cygri’s notes on web data

Ivan is here today for the W3C Tag and gives a keynote to the general conference audience (Slides). If you read this through Planet RDF you won’t find much new in this talk, but I think it is very appropriate for the audience of XML and Web folks here.

Is the Semantic Web AI on the Web? No. “Beware the Hype!” The Semantic Web doesn’t intend to be the Future of Human Knowledge. It is infrastructure for building software that acts as “intelligent secretaries”, but only in the sense of reducing drudge work. Database integration might be the most important application of the Semantic Web.

What is the Semantic Web, then? It is the Web of Data.

What is the relationship to AI? Some SW technologies have benefited from AI research, and the SW has brought new concerns and use cases to AI.

What is RDF? It’s about creating relationships among resources on the Web and to interchange them. It’s pretty much like the hyperlinks on the traditional Web, except there is no “current” document, there is no user-interface action for “clicking” a link, and links are typed.

But isn’t RDF simply an (ugly) XML application? No. RDF is really a graph. RDF/XML is just one way to write it down. Think in terms of graphs, the rest is syntactic sugar. Yes, RDF/XML is ugly because it was developed in the “prehistory” of XML, and now there’s too much legacy code to change it. If you prefer, use Turtle.

What about ontologies and rules? They are a “glue” to help with data integration. Ontology and rule processors can automatically find out that different concepts in graphs are actually the same thing.

Is all this surprising? No, we do this kind of data integration all the time on the Web, using our brains. This is just about adding some bits to the Web that are needed to allow machines to do a part of it.

“One has to learn formal logic to understand and use the Semantic Web.” No, it doesn’t have to be very complex. A little glue can take you quite far. There is an “onion” with RDF in the middle, then RDF Schema, OWL Lite, OWL DL, OWL Full. An application may choose the complexity it wants. Compare to SQL, where the formal semantics is very complex, but 95% of SQL users never looked at it. Developing an ontology may require more knowledge, but that’s for a small percentage of users.

Isn’t this research only? Does this have any industrial relevance? There’s a lot of activity in health care, life sciences, digital libraries, defense. There are lots of tools now, so lack of good tools is no longer an excuse. Remember the original Web started at CERN …

Ivan pointed to lots of areas of current development throughout the talk: GRDDL, RDFa, Database-to-RDF mapping (he mentions our work), SPARQL.

From the problem of the NP-completeness of ontologies, we now move to an ontology of NP-complete problems.

The famous “P = NP?” is one of the big unsolved questions of computer science, with a large body of existing work. The talk is about formalizing the domain of NP-complete problems.

The ontology has classes like “Decision Problem”, “Complexity Class” and “Algorithm”. The ontology was populated with 350 problems from a standard text book on NP-completeness. Instances are annotated with further bibliographical references from some 1000 papers. Everything was entered by hand. (I think.)

Why do this? To assist bibliographical searches. There’s a web interface. (I think it’s supposed to be here, currently down.) In addition to the usual full text search, users can navigate the domain space. The web interface includes a cool widget that shows the current item’s relation to neighbouring items and looks quite usable – a good domain-specific search interface.

Implementation: Jena, servlets, Google/Yahoo API for full text search.

(This is an example of the “80% of Semantic Web projects” Heiner Stuckenschmidt was referring to earlier today. I think the value in this project is the aggregated knowledge collected by a number of experts over years. This would become really interesting if it would be possible to link this with knowledge bases in neighbouring domains maintained by other groups of experts.)

In this short talk, Nicole describes a system that extracts knowledge from mathematical texts in order to build an intelligent encyclopedia.

In mathematical texts, there are sections with a fairly regular structure: conditions, conclusions, properties. “Let x be … Y is a Z ifff …” These can be extracted with text analysis.

The goal is an ontology of mathematical texts, not mathematics.

They use the TRALE system, a grammar analyser for the German language. The implementation uses Jena. They used to output topic maps, but changed to OWL.

Q: (Rainer Eckstein) So how much effort is this?

A: A lot. We had to build a lexicon for the TRALE parser first, that took about a year.

Q: Implementation?

A: Convert input texts to TEI, then a lot of Java, some Python scripts. TRALE is Prolog.

I’m at Berliner XML Tage, an annual German-language XML and Semantic Web conference in Berlin.

The keynote by Heiner Stuckenschmidt of Universität Mannheim features a provocative title. After a recap of the 2001 SciAm paper and a short journey up and down the Semantic Web technology stack, it gets more interesting later on.

Typical Semantic Web applications of today work something like this: Information is extracted from existing data, like web pages. An ontology is built with Protégé. Some inference is done over everything, and some views generated (“Semantic Portals”). Heiner estimates that this describes 80% of current Semantic Web projects.

But that’s not “living in the Real Web”, Heiner says. We want distributed information, a P2P-like architecture, and have to deal with the possibility of failure, inconsistency, incompleteness and heterogenity of data.

Don’t do big monolithical ontologies. Ontologies should be modular.

Don’t do big monolithical servers. Things should happen distributed, e.g. based on ontology structure, that is, put related information on the same server but not everything on a single one, or leave the data and ontologies “out there” in the Web.

Don’t do big monolithical reasoning. Gathering all the data in one place to do inference is not optimal. Inference engines should sit on every node in the distributed system. (He points to C-OWL, Distributed Description Logics and the DRAGO system as possible approaches to distributed inference.)

Distributed Ontologies: Heiner thinks that OWL works well for the old scenario where ontologies are built by individual persons or organizations, but is not very good for working with distributed ontologies. There are many situations where “A is almost subClassOf B”, but in OWL, a single outliner prevents us from relating the two classes. A possible way out would be probabilistic logic.

So, what’s wrong with the Semantic Web? Too little attention has been paid to the specific needs of a distributed environment. The Semantic Web is a part of the World Wide Web.

Q: You talked about distributed ontologies. What about distributed RDF knowledge bases?

A: They exist, the need for distribution has been realized earlier in this area. There are many systems that do centralized reasoning over distributed data.

Q: Shouldn’t we concentrate more on RDF data and schema than OWL?

A: Maybe – but I’m a logician, not a database guy.

Q: (Klaus Schild) Scalability? OWL is NP-complete.

A: Distribution helps. Certain combinations of operators are deadly, but if the two operators happen to end up on different nodes, things can be much faster. But in general it’s a problem.

Q: But if you take the Web seriously, you need sub-linear complexity.

A: You can’t have that with RDF and OWL.