Wednesday, December 21, 2011

Ontologies, named graphs and links

Normally our searches undergo a process like:

@(@(MyGraph,OptimizerGraph),GraphLand)

The major search engines take a sloppy TE list of key words, run it through the graph optimizer which find two or three orderings for the words, adds a few synonyms. The resulting opmized search graph is then convolved with G land.

In the expression, all three graphs are named graphs. How do named graphs and URL jumps coalesce? A named graph is either a node index into a local table, or a node index into a remote table. How does nested stores maintain the distinction, and do URL jumps always require a named graph? Is their special indexing used for named graphs?
Dunno, but a few posts and I can drive the industry toward a consensus on this. It is a big problem right now with everyone doing semantic nets and convolution models.

One thing I can add, the table is not known to the client, who knows only graphs. A table is an internal construct, mainly because my graph layer is sqlite3. Graph layers will, in general, have to have an internal construct, like a table, a collection, something the query system can latch onto. We now with certainty, I think, that the table will at least be a complete and correct nested store.

So let's define the general philosophy of meaning and look ups in ontology networks.
The idea is series of convolutions of a search graph transforming it from a less informative structure to a more informative structure. In the example above, for example, the philosophy would state the the query optimizer is a local ontology network familiar from the past with user keywords and click thrus. So the query optimizer takes search graph through a revision process, translating loose TE into tight TE, based upon local knowledge of the client. The result is sent into G land where the data structures may be very square, very well indexed, and using very proprietary search methods. Do not lose meaning when it has been captured. Under different clients, bedpan may refer to products, or it may refer to a set of very badly designed URLs; who knows what ontologies the client has created with his key words? The local optimizer knows.

What is my underlying assumption?
The graph layer and BSON layer support web bots, bots that operate within g land and recognize meanings via click thru statistics and keyword sets. Consider attributes. In my local machine I have enough indexing space to cross a thousand URLs with a thousand attributes. But this attribute list is unavailable to anyone but me, easy to do. The web bots are the individuals who determine whether an attribute link set should be promoted. Promoted up the net so other clients access it, and promoted into square schema wherever possible.

No comments: