Artificial Intelligence and the Semantic Web

Tim Berners-Lee, <timbl@w3.org>

Artificial Intelligence and the Semantic Web

http://www.w3.org/2006/Talks/0718-aaai-tbl/

Tim Berners-Lee

Decentralized Information Group
MIT Computer Science and Artificial Intelligence Laboratory

AAAI, 18 July 2006

AI and SW

SW is not AI and AI is not SW
AI is a field; SW is a project
SW owes a debt: used much from AI
SW should be a great playground for AI
AI projects should use SW to interoperate

What is AI anyway?

Lots of bits of CS including functional languages, machine learning, recognition, speech, etc., etc.?
An attempt to make smart machines?
An attempt to make machines which behave like people?
An attempt to be collectively more powerful?

What is SW?

Data Interoperability across applications and organizations (for IT)
A set of interoperable standards for knowledge exchange
A architecture for interconnected communities and vocabularies

The RDF data bus

The RDF bus connects data sources and applications

db to sw

Sem Web architecture 101

Give important concepts URIs.
Each URI identifies one concept.
Share these symbols between many languages
Support URI lookup

Define symbols:

Using natural language (bootstrap?)
By reference to existing systems (eg GPS)
By mathematical relation to others (raft)

Chis Welty/IBM: "In the Semantic Web, it is not the Semantic which is new, it is the Web which is new".

Web architecture 101

Things are denoted by URIs.
Use them to denote things.
Serve useful information at them.
Dereference them.

Example

http://www.w3.org/People/Berners-Lee/card#i

http://www.w3.org/People/Berners-Lee/card (in N3, summarized):

@prefix foaf: <http://xmlns.com/foaf/0.1/> .

@prefix : <#>.



:i a foaf:Person;

 foaf:family_name "Berners-Lee";

 foaf:givenname "Timothy";

 foaf:homepage <http://www.w3.org/People/Berners-Lee>;

 foaf:mbox <mailto:timbl@w3.org>.

URI + HTTP architecture 1

The hash is an operator which joins a local identifier to a document URI to give a global identifier.

http://example.com/foo#bar

Strip off #bar
Look up http://example.com/foo using HTTP:
1. Look up example.com giving 128.0.0.1
2. Request foo from 123.0.0.1
3. 200 OK is returned
4. Parse the result according to the Internet content type
This gives you information about bar

URI + HTTP architecture 2

Post TAG resolution of HTTPRange-14, an optional possible operation is:
given http://example.com/foo/bar

Strip off #bar
Look up http://example.com/foo using HTTP as amended by TAG:
1. Look up example.com giving 128.0.0.1
2. Request foo from 123.0.0.1
3. You get a redirection 303 See Other response, indicating that the URI did not denote an information resource, but mentioning a new resource http://example.com/foo-schema.rdf
4. Request http://example.com/foo-schema.rdf
5. Get a 200 OK response
6. Parse the result according to the Internet content type
This gives you information about <http://example.com/foo/bar>

Not recommended by me.
Other issues: Content negotiation between HTMl and RDF. LSIDs

Breadcrumbs ethos

Leave information for others to follow
a Breadcrumbs protocol = what to leave where + what links to follow when
Delegated query is a special case of breadcrumbs
Challenge: For delegated query, how to describe the sort of query a service or document can answer

Mythbusting

Myth: "The Semantic Web technology is Description Logic"

No, OWL is one semantic web language.

It is important that applications which need different expressiveness can use it.

But other languages must interoperate to the greatest extent possible.

They should use URIs

They should not reinvent functionality already provided by standards.

SW Arch: Same symbols, multiple languages

architectural layers

Mythbusting: Not just public data

The SW is not just about public data.
It also about personal, group, agency and enterprise data.
Historically, intranet servers preceded extranet servers.

Its like a metro, the way the lines of common concepts connect the stations of different applications

For example in biopax

When will the patterns all connect? Venn diagram showing ontologies overlapping by certian common terms

[Diagram: Joanne Luciano, Predictive Medicine; Drug discovery demo using RDF, Sideran Seamark and Oracle 10g]

Other myths

"The semantic Web is metadata for classifying documents"
"The semantic web is about hand-annotated web pages"
Such pages are interesting, but not the mainstay of semantic web: too much trouble!
"The semantic web is mainly about content extracted from text"
No, it is primarily an interlingua for relational data and logic. bridges will always be important
"The Semantic Web is about making one big ontology"
The semantic web is about a fractal mess of interconnected ontologies....
"The semantic web ontologies must all be consistent"
Only the parts I am using together

"Ontology": two patterns

Schema	Taxonomy
E.g. Bank statement: account, date, amount	E.g. Human anatomy and diseases
Simple ontology. Often documents existing practice	Complex ontology, difficult to make
Changes rarely	Changes continually
Domain knowledge is in the data not the ontology.	Domain knowledge is in the ontology itself
Use now in enterprise and science IT	Specific fields e.g. life sciences

Both patterns are important. Both patterns use OWL.

NLP vs Semantic Web

NLP	Semantic Web
Words	Terms of logic
Meaning is use	Meaning is defined in words or code or specific use.
Word reused by everyone, no ownership	URI ownership - go get your own
"Hydrogen"	http://example.com/foo.rdf#Hydrogen
Defining words in terms of ontology complex, unsatisfying, never complete and a waste of time.	Defining terms using words is never perfect but useful.
Natural language constantly changing	Ontologies basically static
Can't benefit from injected logic	Can't benefit from cloudy stats from corpora data
Machine can find stuff	Machine can make widespread inference

Distractions: Meaning of meaning

Meaning as definition	Community Standards	Meaning as Use
Everyone uses the same precise definition of each term.	Specific communities agree to share good-enough definitions	Language changes with time
Works in small close systems	If communities overlap, can be global	Works for poetry, makes rich natural language.
Very hard work to set up	Finite work to set up	No effort up front, much afterward.

When we build a system (the SW), this is a choice, not an observation

The fractal tangle

Communities will be of many sizes.
There will be very many small ones (6.10^9 of size 10^0) and a few global ones (e.g. W3C Rec'n)
Kleinberg shows that fractal (1/f) distribution is optimal under some assumptions
Swoogle results for example (right)
We have less experience when fractal is not constrained to a 2D surface.

Total Cost of Ontologies (TCO)

Assume :-) ontologies evenly spread across orders of magnitude; committee size as log(community), time as committee^2, cost shared across community.

Scale	Eg	Committee size	Cost per ontology (weeks)	My share of cost
0	Me	1	1	1
10	My team	4	16	1.6
100	Group	7	49	0.49
1000		10	100	0.10
10k	Enterprise	13	169	0.017
100k	Business area	16	256	0.0026
1M		19	361	0.00036
10M		22	484	0.000048
100M	National, State	25	625	0.000006
1G	EU, US	28	784	0.000001
10G	Planet	31	961	0.000000

Total cost of 10 ontologies: 3.2 weeks. Serious project: 30 ontologies, TCO = 10 weeks.
Lesson: Do your bit. Others will do theirs.
Thank those who do working groups!

User Interface challenges

Domain-specific user interfaces are blossoming... but what about generic ones?

Generality: can browse any data anywhere
Dynamically pick up from ontology: Lenses, style, forms
Independent control of: style, provenance, domains (vocabulary groups)
Blow spreadsheet tools away

Goals for Rules

Interoperability between Rule-based systems
Extend the expressive power of shared knowledge
The next step in the Semantic Web roadmap
New markets for rule systems
Mapping between ontologies
Interlocking with other Semantic Web languages RDF, OWL, SPARQL
Serendipitous re-use of knowledge in Rule form