Friday, June 03, 2005

Sure enough, the sore throat developed. I'll sure be glad when my immune system is used to seeing all these bugs that Luc brings home. It slowed me down a little in the last couple of days, but I still wrote a lot of code. Well, I think it was a lot. :-)

The main impact was that I didn't go out to exercise (which I really need to do in order to work efficiently) and I was also too tired to blog. I'm too tired again tonight, but sometimes you have to push the envelope.

The code over the last few days has been building data structures based on RDF. I had already done some of this work before, but it turns out that integration has needed a more thorough exploration of the data store.

I've realised that there are two ways to write this code. I can write it with a full knowledge of the data structures I'm reading, or I can write it with very little knowledge, and use the ontology to tell me how to build the data structure.

I started out by using the ontology of the rules just a little, while mostly relying on my own knowledge of the data structure and putting that explicitly into the code. This is fine, but it's not very extensible. It was while building the constraint tree that I started to see some other potential problems with the explicit approach.

Each node in the constraint tree is either a leaf, or refers to two or more child nodes. To query an RDF structure about a tree of arbitrary depth it is necessary to get all the links from parents to children as a set, and to connect them together. As each node is found for the first time, the associated Java object must be created to go with it. A problem can arise here if the node is of a type that extends another concrete type. In that case it is necessary to put off creating a node until all of its types have been found, and to then build the most specific type. This is where the ontology starts to play a part.

For a start, simple RDF will just give me the concrete types of the nodes, with no information about the superclasses. It is only with an ontology that the other types would become available (I don't know about anyone else, but I'll probably end up running the inferences against the rules themselves, just to see if I can). At that point I'll need to structure the types together (using transitive owl:subClassOf) and check that there are no loops (except the obligatory subClassOf(A,A)), before finding the most specific type to instantiate. Instantiation will be of a class that would be associated with the node (I don't have that yet, but it would be trivial to add). Of course, multiple inheritance makes it that little bit harder.

The advantage of such an approach would be that extra classes in the structure would just require an OWL definition. Then they could be used in the RDF with no changes to the Java code.

This would apply to any kind of data structure that is expressed in RDF, not just my queries. Building a system like this would be similar to a UML modeling tool with runnable objects (something that I know some tools do). It's cheating a little to link it to a class name, but that is where the idea of storing a Java AST (Abstract Syntax Tree) in RDF would come in. In that way, the basic ontology of a set of classes could be written in OWL, with RDF annotations to describe the complete implementation of the classes. An RDF structure with rdf:types referring to these classes would describe an instance graph.

I really like this idea, and it shows some of the modeling power that comes with OWL and RDF. Unfortunately this would take a few weeks, and I don't have the time for it right now. Doing a Java AST representation in RDF would take a lot longer, though the compiler would be fun.

For the moment I'm sticking to what I know of the data structure, and almost ignoring the ontology. That's a shame, but it's letting me finish it in just a few days, instead of weeks. Fortunately the none of the instantiable classes in this system have descendents, so there won't be any potential confusion about the class to instantiate for any nodes in the constraint tree.

I spent Wednesday morning with Brad. He offered to show me how to get Kowari working with Eclipse. I had a go at this last year when we had the jars-inside-jars class loader in Kowari, and configuring this was very difficult, highly manual, and didn't really work well. It seems that the new flattened class arrangement works much better.

Brad was also able to show some of the useful tools included with Eclipse, such as refactoring, which certainly makes it seem quite compelling. I haven't made the transition just yet, as I haven't got CVS going yet. Eclipse does not seem to understand CVS directories that were created outside of it, so I'll need to get a fresh checkout, and port over my modified files. That might be a job for the weekend.

More importantly, it was very good to have a chat with someone who is using Kowari. It helps to keep focus on what I'm doing it for when I speak to people who actually interact with this stuff. In a similar way, I worked with Andrae this morning, and I appreciated hearing what another developer is doing.

Final Word
Oh, and I promised to make a comment about a "strange Luigi guy". If he's reading, then "Ciao".

No comments: