Step 1. Generate a thesaurus by using a linked data source like DBpedia

As recently reported SWC has developed a tool called SKOSsy which can be used to extract seed thesauri from DBpedia. In our example I will generate a knowledge model describing the domain of “digital photography“. This step took around 15 minutes.

Step 2. Load the thesaurus into PoolParty and improve it to your needs

After the seed thesaurus has been loaded into PoolParty Thesaurus Manager you have many possibilities to enhance the knowledge model further: Add more categories, synonyms, relations etc. In this example I use the seed-thesaurus without any further improvements. This step took approximately 2 minutes.

Step 3. Generate an automatic text extractor on top of your thesaurus

This step took a couple of seconds and ended up in having generated a fast and reliable text mining application on top of PoolParty Extractor, ready to be used to enrich your documents with data from the LOD cloud.

You can try it out here: PPX Live-Demo

To try the extractor on your own, please take a look at the image above which shows a proper configuration, you have to insert the following UUID in the form: d35d4ddb-adc3-4ea5-b027-deacac03e391

Since our example is all about ‘digital photography’, we recommend to use text samples (or some fragments) like these ones to test the quality of PPX based text analytics:

• Digital Camera Image Noise (Results as HTML, RDF/XML)
• Nikon D3S In-depth Review (Results as HTML, RDF/XML)
• Introduction to Shutter Speed in Digital Photography (Results as HTML, RDF/XML)
• Digital Camera Sensors (Results as HTML, RDF/XML)

Let us know what you think about this straight-forward approach and your opinion about the quality of the results. We believe that thesaurus based text mining is in many cases an alternative to some other approaches, especially if you want to to enrich your content with information from the upcoming web of data.

Of course we would be happy to generate other demos in the areas of your interest! Just get in contact with us by using our contact form.

• automatic annotation of all blog entries displayed as tooltips
• a comfortable search facility with auto-complete over all concepts from the linked thesaurus including semantic search over the whole blog
• an integrated thesaurus browser, plus
• a corresponding linked data frontend including RDF/XML serialization of the underlying thesaurus + SPARQL endpoint

All details about the new version 2.2.3 can be read here.

Related articles

• Using DBpedia to generate SKOS thesauri (ablvienna.wordpress.com)
• PoolParty 3.0 and its all new Linked Data framework (semantic-web.at)

Let me introduce you to SKOSsy. I will explain what this web service can do for you:

SKOSsy generates SKOS based thesauri in German or in English for a domain you are interested in. Not any domain but nearly any: SKOSsy extracts data from DBpedia, so it can cover anything which is in DBpedia. Thus, SKOSsy works well whenever a first seed thesaurus should be generated for a certain organisation or project. If you load the automatically generated thesaurus into an editor like PoolParty Thesaurus Manager (PPT) you can start to enrich the knowledge model by additional concepts, relations and links to other LOD sources. But you don´t have to start in the open countryside with your thesaurus project.

Let me give you an example: Imagine you are working for a company which is an international plant builder and you would like to index several thousands of documents the “semantic way”. You have to walk through the following steps:

1. Identify proper categories in Wikipedia/DBpedia which describe best what your business or your domain is all about. Those categories should contain pages / resources which are related to the documents you would like to index. For example: http://dbpedia.org/resource/Category:Metalworking or http://dbpedia.org/resource/Category:Industrial_automation
2. After you have selected proper categories SKOSsy will traverse DBpedia for you and collect all resources, their hierarchical and non-hierarchical relations, alternative labels, definitions and other properties and put them together as a valid SKOS thesaurus; this step will last a couple of minutes. (Find the resulting vocabulary here)
3. Load the resulting thesaurus into PPT, explore it, improve it and enrich it with additional facts.
4. After you´re done you can generate a tailor-made text extractor by using PoolParty Extractor (PPX) which is the second component of PoolParty product family
5. With PPX and its extraction model especially curated for your special use case you can extract named entities from your documents automatically and index your documents in a meaningful way.
6. After a few seconds your semantic search engine is ready to be used. PoolParty Semantic Search (PPS) which is the third PoolParty component will offer some nice facilities like categorized auto-complete, faceted search, content recommendation (similarity search) and smart search suggestions to ease your life as a knowledge worker.

We have constantly discussed the application of thesauri and other knowledge models to improve search over the last years. Many people understood straight away why thesaurus based search is most often much better than search algorithms purely based on statistics. Of course the big contra always was, “the costs are too high to establish a “good-enough” thesaurus or even a “high-quality” one”.

With SKOSsy in place those kinds of arguments become weaker and weaker. To sum up,

• SKOSsy makes heavy use of Linked Data sources, especially DBpedia
• SKOSsy can generate SKOS thesauri for virtually any domain within a few minutes
• Such thesauri can be improved, curated and extended to one´s individual needs but they serve usually as “good-enough” knowledge models for any semantic search application you like
• SKOSsy based semantic search usually outperform search algorithms based on statistics since they contain high-quality information about relations, labels and disambiguation
• SKOSsy works perfectly together with PoolParty product family

If you are interested in the results produced by SKOSsy, just send us a short note about your domain or your project and we will send you an invitation as beta-tester or prepare a demo for you.

Related articles

• Geological Survey Austria launches thesaurus project (semantic-web.at)
• PoolParty 3.0 and its all new Linked Data framework (semantic-web.at)
• PoolParty DemoZone Content Extractor Semantic Search Thesaurus Manager (poolparty.punkt.at)
• Query DBpedia for multiple keywords (stackoverflow.com)

In PoolParty 3.0 we created a Linked Data interlinking editor, making it easier than ever to add your own lookup and interlinking services (even for non-RDF sources) and made the Linked Data publishing front-end fully customizable in design, layout and regards to which parts of your content will be displayed.

But let’s start at the beginning:

Step 1 – Hook into the Linked Data Cloud!

In the era of the rapidly growing Linked Data Cloud your knowledge models don’t need to stay isolated from the outside world anymore. Simply use PoolParty’s new and improved lookup service to find matching resources from the Linked Open Data Cloud (e.g. from DBpedia).

Imagine having different data models that all refer to the same product categories and world regions. Once you have them represented in PoolParty you can use its lookup service to find matching resources from the Linked Data Cloud. In this way you will get globally used identifiers for your product categories and regions, usually in the form of a URI like http://dbpedia.org/resource/Berlin. This eases your internal data integration efforts, and it can aid the data exchange with partners or customers and enables hassle-free distributed management of knowledge models.

Image 1: Lookup of concept ‘Austria’ and selection of properties and values to be imported

With PoolParty 3.0 we increased the number of included lookup services: DBpedia, Geonames, Wordnet, Umbel, Yago, Freebase, Sindice, dmoz and LCSH – BBC Wildlife, Enis and Gemet are available on request.

Step 2 – Pull in Semantic Data!

There is a vast amount of Linked Data out there just waiting to be leveraged for thesaurus creation and extension. To meet that end we had a close look at our interlinking module and decided to enhance it a way that it becomes more of a Linked Data editor.

Once you have a base thesaurus in PoolParty and hooked a couple of your concepts into the cloud as described above, you can proceed to pull in the good stuff that comes with the Linked Data resources you have found.

Image 2: Imported Linked Data for concept ‘London’

As you can see in the image above, you can extend your local thesaurus with labels, definitions and all kinds of other information like e.g. in the case of countries their population, GDP, spoken languages, famous people born there, newspaper articles related to the political situation, and so on.

Now PoolParty 3.0 takes this approach a couple of steps further. You can not only specify which of your local concepts corresponds to which Linked Data resource and grab all semantic information that comes with this resource, but now you are able to selectively pick out the data items you are interested in and even transform the predicates they originally came with. Just switch them to whatever custom properties you created or want to re-use from any ontology (see an example in Image 1).

In this way you can easily enrich your own knowledge models with external information – which in turn can be utilized for better content recommendation, easier data integration and improved search services.

Step 3 – Publish your Linked Data in Style

Previous PoolParty versions already offered the possibility to instantly publish your thesauri, taxonomies or vocabularies and display their concepts as HTML while additionally providing machine-readable RDF versions for them. This means that anyone using PoolParty intuitive GUI can become a W3C standards compliant Linked Data publisher without having to know anything about Semantic Web technicalities.
Of course you don’t need to publish all your valuable models, just choose the parts that safely can be shared with the public and keep everything else behind your firewall, available only to you and trusted partners!

In this new release of PoolParty the design of all pages on the Linked Data front-end is now under your full control. You can use your own style sheets and create views on your data with velocity templates. It is even possible to develop project- and thesaurus-specific templates and layouts, so they can have an individual look and display different predicates and their values.

Take a look at PoolParty´s standard linked data frontend!

The following images show a PoolParty default Linked Data page and a custom-made Linked Data page of a PoolParty concept that has some DBpedia info imported.

Image 3: PoolParty default Linked Data page

Image 4: Custom Linked Data page of ScOT thesaurus (courtesy of Educational Services Australia)

Step 4 – Unlock new Linked Data Sources

With PoolParty 3.0 you are in no way limited to DBpedia, Freebase, Geonames and the other lookup services that PoolParty provides out of the box: you can add your own non-Semantic Web data sources to the mix, thereby enabling you to boldly go where no Linked Data tool has gone before.

Maybe you have a product thesaurus and want to specify which products are related to patents that can be found with Google Patents?
Or you want to interlink concepts from a company taxonomy with related articles from the Guardian’s search service or any other newspaper that provides a search API?

All those sources are not available as RDF, so how can you re-use them easily as data sources for Linked Data style interlinking? For such cases PoolParty introduces the Unified Lookup API, which makes it easy to turn almost any third party Web API into a source for interlinking your concepts with third party resources as described above.

This makes it possible to interlink your concepts with many kinds of data out there, be it New York Times articles, UN data, synonym services, abbreviations, press releases, juridical information – or any web API important for your knowledge domain.

That being said, if you have suggestions for additional lookup services that you think are interesting, let us know!

To gain a first hand impression of the new PoolParty just apply for a demo account!

How to model divergent views on one concept?

Such a central model is not only much harder to manage, but there is also the general problem that differerent departments like marketing, quality assurance, R&D, etc. will have divergent views on the model and its concepts. These different perspectives on one and the same concept are hard to unify in a single model.

Think of a company that sells mobile phones and wants to create a model of its line of products. It wants to utilize this model in the context of its online shop as well as in the context of its user support forum. While the structure of the model (i.e. the relationships between the products) might be very similar or the same in both contexts, there will be differences in which properties of the products are actually relevant in the respective contexts.

In the model of the marketing department there might be a concept for a “Phantastax StamiMaxx” cell phone with a definiton “The StamiMaxx has a powerful battery and is great for professionals who travel a lot”. They might relate it to manufacturer “ACME Corporation” and to several concepts representing different features like “Android OS”, “Multi-touch touchscreen”, etc.
The very same phone has different properties that are interesting from the Quality Assurance departement’s perspective. They might call it by a more specific name like “Phantastax i3000 StamiMaxx S”, have a different definition for it like “3G cell phone implementing the new WTF3000 protocol, …” and relate it to concepts representing known problems and their solutions.

Now they face the task to integrate these different models, as it is not desirable to use a bunch of isolated models within one company.

Support of collaborative work on distributed models

To support this kind of collaborative work on distributed knowledge models, we would like to link the concepts of the models, just as is we link documents in the World Wide Web. Fortunately the Simple Knowledge Organisation System (SKOS) offers mapping properties that can be used to define relationships between concepts from different knowledge models.

E.g. when we want to say that concept “Phantastax StamiMaxx” in the product line thesaurus refers to the same real world entity as concept “Phantastax i3000 StamiMaxx S” in the Quality Assurance thesaurus, then we can use skos:exactMatch to express that. If we want to express that the concepts are merly similar, skos:closeMatch could be used.

The other SKOS mapping properties express a hierarchical (narrowMatch, broadMatch) or an associative (relatedMatch) mapping relation between concepts from different concept schemes. With those we can say that my Samsung Galaxy concept has a skos:broadMatch “Smartphone” in the product line vocabulary and a skos:relatedMatch “ACME Corporation” in a controlled vocabulary about Tech companies.

Modularisation of knowledge models

In this way SKOS thesaurus management systems like PoolParty make it possible to modularise knowledge models, represent concepts in their different contexts and consequently enable collaborative work on those models: The marketing guy can work on his model with the concept properties focused on sales without disrupting the work of the quality assurance expert on her own thesaurus. Later one or both of them can create the skos:exactMatch link between the concepts that are the same, like seen in the “Exact Matching Concepts” box in screenshot of PoolParty below.

Enrich your knowledge: Get connected with the LOD Cloud

Going a step further the models could be connected to external knowledge, e.g. a source from the Linked Open Data (LOD) Cloud. Once we establish links to LOD hubs like DBpedia, we can import additional information for their concepts or use it to establish whether similar concepts from different models really refer to the same real world resource.

Purpose

Poolparty was conceived to facilitate various applications like

• Semantic search engines
• Recommender systems (similarity search)
• Corporate bookmarking
• Annotation- & tag recommender systems
• Autocomplete services and facetted browsing.

These use cases can be either achieved by using PoolParty stand-alone or by integrating it with existing Enterprise Search Engines and Document Management Systems or Enterprise Wikis.

Thesaurus Management

PoolParty is aiming to be easy to use for people without a strong Semantic Web background or special technical skills. The GUI is entirely web-based and utilizes AJAX so the user can e.g. quickly merge two concepts via drag & drop. An overview over the thesaurus can be gained with a tree or a graph view on the concepts.

PoolParty also helps to semi-automatically add concepts to a thesaurus as it can be used to analyse documents (e.g. web pages or PDF files) relevant to a thesaurus’ domain in order to glean candidate terms. This is done by the key-phrase extractor of KEA. The extracted terms can be selected by the user, thereby becoming “free concepts” which later can be integrated into the thesaurus, turning them into “approved concepts”.

Documents can be searched in various ways – either by keyword search in the full text, by searching for their tags or by semantic search and similarity search. The latter takes not only a concept’s preferred label into account, but also its synonyms and the labels of its related concepts are considered in the search. The user might manually remove query terms used in semantic search. Boost values for the various relations considered in semantic search may also be adjusted. In the same way the recommendation mechanism for document similarity calculation works.

PoolParty by default also publishes a Semantic Wiki version of its thesauri, which provides an alternative way to browse and edit concepts. Through this feature anyone can get read access to a thesaurus, and optionally also edit, add or delete labels of concepts. Search and autocomplete functions are available here as well. The Wiki’s XHTML source is also enriched with RDFa, thereby exposing all RDF metadata associated with a concept to be picked up by RDF search engines and crawlers. (See two examples: Cocktail thesaurus &  Standard Thesaurus for Economics)

PoolParty also supports the import of thesauri in SKOS (including several consistency checks) or Zthes format. Those functionalities can also be consumed as stand-alone web services via PoolParty SKOS Services. Additionaly, lists of concepts and their labels can also be imported via CSV files.

Linked (Open) Data

PoolParty not only publishes its thesauri as Linked Open Data (in addition to a SPARQL endpoint), but it also consumes LOD in order to expand thesauri with information from LOD sources.

Concepts in the thesaurus can be linked to e.g. DBpedia  via a service like Georgi Kobilarov‘s DBpedia lookup service, which takes the label of a concept and returns possible matching candidates. The system suggests relevant resources from DBpedia and the user can select the one that matches the concept from his thesaurus, thereby creating a skos:exactMatch relation between the concept URI in PoolParty and the DBpedia URI. The same approach can be used to link to other SKOS thesauri available as Linked Data.

Other triples can also be retrieved from the target data source, e.g. the DBpedia abstract can become a skos:definition and geographical coordinates can be imported and be used to display the location of a concept on the map, where appropriate. The DBpedia category information may also be used to retrieve additional concepts of that category as siblings of the concept in focus, in order to populate the thesaurus.

PoolParty is capable of importing a SKOS thesaurus from a Linked Data server, and may also receive updates to thesauri imported this way. This feature has been implemented in the course of the KiWi  project funded by the European Commission. KiWi also contains SKOS thesauri and exposes them as LOD. Both systems can read a thesaurus via the other’s LOD interfaces and may write it to their own store. This is facilitated by special Linked Data URIs that return e.g. all the top-concepts of a thesaurus, with pointers to the URIs of their narrower concepts, which allow other systems to retrieve a complete thesaurus through iterative dereferencing of concept URIs.

Additionally KiWi and PoolParty publish lists of concepts created, modified, merged or deleted within user specified time-frames. With this information the systems can learn about updates to one of their thesauri in an external system. They then can compare the versions of concepts in both stores and may write according updates to their own store.

This means each system decides autonomously which data it accepts and there is no risk of a system pushing data that might lead to inconsistencies into an external store. Data transfer and communication are achieved using REST/HTTP, no other protocols or middleware are necessary. Also no rights management for each external systems is needed, which otherwise would have to be configured separately for each source.

Technology

The software is written in Java and utilizes the SAIL API, so it can be used with various triple stores. The thesaurus management itself (viewing, creating and editing SKOS concepts and their relationships) can be done in an AJAX Frontend based on Yahoo User Interface (YUI). Editing of labels can alternatively be done in a Wiki style HTML frontend. For key-phrase extraction from documents PoolParty uses a modified version of the KEA 5 API, which is extended for the use of controlled vocabularies stored in a SAIL Repository (this module is available under GNU GPL). The analysed documents can be stored and indexed in Lucene/Solr or any other (enterprise) search system along with extracted and semantically related concepts.

Jamie Taylor, Metaweb

Andreas Blumauer from Semantic Web Company (SWC) talked with Jamie Taylor, Minister of Information at Metaweb Technologies Inc. about Freebase & Linked Data and Google´s announcement to use RDFa.

SWC: At ISWC 2008 Freebase became “officially” part of the LOD Cloud. What exactly has changed since that time?

Jamie: Since Freebase is a community writable semantic database, the addition of the RDF interface allows anyone to publish data into the LOD cloud. LOD Applications can access any Freebase Topic through the RDF interface by constructing a URI from the Freebase identifier.  But perhaps more importantly, because entities in Freebase can be annotated with multiple identifiers, Freebase Topics can be retrieved by constructed URIs using the identifiers used by other systems and data sets.
For instance, the movie Blade Runner can be referred to as http://rdf.freebase.com/ns/en.blade_runner, but it can also be referenced as http://rdf.freebase.com/ns/authority.netflix.movie.70053131 using the Netflix identifier, http://rdf.freebase.com/ns/authority.imdb.title.tt0083658 using the IMDB identifier, or as http://rdf.freebase.com/ns/wikipedia.en.Dangerous_Days using a Wikipedia wikiword (which in this case is a Wikipedia redirect to the wikiword Blade_Runner).
Freebase also provides a user maintained mapping of how these identifiers can be used to address resources in other LOD systems. The sameas.freebase.com schema can tell an LOD user that the Freebase Blade Runner Topic can also be found in DBpedia using Wikipedia identifiers or how musical artists can be found at the BBC using Musicbrainz identifiers.  In fact, the Freebase RDF interface uses the sameas.freebase.com schema to create the owl:sameAs links in the RDF output allowing the user community to expand the interconnections between Freebase and the LOD Cloud.
Linked Data providers are also using the strong identifiers in Freebase to identify entities such as companies and locations in their own data sets.  When they find an entity that is not represented in Freebase, they simply add the entity to Freebase and use the newly minted Freebase identifier.  This permits anyone using their data to understand how their entities relates to any of the more than 5 million things interconnected within Freebase.

The RDF interface can also be used to reference the Freebase type system, giving LOD data set providers vocabularies across a wide range of subject areas.  And because anyone can expand Freebase’s data model, data providers can use our schema development tools to build and extend these vocabularies to suite their needs.
Freebase was not designed for ephemeral or fast changing data, like weather conditions or stock ticks.  But this type of information is well suited for publication as Linked Data.  Freebase entities representing a location or company can be annotated with references to LOD services that provide these types of volatile data.  Similarly, Linked Data provides a great way to disseminate very fined grained information that might be associated with a scientific study or financial report.  Linked Data provides a seemless transition from Freebase, where a user (or application) can run a query with constraints that run across a wide range of types to find entities of interest along with the LOD services that provide access to temporal or high resolution data not available in Freebase.
We recently demonstrated MQL Extensions which allows the Metaweb Query Language to use data from other systems as a part of the query constraint and result set.  While MQL Extensions are user extensible and work with a wide array of systems,  this capability makes the connection between Freebase and the LOD Cloud even more transparent.
For example, because US companies that are registered with the SEC are annotated CIK code in Freebase and the sameas.freebase.com schema indicates that the CIK annotation can be used to create a URI that is dereferencable at rdfabout.com, it is possible to write a MQL query that asks who is on the board of financial services companies that trade on NASDAQ and are  headquartered in California (and using another MQL Extension, you can ask for their stock price as well!)

SWC: Many organisations are very interested in Linking Open Data now but they are still not sure if they can benefit from publishing data on the web – what´s your experience so far?

Jamie: Linked Open Data provides a simple, standard way for organizations to distribute structured data.  For most organizations, providing access to data is another important outlet to announce the availability of higher value services.  For organizations involved in building or selling physical goods, the bits representing what they provide are not the goods themselves, but a way of attracting potential customers.  Making catalogs and specification sheets available in electronic form, so other applications can connect buyers to their physical goods is simply an effective marketing system.  Even for firms involved in electronic services, providing access to open structured data is generally a lead-in to value added services.  For instance, if I ran a service collecting hard-to-find information about manufacturing relationships between medium sized businesses, I would publish open company profiles covering things like market size, industry, location for the medium-sized businesses I tracked, so potential users the premium data would know I had the coverage they were looking for.

SWC: Just recently Google has announced to use RDFa to enhance their search results. What do you think?

Jamie: We are excited about Google’s announcement. Yahoo’s use of RDFa for Search Monkey and Google’s announcement gives RDFa users tangible benefits. The Search Monkey team was very quick to realize that because users can create data models in Freebase, and because the elements of those models all have strong RDF identifiers, Freebase provides a large and user extensible vocabulary for RDF/RDFa (see the list of vocabularies). When a user wants to create a Search Monkey application that works with their film review site, they need not invent a new vocabulary (that will probably be used only once),  they can use the Freebase Film Domain vocabulary which supports over 63,000 instances in Freebase alone.
Similarly, with over 5 Million well described Topics in Freebase and over 14,000,000 Named Objects (Topics, images, musical tracks and documents) when a user wants to unambiguously identify a subject or object in RDF/RDFa, Freebase has an extremely large collection of identifiers to draw from.  These cover people, places, companies, movies, music, books and wide variety of other subjects.  If Freebase doesn’t have the entity the user is looking for, they can of course add it themselves and make use of the identifier immediately. I think this is why Google used some Freebase identifiers in their examples. We hope that with Yahoo and Google’s support for RDFa the web will become a strongly annotated source of data which can support a wide range of user applications.

SWC: Thank you, Jamie!

Related articles by Zemanta

• Putting Freebase in a Star Schema (gen5.info)
• RDFa, Google (ivan-herman.name)

In the meantime I had the pleasure to listen to Talis´ Semantic Web Gang Podcast (January 2009 with Tom Tague from Calais) and some answers came into my mind:

1. Some service providers will provide the highest accuracy regarding the links or tags (and the “things behind them) they provide for a given ressource or document (like Open Calais does). Tom Tague mentioned in the podcast quite often how important disambiguation is to provide the highest quality.
2. Some will provide end-points to a given “thing” like a company, a person etc. in addition to free ones like DBpedia, but they always will try to refer to established URIs like the ones in DBpedia or Open Calais URIs, e.g. IBM´s URI @ Calais). Those companies will provide more facts, for example about a person, as those which are available now for free. They will build on the LOD infrastructure and will live in symbiosis with group number 3. They will control to whom additional facts will be given to but they will build exactly on the same interoperable framework as the “Linking Open Data” community does.
3. Some companies will build applications on top of the linked data infrastructure. They have two kinds of knowledge: Who has the best end-points to a complex “thing” which consists of a couple of other atomic things (which necessarily exist in the web of data)? Who is interested in such a mashup?

My prediction: One possible business model will be pretty much the same as iTunes is built upon at the moment: You can listen to a song for free – but only a couple of seconds , if you want more, you pay 99 cents.

If you want to know a little bit about Werner Faymann (who is Austria´s prime minister) you go to an application which makes use from DBpedia (or the like) starting at http://dbpedia.org/page/Werner_Faymann.

If you pay 99 cents (or a bit more…) you get even more facts about Mr. Faymann, nicely mash-uped with other facts from the LOD cloud and together with special content from some other linked data sources, produced with relatively low costs due the high interoperability the Semantic Web provides – thanks to W3C and the whole community.

Sören Auer and his team from AKSW at Uni Leipzig are the coordinators of this EU funded project which supports the development of innovative software products. All industry partners are SMEs which offer services for different fields like E-learning, E-tourism or Business Intelligence. Leipzig and OpenLink Software will work on an integration of OntoWiki & Virtuoso.

The first day of the meeting was, of course, dedicated to socialize and get to know each other. The mixture of the project team turned out to be well chosen – and in the evening we flew at higher game: We had a nice overview over Leipzig standing on the highest building of the town.

On the second day of the meeting Orri Erling, Program Manager at OpenLink Software, came up with an idea which is pretty forward: Why shouldn´t we provide OntoWiki as a Linked Data Browser, e.g. on top of DBpedia etc.? One possible outcome of this project.

Some other use cases which make already use of the existing OntoWiki system were demonstrated: Take a look at Vakantieland (…and start to plan your holidays in the Netherlands) and also at LinkedGeoData where a nice user interface can be tried out.

The Kick-Off Meeting will proceed with two workshops dedicated to semantic technologies and to Application Development with the OntoWiki Framework. Thanks to Sören and his team for the excellent hosting of this event!

DBpedia has become the largest RDF repository for encyclopaedic knowledge, extracting structured information from Wikipedia and making it available on the Web of Data. UMBEL, on the other hand, provides an OpenCYC-based, light-weight ontology structure for relating Web content and data to a standard set of subject concepts, with a number of 20,000 concepts currently reached. In the Linked Data Cloud, DBpedia and UMBEL map and cross-reference each other.

In practice this means that UMBEL provides classes to describe the concepts to which “things” are members. For instance, named entities from Wikipedia such as “John F. Kennedy” are mapped with subject concepts such as Leader, Person, Administrator and Graduate, with broader and equivalent classes in CYC and FOAF and broader subject concepts within UMBEL. A link is set to Wikipedia, as well as a ‘same as’ reference to DBpedia. A class structure enables faceted browsing and extraction, inferencing, and navigation and discovery for all datasets linked to that structure.

DBpedia, in turn, returns properties of ‘John J. Kennedy’ (e.g. abstracts in available Wikipedia languages, demographic information such as birth date and place, alma mater, predecessors and successors), and ‘same as’ references, e.g., to the JFK entry in Freebase (who recently released their RDF service) and the aforementioned page in UMBEL. Furthermore, DBpedia maps the URI with available RDF types, for instance foaf:person or yago:AssassinatedAmericanPoliticians and, once again, with UMBEL’s subject concepts Person, Administrator, Graduate and Leader.

Due to its reliance on Wikipedia, DBpedia does a great job at covering a bandwidth of knowledge as broad as the spectrum of the interest of people participating in Wikipedia; it’s within the area of named entities, i.e. entities such as persons, organizations, locations, which have a proper name, but are not necessarily and specifically part of a particular, acknowledged domain or discipline. UMBEL, on the other hand, has as its most apparent advantage its reliance on OpenCyc and with that the strong inferencing and logic capabilities of the CYC knowledge-base which are thus also brought to the Web of Data. DBpedia is a community project started by the University of Leipzig, Free University Berlin and OpenLink Software, while the open and free UMBEL is developed and hosted by Zitgist with support from, again, OpenLink Software.

Now, and in particular with the recent release of Zitgist’s web service endpoints and with the incorporation of UMBEL classes in DBpedia, questions arises as to the relationship of the two projects, and regarding the role of OpenLink Software in the further process. To draw a distinction:

One could say that DBpedia’s goal is to lower the barrier for web developers and end-users in the actual use of the semantic web, while UMBEL aims at bringing “order to the chaos” that is inherent to user-generated, collective knowledge.

Would you agree with this description – and is it a contradiction at all or the kind of dynamic the Semantic Web community has been waiting for?

Mike Bergman: Yes, I would agree with this description, though we have tried many others. For example, in various writings in the past, we have described UMBEL as a roadmap, or middleware, or a backbone, or a concept ontology, or an ‘infocline’, or a meta layer for metadata, and others. Today, what I tend to use, particularly in reference to DBpedia, is the TBox-ABox distinction in computer science and description logics. UMBEL is more of a class or structural and concept relationships schema — a TBox — while DBpedia is more of an an instance and entity layer with attributes — an ABox. I think they are pretty complementary…

Sören Auer: I very much agree with Mike, but would like to add that Wikipedia authors do not have in mind to create a coherent and consistent knowledge base when working on Wikipedia. I think the more we demonstrate the benefits of the semantic representations in DBpedia to the Wikipedia community, these people will start to organize and rearrange content to enable the use of Wikipedia as a knowledge base. Right now, Wikipedia authors just have not yet been confronted with the problem of synonymous infobox properties or the uncleanliness of the category system, for example. I think with a few small and non-invasive changes to Wikipedia, much of the current chaos can be already resolved.

Mike Bergman: I agree, too, with Sören’s adder: I think it is difficult for Wikipedia authors to be consistent or coherent across the entire Wikipedia knowledge base. I think, then, the real question is where does that coherence or structural consistency come from? I think the nature of that task is quite different than creating or editing instance articles.

As for the dynamics and drivers of the community, the role of DBpedia for practical, linked data can not be overstressed. It was the first, remains the biggest, and has brought much visibility and awareness to linked data. I think I was one of the first to give DBpedia some press shortly after its release nearly two years ago, which is still one of my most popular blog posts. I’d like to think that UMBEL is now timely and well-positioned to help provide a complementary resource of concept classes, but that is not proven. DBpedia is.

DBpedia relies on user-generated content, UMBEL, with CYC, is expert-driven. How will a system that combines these divergent approaches continue to grow?

Mike Bergman: DBpedia is in the enviable position of being able to leverage two things: 1) the phenomenal success and growth of its source content, Wikipedia; and 2) the growing sophistication of information and structure extraction techniques built around that phenomenon. In fact, my most recent SWEETpedia listing of research projects based on Wikipedia now exceeds 170 or something projects and we discover and learn about more daily. We now see major research efforts in Germany, Austria, Japan, New Zealand, the US and England (among those I know) aggressively mining and learning from Wikipedia. I’m sure DBpedia will learn more, but others will increasingly contribute as well. This is unprecedented and very exciting. Wikipedia may have as strong a heritage in contributing to research and language and structure understanding as it does as a reference encyclopedia.

But, that is largely instance and attribute data, the ABox to use my earlier terminology. For structural and conceptual relationships — the coherent way to organize and relate things in the world; that is, the TBox — I think there is much subtlety and thinking required. I’ve used the phrase before that creating structural schema is “not like flinging hash”. Efforts such as Cyc, with nearly 1,000 person years of consistent testing and effort behind it, or perhaps others such as SUMO or what is coming out of the biology community with OBO (Open Biomedical Ontologies), offer better coherency and the ability to interoperate across diverse datasets and domains. Perhaps Wikipedia and its data extraction offshoots may someday get to this point — and I truly hope so — but are not anywhere near that today in my opinion.

Sören Auer: I’m actually not so concerned about the lack of structure and coherency as Mike. If we look at the current mostly textual Web and the search engines making it accessible to us humans, there is almost no data, few structure and even less coherency, but search engines still manage to provide an enormous added-value. If we add more data to the Webm much more sophisticated browsing, searching and data integration interfaces can be built. Structure and coherency will then emerge automatically, once people see how their content is indexed and can be easily found (or not).

The same happened by the way with the traditional Web 1.0 – in the beginning nobody used HTML’s meta-data tags. Once search engines started to interpret those for ranking results, meta-tags shifted to the center of attention of every Web content manager. Applied to the Semantic Data Web: once search engines understand foaf:person, everybody will use this concept for describing people.

A little experiment – Web 3.0, Semantic Web, Web of Data, Linked Data: Can you think of an ontology that is able to connect these terms and reveal the concepts behind them?

Mike Bergman: Well, I’d like to think UMBEL is that ontology (smile). That is certainly our intent, though truthfully we are still working out structural details and have not added all of this nice SemWeb terminology. But it is coming shortly (smile).

Sören Auer: A precise definition of these terms in the mathematical sense does not (and probably will never) exist, so articles such as those in Wikipedia (or many other publications) about the terms are from my point of view completely sufficient to reveal the concepts behind them to us. Of course it’s nice to have pretty and world-wide unique identifiers (such as provided by UMBEL) to annotate articles about these terms.

Mike Bergman: Well, UMBEL is about linking to concepts, though we welcome anyone thinking our identifiers are pretty (smile). One key aspect we will see moving forward is how we can translate those concepts into one of the 250 languages now used by Wikipedia while retaining existing structure. That is a real exciting prospect.

What are the concepts you would personally want to employ to explain the over-arching idea of these terms to a newbie?

Mike Bergman: Structured data on the Web is becoming like newly visible stars as nighttime darkens in the desert. Structured data are points of light in a global information space. We need fixed reference points in that sky to find specific stars. We need linkages to extract meanings and constellations from them. So, we both need to expose those stars — as linked data — and to provide fixed references to find them again and connections to draw meaning from them. Objects, references and connections all work in concert to expose the wonder.

Sören Auer: Its difficult for me to add something after Mike’s truly poetic description of pretty technical terms (smile). I see Linked Data or the Web of Data as the next milestone on the road of realizing the vision of the Semantic Web. In this regard, I’m Marxist (smile) and think Marx’s Law of transformation of Quantity into Quality applies: once we have a sufficient quantity of data out there on the Web, a new quality will emerge. Unfortunately, we are still far away from reaching a critical mass, since the Semantic or Data Web as we recently found out (cf.: Triplify – Lightweight Linked Data Publication from Relational Databases, PDF, 332 KB) is effectively shrinking if compared with the growth of the traditional Web.

Kingsley Idehen from OpenLink Software was the first to announce on the W3C mailing list that DBpedia & UMBEL are now “fully connected.” Is Kingsley the bridge-builder between the two projects?

Mike Bergman: Without question. Kingsley has backed both efforts in a big way with people and resources. His company actually did the first RDF linkage between the two projects. What is more remarkable, however, is that DBpedia and UMBEL are but a small slice of the things Kingsley has been backing. He has been a leader in middleware, scalable clusters and cloud computing, RDFizing all data forms, converting relational legacy data to linked RDF data, and providing demos and teaching to newbies on mailing lists. I’m glad you asked this question because Kingsley is a real catalyst and visionary.

Sören Auer: I agree, Kingsley is a mover and shaker in many areas of technology and innovation and in particular the Data Web. However, we should not forget his marvelous team at OpenLink with the database mastermind Orri Erling, Hugh Williams, Ivan Mikhailov, Yrjäna Rankka and all the others.

In the same vein: What are roles that are vital for the LOD-engineering process? Are there also “gardeners” or “bureaucrats“, as Wikipedians would put it?

Mike Bergman: I think the LOD (linking open data) and DBpedia mailing lists have been very effective, and there continues to be good community and organization around those efforts. We know that Wikipedia is not a free for all, with a kind of self-policing plus type of governance. I think that works well at the instance level.

However, structure decisions at the level of conceptual schema such as UMBEL or OpenCyc or even the mapping of classes between ontologies or datasets requires more skill and care. Others may not agree, but I think the schema aspects essential to UMBEL’s purpose — while definitely needing to be open and participatory at the suggestion or input level — possibly require roles more like “priest” or “professional” or “authority” at the actual roll-out level. Without quality, structure is nothing, and all of this is just an elaborate toy.

Sören Auer: Again my philosophy here differs a little from Mike’s: I’m pretty skeptic there will be one ontology or organization scheme for the Semantic Web. Rather, I think structure and homogeneity will be achieved on a peer-to-peer basis first and a community consensus will emerge later.

Mike Bergman: Yeah, this is an excellent point and I’m glad Sören raised it. While it is true we have put much effort into creating a lightweight structure of concept classes for linking disparate datasets, we too do not see “one ontology to rule them all.” I suspect there will often be none, and then many times other frameworks chosen. It really depends on the use case and purpose. UMBEL’s specific purpose is to provide a coherent framework for serious knowledge engineers looking to federate data. After that, other frameworks with a different purposes may then need to do the heavy lifting of actual data interoperability.

With its recent RDF service release, Freebase has risen to the level of a major SemWeb knowledge base, too. Where do you see its role in the future SemWeb ecology, also in relation to your own projects?

Sören Auer: To be honest, I was not very convinced of Freebase from the very beginning although their technology in particular the user interfaces are impressive. From my point of view the Freebase approach was too centralized and proprietary. A better strategy would have been to develop an open (source) technology, which people can deploy on their own Websites combined with server side crawling and search facilities. The first part of this equation is by the way exactly what we are aiming for with our OntoWiki and Triplify projects. However, if Freebase now moves towards more openness and interoperability, this can be only applauded.

Mike Bergman: I think early on that Freebase was needing to get its own feet set and did not do much with regard to open standards or external interoperability. These are good signs we are now seeing. However, I think the revenue model around user-supplied data remains highly suspect. Heck, Wikipedia with all of its tremendous success is daily soliciting contributions from users. It’s hard to get traction without being open and free, and its hard to make money when you are open and free even with traction. But these new announcements now make it much easier for us to use Freebase should our customers request it.

About DBpedia:

DBpedia is a community effort to extract structured information from Wikipedia and to make this information available on the Web. Wikipedia is the by far largest publicly available encyclopedia on the Web. Wikipedia editions are available in over 250 languages with the English one accounting for more than 2.49 million articles. Unfortunately, Wikipedia’s search capabilities are limited to full-text search, which allows very limited access to this valuable knowledge-base. Semantic Web technologies enable expressive queries against structured and interlinked information on the Web. DBpedia allows you to make sophisticated queries against Wikipedia, and to link other data sets on the Web to Wikipedia data. The DBpedia data set currently provides information about more than 2.49 million “things”, including at least 108,000 persons, 392,000 places, 57,000 music albums, and 36,000 films. Altogether, the DBpedia data set consists of 218 million pieces of information (RDF triples).

dbpedia.org (general website)

wiki.dbpedia.org/OnlineAccess (DBpedia Wiki – Online Access)

About UMBEL:

UMBEL (Upper Mapping and Binding Exchange Layer) is a lightweight ontology structure for relating Web content and data to a standard set of 20,000 subject concepts. Its purpose is to provide a fixed set of reference points in a global knowledge space. These subject concepts have defined relationships between them, and can act as binding or attachment points for any Web content or data. UMBEL is like a map of an interstate highway system, a set of roadsigns to help find related content and a way of getting from one big place to another. Once in the right vicinity, other maps (or ontologies) — more akin to detailed street maps — are then necessary to get to specific locations or street addresses. By definition, these more fine-grained maps are beyond UMBEL’s scope. But UMBEL can help provide the context for placing such detailed maps in relation to one another and in relation to the Big Picture of what related content is about.

umbel.org (project website)

umbel.zitgist.com (UMBEL webservice – sandbox)

The interview was led by Andreas Blumauer, SWC