SWAD-Europe: Semantic web applications - analysis and selection
Appendix B - Application Survey

Annotea requires an annotation-aware browser to work. Currently Annotea is built-in to the W3C's reference/experimental browser Amaya. When browsing a web page, Amaya will query the annotation server that the user has set up (may be local to a workgroup, or may use the public W3C server), and add a graphic to the rendered page indicating that an annotation exists. Annotations may be specific to a region of a page; regions are identified using XPointer. Once an annotation has been defined, other annotation users can reply to it, similar to a threaded email discussion.

At present information on the media accessions (technical information, rights information, provenance information, basic categorization) is stored internally using a customized database schema. This is an example large scale repository and content management problem where RDF could be applied for representation of the metadata. This might have advantages - easier extension, simpler processing of imported metadata, ability to export subsets of the metadata to support external user communities.

This specific example is included as a concrete instance of the general class of metadata for media repositories and content management systems.

The RDF level of the semantic web offers a common format for representing such loosely defined "pieces of information" and the ontology level offers a tool for representing the domain model through which such items can be indexed and structured.

Work has been done to apply DAML+OIL to the task of representing the offers and requirements and thus be able to exploit description logic reasoning engines to perform the matchmaking.

The semantic web, especially the ontology layer, could be used to describe the semantics of these data structures and the underlying conceptual models to permit a mediation layer to compose services together.

Full service composition will require other functionality, in particular the ability to mediate services with different conversation models (different message sequences).

RDF would have value in building a personal information management tool that support bibliography management. The big advantage would appear when such tools were linked together, to allow personal collections to be merged across a community while still allowing local differences in structure. It would also be valuable if the subject hierarchies themselves could be extended by the community and related to centralized term Thesauri without undue load on individual researchers. This dual management of metadata (the bibliograpies) and term vocabularies across a collaborating network of researchers does seem to illustrate many of the facets of the semantic web vision.

What makes this a challenging problem is that the thesauri change (UNSPSC changes regularly with categories disappearing, categories appearing, and entire branches moving to different parts of the tree) and the same items may need to be classified into many catalogues. The challenge is to reduce the cost of managing this. For example, a deeper ontology might enable terms to classified once and then have the classification resued by mapping the deeper ontology onto specific taxonomies as they arise and evolve.

A significant issue in business integration applications is to map different product catalogues together. A market intermediary may receive product information in several catalogue forms and yet needs to provide a common access solution for the customers. It may do this by attempting to translate all the catalogues into a common unified catalogue or by translating queries into the different source taxonomies.

The semantic web ontology layer might provide a means for capturing partial mappings between catalogues. In particular, the a common ontology mapping representation would allow such partial mappings to be discovered and reused.

• How do you extract claims from a paper (perhaps the author must do this, but then there is a higher barrier to entry).
• The ontology for scholarly discourse (claims, refutes, support etc) may not be stable over time.
• The semantic links might be time dependent - a theory may reconcile two previously apparently contradictory findings.
• A simple modeling be may insufficient: eg '[Paper] :claims [XYZ]' is not necessarily a simple (uncontroversial) ground fact
• Claims can be grouped into a set. Linking those sets may be problematic if set membership changes over time.

1. Find claims made by this paper.
2. Extend concepts (those claims positively linked) to provide an 'intellectual framework' for this paper.
3. Find claims challenging the intellectual framework of this paper.

Claims can also be collected into related 'sets' which can themselves be linked together.

Future work envisaged by the group include query (both intuitive, for the novice, and rich, for the expert) Also, visualisation of the claim network

Notes: Knowledge Media Institute, The Open University, Milton Keynes, MK7 6AA, UK {g.li, v.s.uren, e.motta, s.buckingham.shum, j.b.domingue}@open.ac.uk http://kmi.open.ac.uk/projects/scholonto/

Many different user communities (teachers, lecturers, researchers) have different needs for how to access such a collection. In particular, to index items according classifications of habitat, behaviours and species taxonomies and to link this information to other information sources (both external and developed by the communities themselves).

The key insight to here is that rather then expecting a globally accepted ontology for things like habitat, we imagine each community wanting to use (or develop or at least enrich) different ontologies based on their different needs. Specialists in slime molds have different needs from lecturers looking for materials on bird behaviours. Yet there will be situations where these communities overlap, where it would be valuable to have a common view of the same set of resources.

The semantic web could be used to implement a decentralized bookmark sharing network to support exchange of annotated and classified personal bookmarks.

The semantic web could offer a decentralized solution to the same problem in which individuals could post information on the skills, interests, and others they already work with. Tools could then walk this personal semantic web to answer such "find an expert on..." questions.

One significant example of this in the RDF world already is the "friend of a friend" schema developed by Dan Brickley and colleagues. This supports explicit "knows" links between people which is valuable both to the users (for exploring social networks) and to the tools (as a means to discovering and aggregating foaf descriptions).

The name and examples are drawn from a paper by Joshua Allen.

There is already a UK Government online portal which links the National Curriculum to available learning resources with a new version "Curriculum Online" due to go live around now. For "Curriculum Online" there is a well defined metadata scheme based on the IEEE Learning Objects Model with 59 elements, 21 mandatory. There is a defined binding for this both into both XML and into RDF. For many of the elements there are then controlled vocabularies. In particular there is a defined structure of around 2000 concepts and themes that can be mapped onto National Curriculum requirements. The XML binding just represents these as strings and has no structure, but the RDF binding creates URIs for all of them and links them into categories. There is a SOAP service for uploading and managing your metadata entries in the catalogue.

There is also an attempt to create a more widespread UK standard that can be used by web sites that might become part of the "National Grid for Learning". This extends the core Curriculum Online components. The Virtual Teacher Centre is both an example of applying this and source of the standards documents. This recommends how web sites should self-describe the content they are offering and place that description in an rdf file accessible from the site.

DMoz.org make the raw data, both structure and content data, available for free in RDF format. As of April 2002 this was a 16 million statement RDF data source containing information on some 3,005,746 web pages categorized into 428,590 categories.

A specific example of this, which has been tackled using semantic web techniques, is work by Boeing Corporation on integration of aircraft maintenance and aircraft design databases. This used RDF as a common data model, a query language on top of RDF to express the target query and a mediator architecture to translate the query both syntactically (e.g. to SQL) and semantically (to the data model of the target database). The common domain ontology was expressed using RDFS and then mapped to the different database schemas.

If the content metadata in these portals were expressed in a common form using either a shared or an interoperable set of ontologies for information categorization, then it would be possible to create an overarching virtual portal. This would give a common view of the entire communities information resources without any single group having to take the burden of creating and managing a main central repository.

EARL is an RDF application - that is, the EARL language is an RDF schema, defining a namespace and well-known properties for EARL's intended uses. A goal in EARL is to be able to declare not just the result, but the provenance for it: who claims the result, what the testing method was, when the test was run, etc. Thus the RDF statements that record the actual result of the test (e.g. page http://foo.com/bar.html passes Test01) are reified, and made the subject of earl:asserts statements.

• RDF-QEL-1: Conjunction
• RDF-QEL-2: As above + disjunction
• RDF-QEL-3: As above + negation of literals (ie Datalog)
• RDF-QEL-4: As above + recursion (transitive closure) (ie SQL3)
• RDF-QEL-5: As above + recursive (in language directly).

The P2P architecture is based on JXTA.

The replication is primarily replication of metadata (not data).

The mapping refers to mapping of vocabularies.

The mediation refers to joining of metadata. There is a separate engine called AMOS-II for this, which uses a common data model ECDM.

The annotation principle refers to the idea that people can 'annotate anywhere'.

Outside of the semantic web there is much work on event sharing and tracking. For example, there is an RSS event module and IETF standards for calendaring information (iCalendar - RFC2445, with a IETF draft XML serialization xCalendar). Apart from using RDF as a composable data format for event data the semantic web could provide the basis for categorization of events (ontologies of event types, of topics, of related events like submission deadlines) that could support richer automated processing.

Within the SWAD-E project there is activity looking at the value RDF brings to this area, see calendaring workshop.

Were each of the financial institutions to adopt a common access protocol and format such portals would be more customizable and more effective. RDF is a vendor-neutral candidate for the common format with its support for aggregation of information items and has already been used in at least one product - iSoco's GetSee (believed to be the leading Spanish Financial Portal).

The most relevant area for the semantic web is applications which deal with these annotations - manage them, use them for retrieval and for 'semantic matching' (as opposed to sequence matching)

There are also a whole range of ontologies available, the most public being the Gene Ontology (and hence there is work in ontology management).

• Gene Ontology: From the Gene Ontology Consortium.
• The TaO: An ontology developed as part of the TAMBIS project at Manchester (UK), using description logic (and the OIL editor).
• Lots of other ontologies to choose from - Riboweb, EcoCyc, Schulze-Kremer

Semantic web technologies could play several roles in this. Firstly, RDF offers a vendor-neutral format for the metadata to be managed by the publishing system - giving a common format into which the different proprietary content management system metadata can be translated. Secondly, the roles and requirements of different user communities and how they map to the different document categories could be explicitly represented by a domain ontology rather than indirectly represented by rules or scripts - thus easing maintenance and extensibility.

The Haystack work represents all structured information in a common semi-structured data format (RDF). They have a rich user interface that is driven from both the RDF data itself and from ontology or schema files defining the data models. In Haystack the type of the item being displayed drives the rendering of items and the mapping from category to view is declaratively defined. Haystack supports traditional personal information management (PIM) tasks, such as calendar and task management, as well as snippet and document discovery.

The primary semantic web related aspect to this is the use of domain ontologies to give a richer index and structuring of the case information. However, this usage is primarily within a single system and it is less clear if this is just use of ontology technology or whether there is a significant web aspect.

The second problem here of course is that not all talks may have this information. Some can be extracted automatically, using an information extraction technique (ITtalks use Lockheed Martin's AeroText system). A talk author can also be assisted in marking up a talk entry (they use a standard bag of words classifier to suggest ACM topic based on the abstract).

Some of the more speculative scenarios push on quite deep research issues in agent communication and collective decision making.

One could imagine an ideas workbench which would exploit the free-form, semi-structured data, capabilities of RDF to allow immediate capture of such data and then support incremental construction of organization and indexes to help the the user built a view of the domain. This would include tools for discovering existing relationships and classification schemes that might be relevant, ability to incrementally add relationships between ideas, tools for generating and exploring different clusterings and views onto the collected data.

At later stages in the project the collected data may be used as part of a more conventional content management system, bibliography database or community portal. However, in the early stages existing content management systems are far too rigid to offer any support.

• KAON API: programmatic access to KAON ontologies and instances
• RDF API: Provides programmatic access to RDF models.
• ENGINEERING SERVER A storage mechanism for KAON ontologies
• RDF SERVERProvides persistent storage of RDF models in relational databases,
• KAON PORTAL: A simple tool for generating multi-lingual, ontology-based Web portals.
• OI-Modeler: a tool for ontology creation and evolution.
• TEXT-TO-ONTO: Supports semi-automatic creation of ontologies by applying text mining algorithms. Currently includes term extraction algorithm, concept association extraction algorithm and ontology pruning algorithm.
• RDF CRAWLER: A simple tool for syndication of RDF-based content from a set of Internet pages.

Kaon adopts a slightly different modelling approach to UML, Flogic, OIL and RQL. In the Kaon OI model, an entity may be viewed as a context or instance in a context-dependent manner. The OI model is a 'user' level mapping (ie corresponds to the model the user has in their head), as opposed to the logical (ER) model and the physical (RDB) model. In the latter (physical) model, the scheme varies slightly from RQL (where one table is assigned per concept) for efficiency (table creation is very expensive, and the total data stored is not expected to be large). It is also different to UML (it doesn't have methods), F-Logic (it is more tractable, especially wrt axioms) and description logics, including OIL (it is more intuitive).

DSpace, like many digital library systems, has one primary data model (how items are grouped into collections) and one primary metadata model (Dublin Core). By using semantic web approaches to the metadata it should be possible to support more diverse metadata formats. In particular, to allow individual user communities to create their own augmented formats the support their specific needs.

• support multiple metadata format simultaneously
• support metadata interoperability (format mapping, provenance issues)
• use metadata to control dissemination policies
• distribution issues (performance if distributed, control if not)

Having constructed this general RDF backend it can then to used to access and display metadata from external sources. For example, Mozilla documents mention a SmartBrowsing facility that aggregates data from trusted-third party metadata sources to provide page annotations - though the status of that particular example is unclear.

There is significant interest and activity in using the semantic web in this area.

There is some suggestion that the webness of semantic web could be relevant here. For example, linking information on the same artifact/creator/topic from many sources - the AMICO case study mentioned in Eric Miller's talk (see below) is an example.

Provenancing may be an issue.

OntoShare is a tool developed by BTexact Technologies as part of the OnToKnowledge project. Practitioners use the tool to add documents to a shared community store along with manual annotations, semi-autmatic classifications into the community ontology and an automatic summarization. The tools use this metadata to recommend the new document to other community members based on their interest profiles.

The semantic web could provide an infrastructure for such recommendation networks. Exploiting URI's for representation of the products or services (an issue in itself), RDF as the representation of the product annotations and ontologies both for the types of annotation (e.g. review-based-on-direct-experience) and the nature of the product/service (to support search and selective subscription).

• How to gain critical mass? Once data on a sufficient range of products/services is usefully available this way the service becomes valuable and attracts contributions but the initial bootstrapping is hard.
• Some existing commercial sites (notably Amazon) do this for their topic area and gain value from it (customer loyalty, improved sales). The benefit to them of opening up this data in such a way that it could be exploited by competitors is not clear.
• Much of the information is aimed at people and not machines and so is largely text based and lacking in structured metadata.

Both RSS 1.0 and 2.0 are primarily designed to be lightweight publication formats. The choice of RDF properties or XML elements is chosen to represent that goal. RSS 1.0, however, has an explicit concept of modules, that plug-in to the standard to provide for more representation that is part of the core. For example, there is no provenance meta-data in RSS 1.0, but there is a Dublin Core module defined as a standard extension.

Conceptually, RSS defines channels of information, along which individual items flow. An item may contain a link to another web resource, plus various metadata and optionally content. There are no formal relationships between channels. However, it is common for blogs to be partitioned into categories, where an item might be part of multiple categories and form part of that site's overall flow.

Partners include University of Karlsruhe, VU Amsterdam, Meta4 (Spain), Empolis (UK and Poland), Dresdner Bank (Germany) and IBIT (Spain).

The digital library will provide traditional digital library services in addition to support for non-textual data types, hypermedia annotation, cross-language search and retrieval, and personalized information dissemination. This testbed will be used to demonstrate how an enhanced digital library can enable members of a networked scholarly community to learn from, contribute to, and collectively build upon the community's discipline-oriented digital collections. It will be built as an extension of the ERCIM Technical Reference Digital Library (ETRDL)

The intention is that the tool provides support for community annotation of digital media. Cross-lingual support and notification mechanisms are planned.

• CNR (Scientific Coordinator, Italy)
• INRIA (France)
• GMD (Germany)
• FORTH (Greece)
• SICS (Sweden)
• Lorasi (France)
• Univ. of Masaryk (Czech Republic)
• ERCIM (Administrative Coordinator, France)

• extracts metadata (from a variety of sources, web, database, XML feeds),
• fills in missing fields, classifies (by standard bagging/boosting methods),
• resolves ambiguities and presents to user.

The components include the use of metadata to describe syntactic (eg language/format) and semantic (domain-specific) information about a document. The normalization of ontologies involves merging concepts. Semantic search provides both disambiguation of words and a useful associative element (one gets related content not explicitly asked for).

Gartner predict this will affect $90b b2c and $350b b2b transactions by 2008 and create a new billion dollar industry to collect/organize/sell these tags 2008-20011 [0.6 probability].

This is a generic category rather than a single specific application but is more tightly defined that the categories like "metadata for description".

The last stage is already well-enough handled on the web but the first few phases of information gathering, assimilation, decision making often leave a lot to be desired. They are well handled in specific product areas (e.g. books) but not in less mass-market areas (try buying a non-standard size of "rat cage" for example) and often the best recommendation information is scattered across many specialist bulletin boards and discussion groups. There have been a large number of shopping bots developed (see botspot). However, these are usually forced to use either web-scraping or proprietary protocols to aggregate price information which results in them only existing for highly popular products and even then coverage can be poor or uneven.

The notion of the semantic web shopping assistant is to combine agent technology for seeking out and organizing the relevant information with the semantic web as an infrastructure over which the product, pricing, features, availability and opinion information could be made uniformly available.

Despite these challenges this is a genuine data integration problem of relevance to many consumers and is one of the few semantic web applications that is is reasonably compelling to non-specialists.

Developed by SUN Global Knowledge Engineering group for intranet use.

The primary encoding for such kinds of lightweight publishing and syndication is RSS (Really Simple Syndication, Rich Site Summary or RDF Site Summary). Various incompatible versions of the RSS standard exist, and its history, though short, is filled with contention and disagreement. Currently, the only version of RSS that is RDF-based (or even RDF compatible) is RSS 1.0. Note that RSS specifically depends on the XML serialisation of RDF; various RSS processors exist that treat the RDF XML syntax, and in particular the RSS schema, as their target grammar.

A specific use of RSS that has grown in popularity over recent years is blogging (from "web-logging"). Blogging is the publication of online chronlogically-organised journal entries, using tools that emphasise ease of content generation for the non-expert. As a result, a huge number of diverse blogs exist on every imaginable subject. Some are of very high quality. The phenomenon of blog-rolling arises from the referencing of one blog from another, allowing syndicated content to flow around a complex network. The blogger network is increasingly supported by underlying protocols such as the Blogger API, and SOAP.

The philosophy is that the semantic web is a about structured objects (people, places and events) and not about text strings. So that searches can use semantic terms and the things that come back can have more structure (e.g. a search for a CD could pull back information on how to buy the CD, pointers to eBay auctions, calendar details of upcoming gigs etc). The notion of semantic search or activity based search is to infer from a normal text search what the underlying user activity is and what the semantic terms involved are. Thus a free text search on "Yo Yo Ma", as well as pulling back normal Google results, would realize "Yo Yo Ma" probably refers to the classical musician and thus would also search sources like Amazon, TicketMaster etc for relevant information.

This "would realize ..." bit above is implemented by having a knowledge base of concepts and looking up the text terms in that knowledge base to see if they appear as, say, the dc:title of some concept.

At the technical infrastructure level then TAP offers a simple standard way of doing remote access to semantic web data (the GetData interface), a notion of address by reference to circumvent the problem of universal naming (similar mechanics to the ePerson QBE approach), a caching architecture, a server implementation (Apache mod) which supports aggregation, a client library and an initial knowledge base.

The sort of information to be exchanged in making such a service practical is well-suited to semantic web representation.

Semantic web ontologies could potentially support richer and more extensible descriptions. For example, supporting multiple cross-linked categorization schemes, service parameterization, description of quality-of-service and related "non-functional" features. Furthermore, description logics offer an approach to formulating queries for matching services in terms of description classes which can match a specific service against a more general constraint description.

The initial test application was a tool for storing, organizing and sharing small items of information (called the Snippet Manager). This includes tools for management of web bookmarks (bi-directional access to browser bookmarks, RDF queriable copy of the DMOZ data set, ability to share bookmark classification schemes) but is also able to store arbitrary web-accessible content and associated searchable metadata. RDF is used for all information storage, an RDF query paradigm "query by example" is used for all distributed searches, DAML+OIL is used to represent the schema and vocabularies for both internal machinery and user visible data.

Such a proposal has clear requirements for data integration, intelligent query and metadata use that would seem to provide a good fit for semantic web technologies.