When Peter Murray-Rust restarted blogging recently, I must admit that to having mixed feelings. His blog is very interesting, often insightful and entertaining. He is, however, rather prolific, making up a considerable chunk of my RSS inbox. This would be okay if he was dull, of course, as I'd just unsubscribe, but it's not true.

As a case in point, he recently talked about Ontological Wars; this lead me to the Upper Ontology page on wikipedia which I'd not read before. Mostly of this page is not about upper ontologies but two sides sniping at each other about why upper ontologies are or are not possible.

Since the whole idea of upper ontologies came into bio-ontologies, I have to admit to being deeply ambivalent about them; I can see the appeal, of course. There is a pleasure at fiddling about at the upper, most abstract levels of knowledge. Career-wise, upper ontologies are high-risk but think about the potential publication rate if every one uses your ontology; of course, the actual value might be small to each individual, but if you get a publication out of each; well, it's like writing Maniatis (famous for having a funky name, as well as the book), or BLAST (which gets cited by everyone).

The flip side is that, I think that there is rather little evidence that using a single common upper ontology actually aids the processes of ontology development, deployment or integration. It can help somewhat, but then people end up spend too much time thinking about the philosophy of upper ontologies, which ultimately can take a lot of time; take a look at the BFO mailing list if you want to see how I have fallen in to this trap. On the other end of the process, how much use is an upper ontology in terms of querying? For example, it might be good to know that the function of a test tube and the function of beta-galactosidase are actually instances of the same RealizableEntity, but does anyone ever query at this level of abstraction.

I think that the core problem here, is that upper ontologies tend to be built without evidence; rather illustrative examples are chosen and then used to derive general truths. An illustrative example of this approach is, for example, Barry Smith's paper on part of; the example is a circle half of which is red, half of which is white. Okay, but where is the evidence that this is a good example? Can we be sure that if we picked a different example, the conclusions would not have been different?

This, I think, covers the key problem. At the moment, are attempting to build upper ontologies from the top down; those people who are interested in upper ontologies tend not to apply them to large scale projects; those people building lower ontologies tend not to discuss the applicability of upper ontologies for fear of getting shot down in flames; see wikipedia if you like flame wars. What we need is an arbiter, some way of determining who is right, who is wrong; as a scientist, of course, I know how to do this; I do an experiment. You can argue philosophy all you like, but having not one illustrative example can never outdo having several hundred actual uses. We don't entirely know how to do these experiments yet, but that's partly because we are not trying. I once got told on the BFO mailing list (and I paraphrase): YOU can do controlled experiments if you like, but I'm too busy doing science for that.

In ontology building, we need to avoid arguments like "is it correct", "is it true" or "is it reality" and replace them with "does it work". And to do this, we need to take a small step back and ask: how do we know when our ontology works; and most importantly of all, how can we guess when its likely to work in the future. Only then can we choose with knowledge between the different upper ontologies or, indeed, none at all.

Enough philosophical ramblings; back to work.