological terms looking something like this:

(defclass a)
(defclass b :subclass a)

(defoproperty r)
(defclass d
     :subclass (some r b))

While this is quite nice, and looks fairly close to Manchester syntax (http://www.w3.org/TR/owl2-manchester-syntax/), ultimately, so far all this really provides me with is a slightly complex mechanism for achieving what I could already do; which raises the questions, why not just use Manchester syntax? Why bother with the Lisp if this is all I am to achieve?

I think I have now got to the point where the advantages are starting to show through, as I have started to create useful macros, which operate at a slightly higher level of abstraction from Manchester syntax. I will explain this using examples, perhaps inevitably, based around pizza (http://robertdavidstevens.wordpress.com/2010/01/22/why-the-pizza-ontology-tutorial/), which I have started to develop using Clojure-owl.

First I wanted to be able to define several classes at once, rather than having to use a somewhat long-winded defclass form for each; for this I have written a macro called declare-classes — perhaps a slight misnomer, as it also adds the classes to the ontology. This example shows the purpose:

  (declare-classes
   GoatsCheeseTopping
   GorgonzolaTopping
   MozzarellaTopping
   ParmesanTopping)

In practice, this may not be that useful for an ontology builder, as it creates a bare class; no documentation, nothing else. It may be useful for forward-declaration (like Clojure declare).

One slightly unfortunate consequence of the decision to use lisp symbols is I know find myself writing a lot of macros. For those who have not used lisp before, most work is done with functions. Macros are only necessary when you wish to extend the language itself. They tend to be more complex to write and to debug, although fortunately are easy to use. Compare, for example, the definition of declare-classes to that of the functional equivalent which uses strings.

(defmacro declare-classes
  [& names]
  `(do ~@(map
          (fn [x#]
            `(defclass ~x#))
          names)))

(defun f-declare-classes
  [& names]
  (dorun
   (map #(owlclass x) names)))

Even in this case, there is more hieroglyphics in the macro — two backticks, one unquote splice and some gensym symbols although Clojure’s slightly irritating lazy sequences and the resultant dorun mean that the two are nearly as long as each other. I suspect that the macros are going to get more complex, however. In most cases, should not be the user of the library that has to cope though.

While this provided a useful convenience, I also wanted a cleaner method for declaring disjoints. Consider this example:

(defclass a)
(defclass b)
(defclass c)

(disjointclasses a b c)

This is reasonably effective, but a pain if there are many classes, as they all need to be listed in the disjointclasses list. Worse, this is error prone; it is all too easy to miss a single class out, particularly if a new classes is added. So, I have now implemented an as-disjoint macro which gives this code:

(as-disjoints
   (defclass a)
   (defclass b)
   (defclass c))

This should avoid both the risk of dropping a disjoint, as well avoiding the duplication. An even more common from is to wish to declare a set of classes as disjoint children. Again, I have provided a macro for this, which looks like this:

 (defclass CheeseTopping)

 (as-disjoint-subclasses
  CheeseTopping

  (declare-classes
   GoatsCheeseTopping
   GorgonzolaTopping
   MozzarellaTopping
   ParmesanTopping))

Although this was not my original intention, these are actually nestable. This gives the interesting side effect that the ontology hierarchy is now represented in the structure of the lisp. Example below is an elided hierarchy from pizza. Lisp programmers will notice I have rather exaggerated the indentation to make the point.

(as-disjoint-subclasses
 PizzaTopping

    (defclass CheeseTopping)

    (as-disjoint-subclasses
         CheeseTopping

        (declare-classes
            GoatsCheeseTopping))

    (defclass FishTopping)
    (as-disjoint-subclasses
        FishTopping

        (declare-classes AnchoviesTopping))

    (defclass FruitTopping)
    (as-disjoint-subclasses
         FruitTopping

         (declare-classes PineappleTopping)))

Of course, it is not essential to do this. The nested use of as-disjoint-subclasses confers no semantics; but it does allow juxtaposition of a class and it’s children.

Being able to build up macros in this way was the main reason I wanted a real programming language; those described here are, I think, fairly general purpose; so, this form of declaration could also be supported in any of the various syntaxes, although it would require update to the tools. However, some ontologies will benefit from less general purpose extensions. These are never going to be supported in syntax specification.

Still, it is not all advantage. Using a programming language means embedding within this language. And this means that some of names I would like to use are gone; http://clojuredocs.org/clojure_core/clojure.core/some [some] is the obvious example. While Clojure has good namespace support, functions in clojure.core are available in all other namespaces; like all lisps, Clojure lacks types which would have avoided the problem. There are other ways around this, but ultimately clashing with these names is likely to bring pain; for example, I could always explicitly reference clojure-owl functions; but writing owl.owl.defclass rather than defclass seems a poor option; hence, some has become owlsome, and comment has become owlcomment. I have decided to accept the lack of consistency and kept only and label; the alternative, taken by the OWL API to appending OWL to everything seems too unwieldy.

Bibliography

With my initial work on developing a Clojure environment for OWL (http://www.russet.org.uk/blog/2214), I was focused on producing something similar to Manchester syntax (http://www.w3.org/TR/owl2-manchester-syntax/). Here, I describe my latest extensions which makes more extensive use of Lisp atoms. The practical upshot of this should be to reduce errors due to spelling mistakes, as well as enabling me to add simple checks for correctness.

The desire for a simple syntax is an important one. I would like my library to be usable by people not experienced with Lisp, although I am clearly aware that this sort of environment is likely to be aimed at those with some programming skills. I have managed to produce a syntax which, I think, is reasonable straight forward. It has more parentheses than Manchester syntax, but is easier in other ways, especially now that I have learnt a little more about how Clojure namespaces work. For example, this defines a class in OWL.

(owlclass "HumanArm"
          :subclass "Arm" (some "isPartOf" "Human")
          :annotation (comment "The Human arm is an Arm which is part of a human"))

One of my initial desires for the Clojure mode was to enable the use of standard tools that we have come to expect from a modern programming language, which should enable us to build a more pragmatic ontology building methodology (http://robertdavidstevens.wordpress.com/2011/05/26/unicorns-in-my-ontology/). The first of these is a unit testing environment. Clojure already has one of these integrated. So far, I have only used this for testing my own code; so, for example, this is the current unit test for the owlclass function used above.

(deftest owlclass
  (is (= 1
         (do (o/owlclass "test")
             (.size (.getClassesInSignature
                     (#'o/get-current-jontology))))))
  (is (instance? org.semanticweb.owlapi.model.OWLClass
                 (o/owlclass "test"))))

There are, however, some limitations to the approach that I have taken so far. Consider this statement:

(owlclass "HumanArm"
          :subclass (some "isPartOf" "Humn") "Arm"
)

This is broken because I have referred to the class Humn which I probably do not want to exist because I have spelt it wrongly. Unfortunately, as it stands my code does not know this and so will create the class “Humn”. Now, this form of error is not that likely to happen; tools such as Kudu (http://robertdavidstevens.wordpress.com/2010/04/24/my-own-ontology-projects) enforce this correctness in the Editor, while pabbrev.el (http://www.russet.org.uk/blog/2161) provides “correctness-by-completion”. None the less, these errors will happen and I do not want them to. There are a variety of ways that I could build this form of checking in — generally, this would involve introspecting over the ontology to see if classes already exist.

However, I have taken a different approach, so that I can use the Lisp itself to prevent the problem. To do this, for each class created, I generate a new Lisp symbol; likewise, object property and the ontology itself. The practical upshot of this, I that I can write code like so:

(defclass a)
(defclass b :subclass a)

(defoproperty r)
(defclass d
     :subclass (some r b))

;; will fail as f does not exist
(defclass e
     :subclass f)

;; will fail as r and b are the wrong way around
(defclass e
     :subclass (some b r))

The advantages are three-fold. Firstly, it’s slightly shorter, and there is no need to use quotes all over the place. Secondly, it is no longer possible to refer to a class that has not yet been defined; Clojure will pick this up immediately; from the user perspective, you can test your statements as you go, as soon as you have written them, by evaluating them. Finally, because the atoms carry values which are typed, we can also detect errors such as using a property when a class is necessary.

Of course, the original functions are all still in place; there would be no point defining symbols if the intention was to use the API entirely programmatically. But, my intention for Clojure-OWL is to have environment for humans (well, programmers anyway) to develop ontologies with.

There is a final advantage to this, that I have not yet exploited. Currently, I have generated the name of the OWL class directly from the symbol name. So, in the above example the class a will have a name “a“. There are some problems with this. Not all characters are legal in Clojure symbol names nor in OWL class names, and the set of characters is not the same. So, while this is a useful default, I will formally separate these. At the same time, I think that this will allow me to address a second problem, that of semantics vs semantics free identifiers (http://www.russet.org.uk/blog/2040). I can call a class, ontology or object property anything at all, and refer to it with a easy to remember identifier. I might use something like this:

(defoproperty has_part
   :name "BFO_OOOOO51")

The is still a significant amount of work to do yet; I haven’t made a complete coverage of OWL yet, just the most important parts (i.e. the bits that I use most often). Next, I need to start building some predicates so I can test (asserted) subclass relationships. So far, however, this approach is showing significant promise.

Bibliography

I have been struggling for a while with OWL development environments. While Protege provides a nice GUI based system, this has the limitations of many such systems; it allows you to do what the authors intended, but not all of the things that you might wish.

It is partly for this reason that I have been developing my own OWL Manchester syntax mode for Emacs (http://www.russet.org.uk/blog/2161); I lose a lot from Protege, but then I also gain the ability to manipulate large numbers of classes at once, as well as easy access to versioning. These things are useful.

Still, the environment is lacking in many ways; recently, while building an ontology for karyotypes (http://www.russet.org.uk/blog/2202), I wanted a more programmatic environment. A trivial example, for instance, comes from the human chromosomes; there are 22 autosomes in all. These can easily be expressed in OWL with 22 classes (plus X and Y). The problem is that all of these classes are likely to be very similar, which produces a code duplication problem. Of course, this is not a new problem; OPPL — the ontology pre-processor language was created at least in part for this purpose (10.1038/npre.2009.4006.1).

The main problem with OPPL, however, is that is a Domain Specific Language; while this makes it well adapted to its task, it also means that it lacks many basic features of a “real” programming language. Another possibility is to use the OWL API (10.1007/978-3-540-39718-2_42) (I am actually on this paper, but I publicly acknowledge that this was a rather generous attribution from Sean Bechhofer; I did do some work on the API, but not much, and I suspect none of my work remains). However, a brief look at the OWL API tutorial shows a problem. This code creates two classes and makes one a subclass of another.

OWLOntologyManager m = create();
OWLOntology o = m.createOntology(pizza_iri);
// class A and class B
OWLClass clsA = df.getOWLClass(IRI.create(pizza_iri + "#A"));
OWLClass clsB = df.getOWLClass(IRI.create(pizza_iri + "#B"));
// Now create the axiom
OWLAxiom axiom = df.getOWLSubClassOfAxiom(clsA, clsB);
// add the axiom to the ontology.
AddAxiom addAxiom = new AddAxiom(o, axiom);
// We now use the manager to apply the change
m.applyChange(addAxiom);
// remove the axiom from the ontology
RemoveAxiom removeAxiom = new RemoveAxiom(o, axiom);
m.applyChange(removeAxiom);

Aside from the intrinsic problems of Java — the compile, run, test cycle is rather clunky for this sort of work, this amount of code to achieve something straightforward makes this a little untenable.

Class: piz:A
    SubClassOf:
        piz:B

However, while Java and the OWL API do not seem a good choice for manipulating OWL directly, rewriting everything from first principles would also be a bad idea.

One solution to this problem came to my attention recently, in the shape of Clojure; essentially, this is a lisp implemented on the JVM. I will not describe the virtues or otherwise of Lisp in great detail; for some reason it is one of those languages that tends to generate fanaticism, and there are lots of descriptions of lisp elsewhere. For my purposes, there were three advantages. The first was personal, which is that I know Lisp reasonably well being an Emacs hacker. The other two are more general: Clojure has good integration with Java, and can manipulate Java objects, meaning I can make direct use of the OWL API; and, second, Lisp has a good degree of syntactic plasticity, which is important as, after all, I am looking for a convenient representation.

Initially, I have aimed at producing a representation which is fairly similar to Manchester syntax (http://www.w3.org/TR/owl2-manchester-syntax/). My initial attempts used the various features of Clojure directly. Consider, for instance, the following two statements:

(owl/owlclass
 "Arm" {:subclass "Limb"})

(owl/owlclass
 "HumanArm" {:subclass ["Limb" "HumanBodyPart"]})

Lisp, in general, uses a prefix notation. There is no obvious and easy way around this; in this case, it actually fits rather well with Manchester syntax which looks similar. The use of frame keywords such as :SubClassOf in Manchester syntax is also fortuitous as lisp uses a similar syntax. However, this syntax is rather too difficult. Even in this simple example we have a statement terminator which looks like ]}) (representing end of a vector, hash and sequence respectively). Lisp’s are often criticised for having too many parentheses; Clojure is unusual in using lots of different styles of parens. In Emacs-Lisp, I just keep hitting ) till I finished. In Clojure, you have to the brackets in the right order. All rather painful.

Fixing this turned out to be quite difficult, with a particularly nasty function I have called groupify. It is heavily recursive, which is apparently, a poor idea in Clojure, as it lacks some recursion optimisations present in many lisps; however, without mutable local variables, I could see no other option. The syntax now looks much simpler.

(owl/owlclass
 "Arm" :subclass "Limb")

(owl/owlclass
 "HumanArm" :subclass "Limb" "HumanBodyPart")

(owl/owlclass
 "Hand" :subclass (owl/some "isPartOf" "Arm"))

Both the :subclass and :equivalent frames support any number of class expressions; so far I have only implemented some or only, but the rest are not hard. Currently, it is only possible to save the ontology in Manchester syntax, but fixing this is trivial; the OWL API is doing all of the work.

Of course, this would not be much help if all I had managed to achieve was Manchester syntax with more parens. However, the big advantage of this becomes clearer with the next example:

(dorun
 (map
  (fn [x]
    (owl/owlclass
     (str "HumanChromosome" x)
     :subclass "HumanChromosome"))
  (concat '("X" "Y") (range 1 23))))

This creates a class for each human chromosome. In this case, I have hard coded the list of classes in, but I could be parsing a CSV or accessing a database. Or accessing an existing ontology; this could be very useful in avoiding maintenance of duplicate hierarchies.

Still, as it stands is just a (under-functional) version of OPPL. To make this worthwhile, I need to build off the language features that Clojure brings. I want to be able to interact with a reasoner, performing tasks in batch. In particular, the next step is to hook into Clojure’s test framework; something I have sorely missed when ontology building as opposed to programming. My experiences so far with combining Clojure and the OWL API suggest this should not be too hard.

These would not be minor advances; in the same way that test-driven programming has had a significant impact on the way we code, having a good test frame work for OWL would mean that we could define our use cases up-front, formally, programmatically and then fiddle with the logical representation till they work. As with test-driven programming, the test cases would themselves start to form part of the documentation for the code. When combined with a literate framework (http://www.russet.org.uk/blog/1213), to link between the ontology, the test cases and the experimental data that we are attempting to represent and model, this would provide a strong environment indeed. It would be a good step from moving from the craft-based approach we are taking at the moment, toward the pragmatic environment that I and others (http://robertdavidstevens.wordpress.com/2011/05/26/unicorns-in-my-ontology/) feel we need.

My code is available on Google code at http://code.google.com/p/clojure-owl/, and will be developed further there.

Bibliography

Recently, I and my PhD student, Jennifer Warrender have become interested in the representation of karyotypes. There are descriptions of chromosome complement of an individual. In essence, they are a birds-eye view of the genome. Normally, they are described using a karyotype string, so my karyotype would be 46,XY (probably!) which is normal male. When describing abnormalities, these can get very complex; take, for example, 46,XX,der(9),t(9;11)(p22;q23)t(11;12)(p13;q22)c,der(11)t(9;11)t(11;12)c,der(12)t(11;12)c[20] which describes a patient with multiple translocations.

There are a couple of reasons why we thought that it would be interesting to turn this into an ontological representation. The karyotype strings are not very parsable, and lack a computable specification which makes it very hard to check that they are correct, or to search and query over them. Having an ontological representation should help with the specification; additionally, using OWL we should be able to reason over by the specification and individual karyotype.

So, far it is turning out to be quite an interesting experience. The definitive resource for these strings is called ISCN (http://www.worldcat.org/title/iscn-2009/oclc/496946115). One of the things that we thought would happen, is that we would detect some inconsistencies in this specification; this often happens when producing a computable specification from a human-readable one, and it is not a reflection on the authors of the original. Sure enough, this is turning out to be the case. One the second page of the specification (after front matter), for instance, we find this statement:

Group G (21-22,Y): Short acrocentric chromosomes with satellites. The Y chromosome bears no satellites

pg7
— ISCN

The two sentences here are contradictory in themselves; either, the Y chromosome should not be in Group G, or Group G should not be defined to bear satellites.

There is an apparently similar statement on the page before, which says:

Not all the chromosomes in the D and G groups show satellites on their short arms in a single cell

pg6
— ISCN

However, this there is a significant difference here; to fill in the cytogenetic background, a satellite is a differentially staining visible body on the chromosome, found near the the centromere of some chromosomes. The name “satellite” actually comes from a different property of satellite DNA, that is is often a different density from bulk genomic DNA, so when spun in a density gradient, appears as a smaller band segregated from the main genomic band. Nowadays, satellite DNA is known to be highly repetitive sequence that we are no longer allowed to called Junk DNA (http://www.bbc.co.uk/news/health-19202141). In humans, the most common sequence is known as the alpha satellite. The different densities sometimes seen in gradients occurs where the GC content is different from bulk. The differential staining patterns seen cytogenetically occurs because the repetitive DNA is normally packed as heterochromatin.

Which leads us to understand the difference between our two quotes from ISCN. Although, repetitive DNA is variable in detail, it is not that variable, and will remain constant within an individual; if one Chromosome 22 has alpha satellite at the centromere, then so will another. The key here is the caveat “in a single cell”. In a chromosome spread from a single cell, an individual chromosome may or may not be at the right stage of condensation for the satellite to be visible.

So, we have two usages of the word chromosome. In the first quote, we are referring to a canonical chromosome; so, canonically, it is true that all human chromosome 22s contain a satellite. In the second statement, we are referring to a single chromosome, in a single cell, from a single human. There is no contradiction here because the existence of a single chromosome 22 without satellites is not inconsistent with the canonical chromosome 22 having satellites.

Actually, ontologically, the situation is slightly more complex still. The second quote says “Not all..show satellites” (our emphasis). It is not an issue of whether the chromosomes have or do not have the relevant DNA, it is whether we happen to be able to see this after appropriate staining.

We will consider both of these issues — canonicalization, and visualisation in future posts.

Authors

This post was authored by Phillip Lord and Jennifer Warrender.

Bibliography

I have described my experiences of using Emacs for writing ontologies previously (http://www.russet.org.uk/blog/2161). I was not entirely happy with omn-mode, even after recent changes (http://www.russet.org.uk/blog/2185), so I have taken the opportunity to update it a little more. This article most describes some implementation changes.

Originally, omn-mode was based on generic.el; this is a package which enables development of simple major modes. However, the emphasis was on simple, and my code was getting a little bit complex; generic.el was starting to get in the way. Moving to the define-derived-mode was not pain-free; it involved redoing already functioning code which is always a bit down heartening but probably worthwhile.

One thing that I did have problems with was getting comments to work properly (Emacs syntax tables are nasty). This never worked properly anyway, as I had defined “# ” as the comment starter; the reason for this is that Manchester syntax also uses URIs, within which any character is basically legal, including “#” but not a space. I think I have now found a better workaround for this. Manchester syntax requires URIs to be <surrounded>. I’ve now defined “<” and “>” to be string delimiters, which means that I can now use “#” as a comment starter without it being recognised in a URI. This isn’t perfect; in particular, Emacs will not recognise “<” and “>” as paired, so <URI> is equivalent to <URI< and >URI>, although only the former is correct.

I’ve also update the indentation logic somewhat. This now uses the syntax-table parser rather than font-lock to work out whether point is in comment or string. Of course, it should have been this way all along but the syntax-ppss function is new to me.

Finally, I have added some electric features. Because the identation engine works on keywords rather than brackets, the identation level really needs to be calculated when a line is finished rather than when it is started; electric mode means that the indentation updates as the user types. It’s a little jarring in some ways, although Python does something similar and for the same reason.

So, no major changes from the point of view of the user, except that it should all just work a little better. Code at Google code or website.

Bibliography