Meaning and the Semantic Web

Meaning and the Semantic
Web

Yorick Wilks
Oxford Internet Institute
and
Florida Institute of Human and Machine
Cognition

Paris, May 2012

18/05/2012

Let’s start with serendipity

 “finding good consequences by chance”
 Just one form of unintended consequences
 Does this apply to the Web?

18/05/2012

Serendipity Unforeseen/Unintended Social
- Consequences
GOOD BAD MASSIVE
IT
POST-ITS SYSTEM
PCs __ FAILURE
+ Designed
THE BRAIN--SEEING Failures
vs. TALKING??? for X but
used for
Y
ABORTION MINITEL?
AND LOWER
CRIME Doing X
RATE politically
and getting Y
Unintended COMP.SCHOOLS AND
but foreseen? WIDER SOCIAL
DIVNS.
Worst
GARDEN
cases
OF
18/05/2012
EDEN?

Berners-Lee on the origins of the
WWW
 …in 1989, while working at the European Particle
Physics Laboratory, I proposed that a global
hypertext space be created in which any network-
accessible information could be refered to by a
single "Universal Document Identifier".
 I wrote in 1990 a program called
"WorldWideWeb", a point and click hypertext
editor which ran on the "NeXT" machine. This,
together with the first Web server, I released to the
High Energy Physics community at first, and to the
hypertext and NeXT communities in the summer
of 1991.
18/05/2012

 After much discussion I decided to form the World
Wide Web Consortium in September 1994, with a base
at MIT is the USA, INRIA in France, and now also at
Keio University in Japan. The Consortium is a neutral
open forum where companies and organizations to
whom the future of the Web is important come to
discuss and to agree on new common computer
protocols. It has been a center for issue raising, design,
and decision by consensus,

 The great need for information about information, to
help us categorize, sort, pay for, own information is
driving the design of languages for the web designed for
processing by machines, rather than people. The web of
human-readable document is being merged with a web
of machine-understandable data. The potential of the
mixture of humans and machines working together and
18/05/2012
communicating through the web could be immense.

What did Berners Lee originally
want?
 What he originally wanted was not a web of
documents but of data and objects that a
machine could understand.
 Accidental features in the WWW not
designed as the basis for universal
knowledge and services
 Initially owned not by CS but physicists
 New attempts to go back to his vision---as
we shall see below with the Semantic Web

18/05/2012

Berners-Lee’s WWW proposal

18/05/2012

Generalizations?
 Unlike huge private sector IT failures, the WWW is
a vast (US) Government IT success (+)
 Elite tools can become democratic leisure aids (+)
– ARPA, then CS, then physics
– Like air travel
 Elitism can fight back (+?)
– Paid information services
– Cf. all-business flights
 The sex wave overwhelms all such media (+/-?)
– Minitel
– Text messages (Vodafone 80%, sex/love/personal)
– Usenet Bulletin boards (alt.sex.fetish.diapers)
– The WWW—Pou Porn now much bigger than the 2000
web
18/05/2012

The Semantic Web: the
apotheosis
of linguistic annotation, or the
basis of eScience and
everything?

18/05/2012

The Semantic web
(Berners-Lee, Hendler, and Lassila, Scientific American
2001)

 A vision of making the Internet as readable by
computers (agents) as it is by us.
 A similar notion to the “ascent” in the semantic web
pyramid---meaning/interpretation somehow tricking
UP it from the bottom (cf. Braithwaite’s view of
scientific theories--neutrinos linked to experiment)
 Is this last what SW people mean/want, or do they
assume that the higher-level structures are self-
interpreting?
 18/05/2012 as the basis of eScience (at the end)?
The SW

 “In the middle of a cloudy thing is another
cloudy thing, and within that another cloudy
and thing, inside which is yet another cloudy
thing…………
 ………………and in that is yet another cloudy
thing, inside which is something perfectly clear
and definite.”

-----------Ancient Sufi saying.
18/05/2012

Forms of knowledge in the SW
 1) Universal Resource Indicators (URIs)
 2) Resource Description Format
 RDF triples---putting all facts in the form:
 John-LOVES-Mary
 Not quite logic yet! Basically IE output;
explicitly called « subject » « object »!
 3) Ontologies---trees of concepts in
hierarchical and functional relations: again
like
 Canary-ISA-Bird
 4) DAML/OIL reasoning languages
18/05/2012

The Semantic Web and Artificial
Intelligence (AI) and Natural
Language Processing (NLP)
 Some look at the top of the SW-pyramid and
say “it’s just GOFAI [Good Old Fashioned
AI] isnt it?”
 My case will be that if you look at the bottom
of the pyramid, the SW rests much more on
NLP than is usually realised.
 Of course, NLP people would say that,
wouldn’t they, but they may be right!
 But here’s another reason for keeping NLP
in mind……
18/05/2012

Sources of Knowledge
 80-85% of a company’s knowledge is contained in
unstructured form,
– i.e. expressed in some forms of natural language.

18/05/2012

The bottom levels of the SW
always sit on NLP annotations of
objects and actions (i.e. IE)
 Classic IE (Information Extraction) detects
named entities--populates SW’s “namespace”
 Does semantic type annotation
 Detects actions
 All recent IE works with an ontology
 It is also, of course, a SW annotator and RDF
finder
18/05/2012

Reminder: the sources of
annotation:
 Roff, nroff, troff--publishing languages superceded
by Knuth’s TeX (about 1972): adding to a text how it
should look.
 The empirical NLP movement starts with POS
tagging (CLAWS4, Leech, about 1979): adding to a
text “what it means”.
 The Text Encoding Initiative (TEI, 1985), then
SGML, adding to a text., but within the text
 The metadata concept came from the DARPA NLP
programs--annotations separate from the text.
 A subset of SGML later becomes HTML, then XML
for the Web, then anything-at-all--ML (e.g. VoiceML).
18/05/2012

“Texts are really programs”:

 Hewitt’s wall-building program in Rustin’s NLP
book 1972: “Natural Language is a side-
effect”
 Longuet-Higgins (and others): “English is just
a high level programming language” (1975?)
 Dijkstra: “ CS cannot really deal with natural
language (i.e. as NLP) , because the latter
isnt really up to its job” [YW’s version of ED’s
views].
18/05/2012

“Programs are really texts”:
 The “Wittgensteinian opposition” contains people like
YW:
 “Understanding without proofs” IJCAI 1973
 “Programs and texts” 1977
 “What’s in a symbol” (with S. Nirenburg) JETAI 2000
 Bad fellow travellers (Derrida et hoc genus omnes)
 This is the parasitic view of programs and logic vs. NL
(I.e. NL could exist without the others, and did for
millennia, but not vice versa). On this view,
logic/programs remain parasitic upon language in
varying degrees, in that both retain language-like
features.
18/05/2012

Annotations as the most recent
bridge from language to
programs and logic
 The switcheroo is from text-infixed annotations to
meta-data considered separate from the text itself
 The former still corresponds to something linguistic
and within NLP
 The latter conforms more to old-AI --the McCarthy-
Hayes program (1969)----and the dispensability of
the text
 This is the key notion: Remember this the
assumption of all true interlinguas (e.g. Schank),
18/05/2012 Sparck Jones’ case against SW/AI/NLP
and is

Annotation Engines
 Manual document annotation was
initiallyexpected to be the main SW vehicle
creation
– Especially for trusted environments (e.g. within a
company) this is expected to provide high quality
material
 Automatic annotation was a vision but now
largely fulfilled:
– To help manual annotation OR
– To replace human annotators
 Producing automatic annotation services
– For a specific ontological component/application
– Constantly re-indexing documents
18/05/2012

(UShef)
 Document annotation assisted by adaptive IE
 Users:
– Provides DAML ontology
– Annotates document samples
 IE system:
– Trains while users annotate
– Generalizes over seen cases
– Provides preliminary annotation for new documents
 Industrial users:
– Solcara (UK): commercialisation+2 applications
– Boeing (US)
– Quinary (I)

18/05/2012

Knowledge Sharing and Reuse
across Media

University of Ljubljana

U Freiburg

 http://support.google.com/webmasters/bi
n/answer.py?hl=en&answer=173379

18/05/2012

European SW projects tend to
accept some form of this NL-
based view of the SW
 E.g. X-Media (Ciravegna, Sheffield) a large
EU 6FP IST 2005-9 on multi-media web
services
 Based on IE/NLP annotation engines and
machine learning.
 BUT, US and eScience projects still see the
SW more in logic/AI based terms.

18/05/2012

Semantic Web getting real??-
Slashdot 11/2/08!
 Reuters just [1]opened access to
their[2]corporate semantic technology
[3]crown jewels. For free. For anyone.Their
[4]Calais API lets you turn unstructured text
into a formal RDF graph in about one second.
I ran about 5,000 documents through it and
played with a subset of them in [5]RDF-
Gravity. The results were impressive overall.
Is this the start of the [6]semantic web getting
real?When big names and big money start to
act, [7]not just talk, it may be time to pay
attention. Semantic applications anyone?
18/05/2012

Remember, two other key
components/interpretations of the
Semantic Web:
 The “Data-linked web” (Berners-Lee and
data bases!) and the huge amount of public
and scientific data being uploaded as RDF
triples—millions a day.
 The ontologies needed to organize all this.

18/05/2012

Mashups as a form of the
Semantic Web already with us
and available to the public

 http://www.police.uk/crime/?q=Oxford,%
20Oxfordshire%20OX2%206DQ,%20UK#c
rimetypes/2011-12http://

18/05/2012

Ameliorating the problem of
grounding the meaning of SW
terms (e.g. ontology terms)
 Inducing ontologies empirically from text
– Can it be done?
– Abraxas/Adaptiva
 Deriving very large language models to yield
RDF-like “forms of fact” from text (later!)
 These are both special forms of IE
(information extraction) from text
18/05/2012

Your very own Semantic Web?
Your very own Ontology?
 Former may be better than the latter
 Though the latter must be in our heads--we just
cant get at them—and individually different
 AI History of Points-of-View and To-hell-with-
consistency
– KRL
– Viewgen
– Hewitt’s Programmar
– De Kleer and Truth Maintenance
 Google-as-truth is also a POV phenomenon
 Scientific theories can have competing websites!
 Ontologies used to be personal and derived from
18/05/2012 thinking!
just

But having your own SW
structure will come at a huge cost
 Negotiating meanings all the time by showing
your ontologies and URIs
 These will have to stay pretty close together
or all communication will fall apart---how
much of conversation do you spend telling
people what you mean by terms (if you’re
NOT a philosopher!)
 But it is the personal data holdings that
companies now want access to and will
organise for you!
18/05/2012

An IR-based critique of the SW
and the idea of a universal
ontology behind it.
 Sparck Jones, K. (2004) What’s new about the
Semantic Web? ACM SIGIR Forum.
 “words stand for themselves”: the basis of
successful IR search in WWW and elsewhere.
 Content cannot be recoded in a general way
especially if it is general content--IR has gained
from “decreasing ontological expressiveness”
 Successful QA and IE are “superficial”

18/05/2012

The problem of Recoding content
 Charniak argued in 1974 that Schank’s
conceptual primitives such as
[PTRANS liquid skin] could not distinguish
sweat/sneeze/spit etc.
 (KSJ auction catalog example)”A Charles II
parcel-gilt cagework cup, circa 1670”
 What, she asks, can be recoded beyond
{object type: CUP}?
 The rest, she says, IS THE ENGLISH (it
could be any language, of course)
18/05/2012

The philosophical problems may
or may not just vanish as we
push ahead with annotations!
 David Lewis and « markerese »: his 1970s critique of
Fodor and Katz, and of any non-formal semantics
(such as semantic type annotation)
 The Semantic Web takes this critique head on and
carries on, hoping URIs and some »popping out of
the symbolic world » (e.g. giving the web your phone
number!) will itself solve semantic problems.
 Can all you want to know be put in RDF triples, and
can they support the reasoning you need to do?
 Note: philosophical critics of seantic primitives, like
Lewis, want just (non-symbolic?!) OBJECTS, whereas
Sparck Jones wanted just WORDS.
18/05/2012

What are the meanings of URI’s?
 “A resource is anything that might be
identified by a URI”(Berners-Lee et al.,
2005) i.e.
 Data, documents, images and real things
outside the Web! E.g.
 http://www.tour-eiffel.fr/
 But this is just another document, just like
the Dbpedia resource for an almost identical
URI.
 URI’s look best for
http://www.yoricksphone.com/”+44778530
18/05/2012

Annotation and the SW trying to break out
to meanings “elsewhere”
(just like David Lewis)
 HTML: <item>EiffelTower</item>
 XML: annotating directly, we might
extend this to:
 <item><semcat:physicalobject>EiffelTow
er</semcat:physicalobject></item>
 Encoding similar information in a
semantic web page might look like this:
 <item rdf:about=”
http://dbpedia.org/page/Eiffel_Tower">Eif
felTower</item>
18/05/2012

The URI issue is close to that of
every object having a name:

 The bar code on every piece of cheese in a
Tesco supermarket
 Every person having a name even though
they are far from unique—especially in
Sweden!
 Hardy’s poem on the iceberg waiting for the
Titanic and its history.

18/05/2012

“A Shape of Ice, for the time far and
dissociate”, Thomas Hardy  

18/05/2012

The issue of “Semantic Web
semantics” recapitulates much of
20C philosophy of meaning.
 Logical semantics alias descriptivist theories
of meaning (from Tarski): set theoretic
description of models, in terms of abstract
objects, BUT
 Any random alternative model is just as
good, even numbers! Commonsense will not
be able to pick out the right model!
 Hayes, P. (2004). RDF Semantics. Recommendation, W3C.
http://www.w3.org/TR/rdf-mt/
18/05/2012

A descriptivist view of URIs
(Halpin’s diagram)

18/05/2012

20C philosophy of meaning--2.
 Causal or Direct semantics alias Kripke’s
“rigid designators”-----historical “baptisms” of
objects.
 Maybe not far from Berners-Lee’s own view.
But it is not true to the way URI’s arise
casually and informally, in a world of millions
of RDF triples.

18/05/2012

A causal theory of reference for
URIs

18/05/2012

 “If it returns no web-page, how can a user-agent
distinguish a URI for a referent outside the Web from
that of a URI for a web-page? This question is
partially answered by Berners-Lee in a solution
called ‘303 redirection,’ where a distinct URI is given
to the thing-in-of-itself, and then when this URI is
accessed by an agent such as a web-browser, a
particular Web mechanism called the 303 Header
redirects to the agent to another URI for a web-page
describing the resource, either in RDF or in HTML,
or both. However, this mechanism has been
considered difficult to use and understand,
“analogous to requiring the postman dance a jig
when delivering an official letter” (Hayes, 1977).”
 Halpin, H. (2011). Sense and Reference on the Web. Minds and
Machines 21 (2):153-178.
18/05/2012

Remember Putnam’s scientists as “Guardians of
Meaning”

– Aluminium and Molybdenum look the same ----people
cannot tell them apart----- but scientists can.
– Therefore scientists really know the meanings of these
terms
– But they should not let this “leak out” to the general
populace or the meanings might change.
– “cats from Mars” and the meaning of “cats”
– Compare: heavy water vs. water
– BUT (said Hugh Mellor): we call heavy water “water”
because it is water-----and that’s simpler and more
democratic than the alternative view
– You cant lock away meanings and keep them safe
anywhere (as Wittgenstein might have said)--it’s all
usage in the end.
18/05/2012

20C philosophy of meaning--3.
 URIs as a public language
 Reversion to Frege’s view that there were
meanings separate from extensions (contra
Tarski and Kripke)—intensions or sense
 Closer to Wittgenstein’s view of a public
language (albeit ambiguous) over which
user’s have control.
 Wilks, Y. (2008). What would a Wittgensteinian computational
linguistics be like? In Proceedings of Convention for the AISB,
18/05/2012
Aberdeen, Scotland.

 “what is apparent from any analysis of the
Semantic Web is that there appear to be too
many URIs for some things, while no URIs for
other things. As differing users export their data
to the Web in a decentralized manner, new URIs
are always minted, and so running the risk of
fracturing the Semantic Web into isolated
‘semantic’ islands instead of becoming a unified
web, as the same URIs are not re-used. The
critical missing element of the Semantic Web is
some mechanism that allows users to come to
agreement on URIs and then share and re-use
them, a problem ignored both by the logical and
direct reference positions.” (Halpin, ibid)
18/05/2012

 But notice this is exactly what we do with
language all the time
 Lots of terms/phrases “for the same thing”
Lots of ambiguous descriptions
 The everyday deployment of language rests
on no formal basis whatever
 Only the “forms of our life” (Wittgenstein)

18/05/2012

A diversion from Wittgenstein to
experiment:
 Wittgenstein said ask for the use not the
meaning..
 Where would be better to look than the Web,
as its usage is now so much larger than any
human’s language (60K years of reading!)
 Roger Moore on a hundred year’s of human
training to get a modern speech recognition
system

18/05/2012

Vague and distant
relationship of Wittgenstein
with classic AI:
 "The solution to any problem in AI may be found
in the writings of Wittgenstein, though the
details of the implementation are sometimes
rather sketchy.”

 Graeme Hirst. "Context as a spurious concept." Proceedings, Conference
on Intelligent Text Processing and Computational Linguistics, Mexico City,
February 2000, 273--287.

18/05/2012

Recent empirical semantics.
Imagine extracting from a vast corpus all forms
of Agent-Action-Object triples (I.e. all examples
of who does what to whom etc.).
 Use these to resolve ambiguity and
interpretation problems of the kind that obsess
people who are into concepts like ‘coercion’
‘projection’, ‘metonymy’ etc. in lexical
semantics.
 E.g. if in doubt what ‘my car drinks gasoline’
means, look at the stored triples about what
cars do with gasoline and take a guess.
18/05/2012

Forms-of-facts as a useful reality?
 This isnt a very good algorithm, but it might stir
memories of Bar Hillel’s (1959) argument against the
very possibility of MT, namely that you couldn’t store
all the facts in the world you would need to interpret
sentences
 AI has always believed you could
 Modern empirical corpus linguistics suggest you can
find vast numbers of facts and use them.
 USC/ISI fact set currently several million….
 Is this also a way of doing classic AI “knowledge-
based understanding”on the cheap? CYC without
People…!
 Is it any different from SW thinking based on RDF
triples, except it delivers bigger numbers?
18/05/2012

The man drove down the road in a car
((The man)(drove (down the road)(in a car))))

((The man)(drove(down the road(in a car))))

18/05/2012

Remember the opening point
about the SW as data or
language:
 Berners-Lee saw the SW as a web of linked
data, and that data is being interpreted by
transformation into the RDF semi-logical form.
 Simultaneously, NLP techniques are scouring
the web for “fact triples” for tasks like question-
answering (think of Apple’s SIRI phone
assistant)—the recent work of Pantel at ISI
 These are two complementary forms of the
same thing— the one populating the other.
18/05/2012

Preliminary conclusion on
empirical results:
 Corpora for empirical semantics are now the
web itself and are so vast they cannot offer a
model of how WE do semantics, so cognitive
semantics remains an open question.
 Perhaps we can refound semantics within a
world consisting only of word tokens, but
which is not merely distributional?

18/05/2012

So much for “meaning as (real,
superhuman) use”—now back
into the cloudy stuff for a
moment:
 All the empiricism is what Sparck Jones
would call doing NLP/CL by IR methods
(AIJ, 2000)
 But what about those who want to go on
talking about concepts and meanings……?
18/05/2012

What is the way out of wanting both
data/usage and concepts?
 Maybe we must take “words as the stand” (Sparck
Jones) but perhaps not all words are equal
 Some words as aristocrats, not democrats
 Perhaps”semantic primitives” are just words but also
special words: forming a special language of
translation, that is not pure but ambiguous, like all
language.
 If that is so, perhaps we can have explanations,
innateness (even definitions) on top of an empiricism
of use.
 I would like this to be what “emergent semantics” is.
18/05/2012

In that case:
 “primitive languages = languages of primitives” may or
may not reflect the whole (game) language
 Wittgenstein seems ambiguous as to whether
(sub)parts of a language can represent the whole (see
Nirenburg and Wilks, JETAI 2001).
 Can we smuggle concepts back into a representation
by empirical methods and “privileged words”:
– Sophisticated information retrieval (time, space, colour,
number….)
– Some linguistic metadescription and innateness (over and
above the learning mechanism that even empiricists allow)
– Is this different from, say, the use of existing thesauri or
18/05/2012
Jelinek’s programme to induce all linguistic structures??

What has any of this to do with
eScience?
 T. Kazic (2006) Putting semantics into the
Semantic Web: how well can it capture biology?
Pacific Symposium on Biocomputing.

 “To sufficiently capture the scientific validity of the
SW’s computations, it must sufficiently capture
and use the semantics of the domain’s data and
computations…it mustn’t confuse reactions of
enzymes with reactions to drugs.

18/05/2012

The data is extraordinary:
 “purine salvage” reactions
 A<--> B
 C<-->D
 An enzyme Z (EC 2.4.2.4) catalyses both
but is not in class Y (purine nucleoside) and
should not catalyse them (note in KEGG
maps), as Z’ (which is in Y) should , but it
does.
 Moreover, Z promotes growth in some
reactions and inhibits it in others (a statin).
18/05/2012

The problem is how such data
can be put in a SW ontology:
 GOFAI had “default logic” but this is more
complex
 Particularly the opposite reactions in differing
contexts.
 Can this be expressed other than by writing
notes in the KEGG maps
 That is the challenge for the SW in eScience
 Otherwise it’ s all LANGUAGE annotations in
the “margins” of the structures which WE
have to interpret (contra the SW hypothesis).
18/05/2012

 Kazic: “..this implicit semantics is less effective
than an explicit, by which I mean a
computationally determinable, semantics”
 The same term in many URIs can only be
disambiguated by definitions, but they are still in
language.
 OR what “gene” means is implicitly programmed
into the code manipulation gene descriptions; I.e.
if NOT in language, then it cannot be understood
or checked by scientists.
 Two ways out:
– The SW learns to understand the language still present
(the NL solution)
– We give up worrying what the programs mean as long
as they deliver (!)
– GOFAI delivers after 40 years on a comprehensible,
comprehensive, logical formalism for the SW.
18/05/2012

Another case (FlyBase): the web
for Drosophilia studies
 FlyBase grounds gene identifiers (e.g. Rutabaga)
in the genome, and associates an ID with e.g.
regulatory regions, introns etc.
 BUT new regions are often identified for a gene
“upstream” in the genome
 Thus the referent of the gene ID changes---and
scientists live with this (I.e. what IS the gene ID of
Rutabaga??)
 Cf. Putnam on guardians of meaning, and the
shakiness of URIs in eScience. The moral is that
there are not always formal URIs outside the SW,
even in such areas of modern science.
18/05/2012

Taking stock : three views of what the SW is:
 1) An updating of the old AI dream of representing
everything in logic for reasoning over the world (GOFAI);
actually the SW is much less sophisticated than that--it has
traded representation power for tractability.
 2) An apotheosis of annotation in Information Extraction,
attempting to build up to concepts in ontologies for e.g.
scientific knowledge by very large shallow computations
over texts: problem of grounding the terms other than in
texts, and tying the general concepts plausibly to the
distributions of usage in text.
 On this view the SW is the WW of text plus meanings.
 3) A system of trusted data bases that ground meanings in
something close to objects (TB-Ls own view?) eScience?
 This is close to Putnam’s view: scientists are Guardians of
Meaning but, like all such views, is philosophically absurd.
18/05/2012

Is there a fourth view?
 An engineering-cum-Wittgensteinian-cum-
democratic view (though with some
reserved aristocratic space for primitives):
 People will use URIs as a language—
whatever its philosophical status
 And simply make its work
 In spite of its ambiguity
 In spite of the vagueness of what’s at the
end of the URIs
 Just as is the case with words themslves!
18/05/2012

Meaning and the Semantic Web

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to Meaning and the Semantic Web (20)

More from PhiloWeb (20)

Recently uploaded (20)

Meaning and the Semantic Web