Category Archives: Semantic Web Applications

Andreas Blumauer

Why SKOS thesauri matter – the next generation of semantic technologies

by , Category: Search Engines, Semantic Web Applications, Text Mining, Tools & Software

As a matter of fact still a lot of “semantic technologies” are around which do nothing else than pure statistical analysis of text. Sure, this is better than simple full text search but there are still quite a lot of opportunities to improve search, especially when it comes to more sophisticated applications like “similarity search”, the search for similar documents to enable cross-reading or recommendation systems.

Providers of first generation semantic technologies calculate rather basic “semantic networks” by co-occurency analysis which results sometimes in  disappointing results. Bearing in mind that Google just bought a company (“Google buys Metaweb“) which has been working on one of the largest knowledge bases in the world, we could assume that some of the last miles towards a semantic search engine can be achieved by applying thesauri or other structured knowledge bases.

A demo application was recently developed by PoolParty team where one can find out how thesauri will improve search results on top of second generation semantic technologies. With PoolParty SKOS based controlled vocabularies can be managed and also can be enriched with linked data. PoolParty Tag & Content Recommender analyzes virtually any text or website to recommend corresponding tags, concepts from (in this case) STW (Standard Thesaurus für Wirtschaft), DBpedia and respective articles from Wikipedia.

STW which was developed by the German National Library of Economics (ZBW) provides vocabulary on any economic subject: about 6,000 standardized subject headings and about 18,000 entry terms to support individual keywords.

This background knowledge is used in this demo app to improve the search for similar documents dramatically:

Similarity between two documents can be calculated not only on a key-phrase basis but also on a rather conceptual basis. Even if two documents do not have one single word or phrase in common they can be identified as “similar documents”.

This can be achieved because thousands of important relations between economic subjects are represented in the domain specific thesaurus. Thus, in this special case best results are achieved with documents from economics (for instance from Econstor) but of course for other recommender systems thesauri from other domains can be used instead of STW.

Nevertheless, also this approach can be improved and this development is underway: SKOS thesauri enriched with Linked Data do an even better job. This kind of third generation semantic technologies are currently developed by LASSO project and LOD2 project, two innovative projects in the area of linked data and the semantic web.

Michael Hausenblas

A Dynamic Web Of Data

by , Category: Linked Data & Open Data, Semantic Web Applications

As a matter of fact things change – the Web of Data is no exception in that respect. While some sources, such as Twitter, are intrinsically dynamic, others change every now and then, potentially in unforeseeable intervals. In the recent Talis Nodalities Magazine, we made a case for Keeping up with a LOD of changes; here I’m going to elaborate a bit more on the current state of Dataset Dynamics and its challenges.

Let us first step a back a bit and have a look what Dataset Dynamics are and why this is important. In the Web of Linked Data we typically deal with datasets, for example, from the biomedical domain or the media industry on the one hand, and entities, such as a certain protein or people on the other. For the entity-level case established HTTP caching mechanism can be leveraged (see the Caching Tutorial and Things Caches Do). Further, with Memento, a HTTP-based versioning mechanisms has been proposed as well as implemented, adding a “time dimension” to HTTP (see Fig. 1).

Fig. 1 Memento Framework (Source: "An HTTP-Based Versioning Mechanism for Linked Data" Herbert Van de Sompel, Robert Sanderson, Michael Nelson, Lyudmila Balakireva, Harihar Shankar, Scott Ainsworth, LDOW 2010)

Dataset-level changes

However, tackling dataset-level changes is a rather new field with no agreed-upon, even less standardised solution handy. The main problem is that a dataset typically talks about many thousands to millions of distinct entities, which makes it impractical to apply entity-level solutions for a range of use cases, such as link maintenance or replication (see also Fig. 2).

Fig. 2 Change frequency vs. change volume

I often hear these days: “it seems there is no solution for handling of dataset-level changes”; nevertheless, I think quite the opposite it true. There are plenty of proposed solutions from both the academia and practitioners, targeting different challenges in the areas of:

  • Change discovery – how do I find out about about dataset changes?
  • Propagating changes - if there is a change, how is the change communciated to a consumer?
  • Change semantics – how do I learn what has changed (has been added, removed, etc.)?

Some proposals on the table are integrated approaches (such as DSNotify, SemanticPingback, Talis Changeset) while others focus on certain aspects (like the dady vocabulary for discovery or the Graph Update Ontology for change semantics) or deal concrete environments, for example sparqlPuSH for SPARQL enpdoints.

A Dataset Dynamics Manifesto

No matter on what (set of) solutions the community eventually agrees on to address the handling of dataset-level changes, it should adhere to the following principles:

  • light-weight
  • distributed and scalable
  • standards-based

Obviously, a light-weight (and ideally RESTful) approach lowers the barriers to adoption and enables a quick uptake. When I say light-weight, I mean it both in terms of protocol and code. It should be easy to integrate in RDF stores and libraries and available in all common Web programming languages including but not limited to Java, PHP, .NET family, etc.

Just as the Web of Data is a globally distributed dataspace, handling of changes should be done in a distributed fashion. There will be many different publishers and consumers (such as agents, indexer, consolidator platforms, etc.) of datasets with different requirements and capabilities. A distributed approach can cope with this challenge in a cost- and performance-efficient way. Tightly connected to this: It has to scale. Today, we’re dealing with some hundreds of LOD datasets. In the next couple of years, this will likely explode into the millions and hence one needs to be able to deal with such a growth. The same, just sooner, is true for the number of consumers of the changes.

Last but not least the Dataset Dynamics solution should be based on standards. It doesn’t necessarily need to be RDF for all of the challenges as outlined above. For example, Atom offers a standardised, extensible and widely accepted format to propagate changes; to take this further Pubsubhubbub can be utilised to enable a standardised, distributed publisher-subscriber scheme (Fig 3.)

Fig. 3 Pubsubhubbub - a standard-based, distributed publisher-subscriber-hub system (Source: http://docs.google.com/present/view?id=ajd8t6gk4mh2_34dvbpchfs)

As I’ve outlined above, it might still be too early for a conclusion on how to deal with dataset-level changes. However, people interested in this area have gathered already in the Dataset Dynamics group where solutions are discussed and implemented, potentially leading to a W3C standardisation work.

As an aside: in case you’re at the WWW2010 in Raleigh (NC, USA) these days, you may want to join the break-out meeting on Dataset Dynamics during the W3C Linked Open Data track on 29 April 2010.

(This blog post was written by Michael Hausenblas)

Tassilo Pellegrini

Interview with Juan Sequeda: “I believe Linked Data will enable new killer apps that are only possible thanks to Linked Data.”

by , Category: Calls & Competitions, Linked Data & Open Data, Semantic Web Applications

Juan Sequeda, co-chair of the Triplification Challenge 2010 and one of the core figures in the Linked Data movement, gives us his view how the Semantic Web might evolve. His central message: “Once there is an incentive to create quality links, these links will start to show up. And then users will start linking to the data hubs of their interest.”

Linked Data itself has grabbed a lot of attention inside the Semantic Web community recently. But what about the outside perspective? Could linked data be called the killer app for the Semantic Web?

I foresee two things happening with Linked Data. One is from the web development perspective (the so-called Web 2.0 developers) and the other is from the enterprise perspective. The web development community will sooner than later realize that Linked Data will enable easy integration of data and therefore will ease the pain of consuming data from different data sources. Thanks to big organizations such as BBC, New York Times, Reuters, Best Buy, etc. web developers will start paying attention to this “new thing” called Linked Data.

What we need is that the inside Semantic Web community starts to create applications on top of current Linked Data so when the outside web development community starts to pay attention, they have something to chew on. We (the semantic web community) needs to start speaking the web development language. There is still a big gap. I have had personal experiences with people in the web development community who think that RDF is XML and because they hate XML, they will never consider it. This is false and this is something that we need to change.

From the enterprise perspective, Linked Data is another data integration solution. Data integration has been a problem since day one of relational databases. I believe enterprises will be open to consider new solutions with new technologies. I’m hoping to see new startups tackling the enterprise domain. Imagine being able to query “get all my clients from cities whose population is greater than 1 million” even though I don’t have the data about population of cities in my database.

Is Linked Data the killer app for the Semantic Web? Before I answer that, I would like to ask, what was the killer app of the Web? Was it the browser? Was it e-commerce? Was it search? Was it Amazon or Ebay or Google? I believe Linked Data will enable new killer apps, apps that are only possible thanks to Linked Data. The browser was only possible because of HTML. So let’s ask ourselves what is possible because of Linked Data, and there we will find our killer app.

One of the core deficiencies of the young open data cloud is the little amount of interlinks between datasets. Is it just a matter of time to overcome this or are there other measures needed to turn the existing datasets into a true giant global graph?

I like to remind myself that this new wave of semantic web technologies is an extension of the current web. Therefore we should analyze how the web evolved in the beginning. Initially, everything were a bunch of documents on the web in which people manually created links to other documents. When Google started, it created an incentive to offer quality links between documents. This also created data hubs. If you write a blog post about a book, most probably you will link to the web document of that book either on Amazon and/or Wikipedia. I believe that this will happen with Linked Data. Once there is an incentive to create quality links, these links will start to show up. And then users will start linking to the data hubs of their interest.

Open Governmental Data is a big issue at the moment. The US and UK government have started to apply Linked Data principles to turn this vision into reality. Lots of other countries are following. What do you expect from this trend?

I believe that Linked Data will take off thanks to the initiative of governments. We always talk about the chicken and egg problem of the semantic web. Once we have organizations that don’t even think about it and are just interested in putting their data on the web, the semantic web will start to grow. If Bookstore ABC puts their data on the web, it may not be so meaningful. But if the US and UK government puts their data on the web, following the Linked Data principles, then people can wake up and say “ok, so this is for real. Let me start paying attention to this”.

You are one of the chairs of the Triplification Challenge 2010. Can you give us a brief insight what to expect from this year’s challenge? What are the conditions to participate?

The Triplification Challenge this year has grown and is very exciting. For the first time, it is offering two different tracks.

The first track, the Open Track will accept submissions on three areas 1) new datasets that are published following the Linked Data principles and that show potential benefit, 2) generic methods, mechanisms and approaches of creating Linked Data from legacy datasets and 3) applications that make use of Linked Data.

The second track is the New York Times track which will accept submissions of applications that make use of the New York Times Linked Data and one or more government dataset. The objective is to create an application powered by Linked Data that would be of interest to any constituent of that government.

I personally believe that the year 2010 is the year of creating Linked Data applications and the Triplification Challenge is the way to be part of it.

Tassilo Pellegrini

Interview with Georgi Kobilarov: “I believe that data publishing must happen in a distributed style.”

by , Category: Linked Data & Open Data, Mashups & Web services, Semantic Web Applications, Tools & Software

Uberblic.org connects structured data from the web. The Berlin-based inventor Georgi Kobilarov gives a brief insight into the mashup service and talks about the challenges when it comes to build applications upon linked data.

You have recently published the service uberblic.org, a Linked Data mashup editor. What was your motivation to develop this tool?

Uberblic.org provides an integrated view of web data. Our goal is to integrate all the structured data on the web, and give web-developers a single point to access to that reconciled data. More than that, we will open up the tools we use to manage the data sources to the community, so that the people can help us curating that repository of free data. We re-publish all the data we import as Linked Data, under the licenses of the original data publishers.

Some of the data sources we import are available in the Linked Open Data cloud as well, but many are not. Linked Data is an elegant way to publish data in a distributed way on the web, but consuming it from that distributed cloud is – at least – impractical. In every real-world application using linked data from the web I’ve seen, organizations built up internal copies of the cloud, and often even reconcile linked data sources. They build their own Linked Data proxies. Uberblic.org helps those users by providing one public proxy for data from the web. Many of our sources get monitored for data changes, and the according data in uberblic is updated in real-time.

uberblic

Can you give us a brief insight how the tool works? What technology is is built on?

My company, Uberblic Labs, has developed a data integration platform that we use to power uberblic.org. We call it the Uberblic Platform (the name uberblic is derived from the German “Überblick” – English “overview”). This platform enables us to do the full process of “data fusion”: Importing and converting external data sources, mapping the data schemas to a central ontology, filtering out data errors, automatically suggesting duplicates to the user, and merging data from different sources into a single, reconciled representation.

Structured and semi-structured data from the web is an excellent use case for our software platform, since there we come across all the interesting cases of real-world data heterogeneity. But what I think is especially powerful and yet missing in other Linked Data projects I know, is the ability to subscribe to update-feeds. We do that extensively, fetching updates in real-time from Wikipedia and the like.

Our platform is built in Scala and runs a on cluster of machines, with workers communicating through a messaging system. We developed an RDF storage layer on top of a distributed key-values store for storing all provenance information used in the extraction process, currently around 100 million named graphs for uberblic.org. That storage layer does not directly provide SPARQL access, so we push all the output data into a SPARQL endpoint hosted by Talis as well.

What have been the biggest challenges in tackling the integration issues of dispersed data?

It was quite a steep learning curve to do Linked Data not only in an academic environment, but in a reliable, industry-strength set-up. In academia, there was always the excuse that things are just research prototypes. Now that excuse is gone. That’s also where it becomes necessary to manually clean up data. And there are two ways to do that: Either you enable the users to change facts directly in your repository after you have imported the external data (that is what Freebase does), or you facilitate clean-up cycles in the original data source and fetch these updates in real-time. That is what we do.

I believe that data publishing must happen in a distributed style, because then each data source gets taken care of by a specialized group of people using specialized tools. And it’s what you see not only on the web, but also inside organizations and enterprises. But consuming data trough centralized APIs is more than just convenient. We all use Google
or another search engine as a central access point to web pages which are published in a distributed way all over the web, don’t we? Can you imagine today researching a topic on the web without the centralization power of search engines, just by following links across web sites, like in the old days?

When we built the Uberblic Platform, some of the things I imagined to be large headaches, like schema mapping, turned out to work really well. Those pathologic cases you often see in academic “challenges” are – well – pathologic. It’s not necessary to solve them fully automatically through super-intelligent algorithms. Much more important than the sophistication of your algorithms are well designed workflows so that the user becomes a part of the solution. And that’s not about crowd-sourcing or swarm intelligence, the editorial curating of schema mappings and object reconciliation can be done just by a small team of people. If they have the right set of tools.

What are the next plans with uberblic.org? Where will the journey go?

Uberblic.org will continue to integrate more interesting and useful data sources from the web, and we will start making more APIs available to web developers to build their applications on top. We are also looking for partners who are interested in developing applications and have been struggling in the past to get the cross-source data from the web they need.

The work on improving uberblic.org will also benefit our Uberblic Platform, and hence our clients who use that same software for integrating organizational data sources with each other and with the web of data.

About Georgi Kobilarov

Georgi is founder and managing director of Uberblic Labs, a company based in Berlin specialized in Linked Data integration. He worked as a research associate in the Web-based Systems Group at Freie Universität Berlin and as a visiting researcher at Hewlett Packard Labs Bristol. As co-founder and lead developer of DBpedia, he was also a day-one contributor to the Linking Open Data project. Georgi is consulting with the BBC on several Linked Data related projects. He organizes the Web of Data Meetup London, a bi-yearly gathering of the UK Linked Data community. Georgi graduated with a Diplom in business administration from Freie Universität Berlin and has many years of work experience as a software developer. Visit his blog: http://blog.georgikobilarov.com