12th-13th September, 2005. Hosted by the British Computer Society, London [PDF]

Page 1
Web Science Workshop 12th -13th September, 2005
Hosted by the British Computer Society, London
Executive summary
On 12thand 13th September, 2005, the British Computer Society hosted a workshop
on Web Science, at their headquarters in Southampton Street, London. Attendance at
the workshop was by invitation only, with 21 participants. The workshop was chaired
by Sir Tim Berners-Lee, Wendy Hall and James Hendler.
There were three types of sessions on the agenda: lynchpin talks from invited
speakers; panels in which short presentations were followed by discussions; and open
discussion sessions.
The purpose of the workshop was to explore what the critical research challenges are
for the future of the World Wide Web, and to outline what major breakthroughs will
be needed to let its exponential growth continue into the future. It was structured to
consider fundamental questions in a range of Web technologies including:
• The networking needs for continued scaling of decentralized technology
• The scientific challenges of managing and searching for information in the
growing Web space
• The means for managing the space of concepts and semantics needed for
understanding and processing the emerging Web of data
• The challenges that arise when attempting to discover, invoke and compose a
globally distributed and constantly evolving set of services
• The development of peer-to-peer and other architectures needed for continued
Web scaling
• The interface needs for exploring and using the increasingly complex
information space that is the future Web.
• Social, cultural and legal challenges raised by the evolving Web and means by
which computer scientists and others can shape technology to address such
Photographs of the event have been posted by Ora Lassila and James Hendler.
Lynchpin talk by Hal Abelson
Panel: Representation (or not) on the Web
Lynchpin talk by Robin Milner

Page 2
Web Science Workshop Report
Panel: Public policy and architecture
Lynchpin talk by Yorick Wilks
Panel: Large scale decentralised (information) systems
Lynchpin talk by Tim Berners-Lee
12th September
• Tim Berners-Lee (chair)
• Wendy Hall
• James Hendler
The participants were welcomed to the workshop. Berners-Lee discussed his
expectations for the scientific agenda over the next 10 years, and described the task in
hand, as we move from tree-based information structures to web-based ones, as
philosophical engineering.
Hall said that Web Science had grown out of work to develop the Semantic Web
(SW), and the need for an intellectual overview of web-like relationships. The aim is
to set up a new discipline, or at least to bring people together in a new
interdisciplinary space.
Hendler announced that the workshop should result in a white paper, to be written by
Hendler, Shadbolt & Weitzner.
Lynchpin talk by Hal Abelson
Hal Abelson’s talk began from the observation that statistical methods and machine
learning tended to produce better results than reasoning from ontologies, and set out 5
challenges to the SW community.
1. We need real applications, not demos that work in the lab.
2. How do we communicate the ideas without being mired in syntax? We don’t
want new languages, new working patterns, or new panaceas.
3. How do we make ontologies robustly outside the lab, with real-world
constraints, on the web scale?
4. Knowledge is not all classificatory, and is not always best modelling by
ontologies. How should alternative methods of representation and reasoning
be incorporated?
5. Can the Web be made more useful by incorporating human-like reasoning?
In the discussion, various types of reasoning or types of knowledge were discussed.
Yorick Wilks recommended information retrieval techniques. Henry Thompson
pointed up the need for metadescription, and warned that that made inference harder.
Carole Goble said that representations of workflow, rather than ontological
knowledge, were needed in her area of e-science.

Page 3
Web Science Workshop Report
Berners-Lee argued that large, complex, internally-consistent ontologies were not
necessarily required on the SW. Actually, we can generally employ small overlapping
ontologies that can be sewn together, and which are consistent with each other in the
requisite areas. Guus Schreiber noted that it was important for advocates of the SW
not to overclaim.
Panel: Representation (or not) on the Web
• Nigel Shadbolt (chair) – position paper
• Carole Goble – position paper
• Craig Knoblock – position paper
• Guus Schreiber
• Henry Thompson – position paper – slides
Shadbolt argued that content is primary, and that we must avoid overanalysis.
Lightweight ontologies work, while distributed representations (cf. Brooks) are
compelling. Labels must work stably for such ideas to work.
Goble has been working in molecular biology, which produces too much information
for web formats to cope with. What does a community want to share? There are lots
of ontologies about, but sharing knowledge demands workflows which are key for
knowledge dissemination.
Knoblock looked at the issue of bootstrapping the Web from syntax to semantics. We
need simple tools and immediate payoff. We need personalised integration tools and
need integration frameworks which allow end users to build applications.
Schreiber discussed representation and consensus. In practice, one needs to work with
pre-existing representations that may be problematic. Major research issues include
partial alignment techniques, the semantics of visual representations, and reasoning
under inconsistency.
Thompson discussed the requirements of a science of the Web: hypotheses,
experiment, falsifiability, consensus methods and standards for measuring quality. As
regards the Web, the key question is how the SW differs from AI, and whether
analytic techniques can outdo, or even match, statistical techniques.
In the discussion, Berners-Lee made the point that underlying much of the potential of
the SW is the large amount of data in structured relational databases, which the SW
will be able to exploit. Wilks, Thompson and Hendler underlined the difficulties in
achieving consensus. Hendler and Goble pointed out that automatically generated
ontologies and hand-crafted ones had different properties, and worked in different
domains and tasks: one task for Web Science would be to create the use cases to link
them. Robin Milner and Thompson highlighted the need to represent provenance.
Lynchpin talk by Robin Milner
Robin Milner’s talk described a multi-layered structure for Web Science, by analogy
to the natural sciences’ three levels of physics, chemistry and biology. At the lowest
level would be the elements of computational interaction, basic processes such as
naming or synchronised actions. In the middle would be non-trivial structures such as
agents, data structures or workflow patterns. At the upper level would be complete
individual systems of software. Software science should be about description as well

Page 4
Web Science Workshop Report
as prescription. Software engineering, in contrast, is about processes of development,
not about what software actually is, because languages had gradually become
divorced from theories in computing. One aim of Web Science should be to keep
languages and formalisms in step with theory. We are in danger of understanding
complex biological systems better than wholly engineered computing systems.
In the discussion, Wilks suggested parallels with Milner’s layered structure with
Braithwaite’s notion of semantic ascent and supervenience of layers. Issues
surrounding reductionism were discussed, including the effects of the Web scale, and
of having humans in the loop.
Panel: Public policy and architecture
• Daniel Weitzner (chair) – slides
• Kieron O’Hara
• Joan Feigenbaum – position paper
• Mark Ackerman
• Jonathan Zittrain – position paper
Weitzner noted the importance of debate in this area. Architecture constrains public
policy in various ways. It allows copying of information, for instance, and allows
anonymity. Similarly, policy has been poorly understood. Two important concepts are
transparency and accountability. Can we build large scale systems that include
transparency (the history of data uses) and accountability (the ability to check whether
the policies that govern data manipulations and inferences were in fact adhered to)?
Can you prove that you have used the information correctly?
O’Hara noted the difficulties in applying political and philosophical constructs to
online spaces, which are analogous to but different from offline spaces. There is also
an important distinction between normative and de facto accounts – how are these to
be related?
Feigenbaum argued that personal information should be termed ‘sensitive’ not
‘private’. The focus should shift from hiding information to saying what misuse is.
How can you balance consultation and correction with automation and not bothering
people all the time? On copyright, creative commons isn’t a total solution, as it
doesn’t work for legacy content.
Ackerman discussed the issue of how things become resources. How can you make
things that can be incorporated into people’s lives? It is hard to understand what
people do and what the intent of systems is in that context.
Zittrain discussed the notion of subversive and enabling technologies. For instance,
RSS is an enabling technology, while the URL is a subversive idea, making every
web address equal and one click away. Rules about privacy or copyright on the Web
would have been wrong at the beginning, because it is hard to predict how people use
In the discussion, a number of points were made. Most privacy policy is concerned
with government intrusion on the citizen’s privacy, although there are other types of
intrusion. An important issue, highlighted by Zittrain, Weitzner, Hendler and Ora
Lassila, is that sites such as Google are bottlenecks, and the functioning of the Web
depends to a large extent on their good behaviour. But often it is the badly-behaved

Page 5
Web Science Workshop Report
(hackers, illegal downloaders, spammers) who put in the effort to understand the
The summary discussion of the first day’s work noted a number of issues.
• Transparency. Cookies or redirects make the Web less transparent for the user.
Links are important invariants of the Web experience, but they are being
undermined in various ways. Preserving such invariants is important for trust.
• Metadata are important. Does the SW knowledge representation architecture
need a coherent story about provenance etc?
• How do we link the technical architecture with the human/social effects that
These three issues are linked by suggestions such as e.g. time-stamping code
encapsulated by a URI, to make it invariant. Changes could be noted in the metadata,
in that case – how would this affect reasoning? TAG at W3C has discussed such
matters in some detail.
• One other opportunity comes from technology to map knowledge, and spot
trends and developments.
13th September
Lynchpin talk by Yorick Wilks
Yorick Wilks’ talk focused on the relationship between NLP and the SW, in the wider
context of the relationship between language and knowledge representation. If the SW
is, or depends on, the traditional AI knowledge representation task, then there is no
particular reason to expect progress in this new form of presentation as all the
problems and challenges of logic reappear and it will be no more successful outside
the narrow scientific domains where KR seems to work and the formal ontology
movement has brought some benefits. Alternatively, if the SW is the WWW with its
constituent documents annotated so as to yield their content or meaning structure, then
NLP/IE will be central as the procedural bridge from texts to KR. This view was
discussed in some detail and YW argued that this is in fact the only way of justifying
the structures used as KR for any SW.
In the discussion, Berners-Lee argued that the SW rests not on NLP but logic. Logic
and ontologies will suffice to extract much of the value from the data held in
structured relational databases. Wilks responded that the unstructured (legacy) part of
the Web needs NLP for annotation, even in other media, at least until visual
recognition systems become more reliable.
Panel: Large scale decentralised (information) systems
• James Hendler (chair) – position paper
• David de Roure – position paper
• Ora Lassila – position paper – slides
• Dieter Fensel – position paper – slides
• Wendy Hall – position paper

Page 6
Web Science Workshop Report
Hendler focused on the information that is not on the Web. Automatic annotation is
important, but basically tells the machine what it already knows. Getting humans into
the annotation loop is important, involving providing applications into which people
are willing to put some work. Semantic enabling will allow unifying approaches to
creating, browsing and searching text, and integrating information that is not stored in
a centralised way.
De Roure discussed the relationship between the SW and grid computing, as a method
of getting to a world where large quantities of data were open to large amounts of
computational power. There is an important need to get semantics into the
middleware, and to allow scientists to formulate and answer more complex queries.
Lassila noted that sharing across application boundaries is very difficult. It is hard to
anticipate all the application pairings. We need to focus on the information, not the
application. Also, implicit information is inaccessible. Ontologies and reasoning
services are needed for such things. Also SW concepts are useful for mobile and
ubiquitous computing, for similar reasons: awkward usage situations and
unanticipated encounters make application-centred approaches difficult.
Fensel argued that engineers relate the usage of resources to achievement of goals.
We can abstract from hardware, and now we are abstracting from software, using
service oriented architectures (SOA). SOAs will not scale without mechanising
service discovery, etc, so machine-processable semantics are required for services.
Later we may even talk of problem-oriented architectures, focusing on the customer’s
problem, not the provider’s service.
Hall discussed the importance of linking, and interacting with big information spaces,
multimedia environments, etc. We need to link things, add metadata to archives, put
links in video, etc. The current web is mostly linkless, and we need to make it better.
A query is in effect a non-existent link. We need the web to disappear behind the
In the discussion, it was agreed that open-ended architectures are essential, but
complicate the task of designing the next generation of the Web. Semantic
specifications of, e.g. services are still an open research question, which makes the
task of working in such a distributed environment very complex.
Lynchpin talk by Tim Berners-Lee
Tim Berners-Lee’s talk highlighted different phases of information structure, moving
from characters, to linear structures, to trees and hierarchies, and lastly webs, which
can grow without a central bottleneck. The goal of a web is serendipitous reuse, but a
minus is that it comes with global obligations, such as maintaining Web content,
which are important for allowing the serendipity to happen. The SW is a web of logic,
very different from hypertext. We need the same standard for symbol space as the
WWW. We need to be able to map URIs to anything. Looking up URIs is still the
critical architecture.
Berners-Lee also offered some comments on the NLP/SW debate stemming from
Wilks’ lynchpin talk, arguing that the two are very separate.
Terms of logic

Page 7
Web Science Workshop Report
Meaning is use
Meaning is defined in words, or code,
or specific use
No ownership of words
URI ownership
Defining words in ontology is never complete
and a waste of time
Defining terms is never perfect but
NL is constantly changing
Ontologies are basically static
Can’t benefit from injecting logic
Can’t benefit from cloudy statistics
Machine finds stuff
Machine infers stuff
Following on from this, some further misconceptions of the SW were set out.
• Is it RDF “syntax”? No –XML etc can also be used.
• Is it OWL? No, there are alternative ontology representation languages.
• Is it built on NLP extraction? No.
• Does it depend on manual annotation of legacy content? No.
• Actual SW data deployment will be mainly on existing RDBs? Yes.
• I will have to do everything? No, it will be a collective effort. The effort of
developing ontologies will scale. If we model the effort of building ontologies,
relative to the scale of the organisation, number of users and number of
cooperating builders, the effort per person of ontology building grows very
small as the scale increases.
• Is it slow? No – many of the prototype systems are not optimised.
Furthermore, simply dumping RDF into a triplestore isn’t going to solve
information processing problems.
• Everyone will have to mark up web pages to generate content? No.
• Management of data will have to change? No.
Finally, Berners-Lee outlined some important large-scale phenomena and discoveries.
• Eigenvector analysis is wonderful: no-one assumed Google would be possible.
• Plume tracing. New trends can be studied to find originators of trends. This
technology should be available beyond government.
Issues to do with decentralisation.
• How decentralised is the web, given for example the US domination of the
• We want to encourage P2P experience. Could we make this an HTTP
fallback? When servers are unavailable, P2P could locate cached pages,
removing the dependence on DNS, to make the web even more decentralised.
Issues to do with user interfaces.
• The ultimate SW browser would be every application rolled into one.

Page 8
Web Science Workshop Report
• Speech and NL interfaces to SW.
• Security/trust models, rules and policies. UI issues are the tricky ones here. It
is easy to write a trust policy, but hard for a neophyte to do it.
In the discussion, the argument about the relationship between NLP and the SW
continued, with Wilks arguing that on Berners-Lee’s view there would be serious
problems with the fuzziness of terms (in domains such as law, for example), and only
certain areas of very hard science would be amenable to the SW approach, and
Berners-Lee responding that some areas of law are relatively non-fuzzy, such as
regulations. On Wilks’ view, even ownership of URIs allowing the retrieval of
relatively definitive information would not remove the fuzziness, and would also
create the problem of trusting the institutions managing the URIs.
On the topic of trust, Hendler pointed out that trust is a problem throughout the
WWW – authority cannot be created, and people make more or less informed
judgements. Understanding the semantics of trust is useful but not necessary.
Thompson worried that on the SW it is the computer that makes the judgement on the
basis of metadata, not a person. Berners-Lee argued that it was another misconception
about the SW that most applications would draw information from a wide range of
unknown information sources (thereby exacerbating the problem of trust); in fact
many applications will draw information from a small number of well-known and
trusted sources. Joan Feigenbaum detected an issue for Web Science, which is that the
architecture is so open that we lack computational models for investigating trust
The final discussion section centred on two topics, architecture and the relation
between the Web and society.
The discussion on architecture, chaired by Shadbolt, followed an issue that had been
prominent in the workshop, together with representation, decentralisation, service
orientation, inference and large-scale structure.
Questions were raised by Fensel and Goble about whether the discussions previously
had focused too much on browsers, whether demand for browsers (e.g. in e-science)
was high or low, and whether an SW browser was vital. Berners-Lee argued that a
generic browser looks attractive, but because of the richness and diversity of
applications integrated by the SW, there will be a great variety of ways that people
will want to interact with it.
Milner introduced the orthogonal topic of process: can we replace procedure names
with URIs? What SW platforms need to be in place, and what can be developed by
users? Berners-Lee contrasted the Web as a (nearly) static information space and Web
services as a different protocol. Milner posited another type of Web Science looking
at interaction calculi, such as Choreography or process algebras.
Various representational requirements were discussed, including probability,
coreference resolution, persistence and process. Thompson raised the strategic
question of whether there are things that OWL + RDF don’t do that other KRLs can.
Milner noted that P2P needs private links between parties, and so models of

Page 9
Web Science Workshop Report
decentralisation were needed. We don’t have the notion of an abstract process, as we
do of computation.
Further issues concerned negotiation, agents and trust; Hendler and Hall argued that
the agent trust literature addressed different questions, and for the SW large-scale
decentralised trust mechanisms were needed. Goble argued that the grid community
was a better source of insight into trust than the agent world. Weitzner argued that
there is much about trust that can’t be modelled totally.
Web and society
The discussion on the Web and society was introduced by the chairman, Weitzner,
reviewing earlier discussions on public policy questions of privacy and copyright, and
arguing that we need to give people control over how they interact with information
resources, while adding transparency. Feigenbaum argued that questions of
accountability and transparency are linked: transparency enables accountability. It
was generally agreed that it was possible to write rules to cover policies and policy
violations, but that a problem was how to get people to use such rules.
Feigenbaum also noted that accountability was controversial, particularly at high
levels of abstraction. Alternative goals, such as fairness, could be pursued. Weitzner
distinguished between accountability and enforcement, and noted that the pursuit of
particular goals needed experimental and empirical work to address the question: what
happens to social interactions with particular architectures?
Berners-Lee and Milner discussed the use of process calculi to model small parts of
systems, as a potential method for helping with these social questions.
Goble noted that any kind of social regulation infrastructure would have to be
lightweight. Feigenbaum suggested that many aspects of the infrastructure, such as
identification and authentication, aren’t hard to use, although Hendler cautioned that
controlling identities is hard in distributed open systems. Weitzner noted the useful
property of the general SW architecture that allows generic high level rules covering
social interactions to be written.
Ben Shneiderman was unable to attend, but contributed a position paper.