Musings on an Ontological approach to Data Selection from Astronomical Datasets

I've often been asked (and wondered myself) how ontologies can relate to the everyday astronomer's work.

In the simplest sense, ontologies are an accumulation of knowledge against data. As a store of data is created, certain assertions can be made about that data and stored in the ontology. If an assertion is made which contradicts previous assertions then the inference engine will tell the astronomer. Then either the new assertions or the previous ones can be corrected. In most cases, I suspect, it will be the previous ones which are incorrect, or if not incorrect, incomplete.

So, why is the astronomer to type all this assertion stuff in if, in most cases, the only result is the need to correct previous assertions?

We have an expectation that ontologies will be useful in storing and retrieving resource information in the registry and in assisting astronomers in constructing a workflow of jobs for submission to an automated job control system. How, I was wondering the other day, might ontologies be useful in the astronomer's day to day work: that of simply retrieving data from an archive?

The main activity against some archive or general dataset (or combination of datasets) is to select rows according to some criteria; those rows then make up a new dataset, possibly a catalog of objects of some description. The selection criteria, in the simplest sense will be a graph of UCD-operation-Value nodes (1) with boolean connectors.

We've spent some time thinking of how we can encode the metadata for a catalog or an archive. In many cases, the solutions are keywords or extensions to the UCD structures. Some of us have extended that thinking to the ontological world: how can we use DAML+OIL objects, properties and assertions to encode the metadata for some set of data.

The ontological approach to this is to store basic assertions about objects (star, binary star, galaxy, cluster of galaxies, etc), properties (binary star has 2 stars, galaxy contains stars, gravitational lensing is an effect of quasar etc) and assertions (I've not got that complex yet :)). Looked at in this way, the ontological assertions provide an ultra-rich form of metadata.

But what if we store ontology-based metadata about the query itself:

The UCD-op-Value graph is metadata in its own right (are there any other assertions you can make about the selection criteria?). The dataset itself has its own metadata in the form of assertions about its content. The astronomer can make assertions about the resultant dataset or catalog and its content to generate its metadata.

In a way, the metadata for the catalog is the result of some mapping of the metadata for the selection criteria (which may be little more than the UCD-op-Value graph) onto the metadata for the dataset. The metadata for the selection profile (the selection criteria applied to one or more datasets) should be that for the catalog.

The real benefit will come if we can find some way to derive the catalog metadata directly from the profile metadata. Or, even better, to describe the results wanted and the dataset(s) you want the results from and have the engine deduce the correct query criteria. Taking it one step even further, might we then state the metadata for the desired results and have the engine infer the combinations of dataset and selection metadata, and from that, to the actual query criteria and datasets.

This, of course, is pure speculation. I'll post a query on the semantics mailing list to see what the experts think.

References

(1) Universal Column Descriptors (UCDs) are a hierarchical set of standard column names which the astronomy community have agreed upon. The hierarchy is a taxonomy of a sort, though rather a simplistic one as it seeks to force a given column name into a single branch, disregarding relationships with other branches.

-- TonyLinde - 25 Sep 2002