In May 2012, the Web search engine Google has introduced the so-called Knowledge Graph, a graph that understands real-world entities and their relationships to one another. It currently contains more than 500 million objects, as well as more than 3.5 billion facts about and relationships between these different objects. Soon after its announcement, people started to ask for a programmatic method to access the data in the Knowledge Graph, however, as of today, Google does not provide one. With SEKI@home, which stands for Search for Embedded Knowledge Items, we propose a browser extension-based approach to crowdsource the task of populating a data store to build an Open Knowledge Graph. As people with the extension installed search on Google.com, the extension sends extracted anonymous Knowledge Graph facts from Search Engine Results Pages (SERPs) to a centralized, publicly accessible triple store, and thus over time creates a SPARQL-queryable Open Knowledge Graph. We have implemented and made available a prototype browser extension tailored to the Google Knowledge Graph, however, note that the concept of SEKI@home is generalizable for other knowledge bases.
1.1
With the introduction of the Knowledge Graph, the search engine Google has
made a significant paradigm shift towards “things, not strings” [
The term crowdsourcing was first coined by Jeff Howe in an article in the
magazine Wired [
Kaplan and Haenlein have defined social media as “a group of Internet-based
applications that build on the ideological and technological foundations of Web
2.0, and that allow the creation and exchange of user-generated content” [
In this position paper, we describe and provide a prototype implementation of an approach, tentatively titled SEKI@home and based on crowdsourcing via a browser extension, to make closed knowledge bases programmatically and openly accessible. We demonstrate its applicability with the Google Knowledge Graph. The extension can be added to the Google Chrome browser by navigating to http://bit.ly/SEKIatHome, the Open Knowledge Graph SPARQL endpoint can be tested at http://openknowledgegraph.org/sparql1.
The remainder of this paper is structured as follows. In Section 2, we highlight related work for the field of extracting data from websites with RDF wrappers. In Section 3, we describe the SEKI@home approach in detail. We provide a short evaluation in Section 4. The paper ends with an outlook on future work in Section 5 and a conclusion in Section 6. 2
Wrappers around Web services or Web pages have been used in the past to
lift data from the original source to a meaningful, machine-readable RDF level.
Examples are the Google Art wrapper by Guéret [
With SEKI@home, we offer a related, however, still different in the detail, approach to lift and make machine-readably accessible closed knowledge bases like the Knowledge Graph. The entirety of the knowledge base being unknown, via crowdsourcing we can distribute the heavy burden of crawling the whole Knowledge Graph on many shoulders. Finally, by storing the extracted facts centrally in a triple store, our approach allows for openly accessing the data via the standard SPARQL protocol. 3 3.1
We have implemented our prototype browser extension for the Google Chrome browser. Chrome extensions are small software programs that users can install to enrich their browsing experience. Via so-called content scripts, extensions can inject and modify the contents of Web pages. We have implemented an extension that gets activated when a user uses Google to search the Web. 3.2
Web scraping is a technique to extract data from Web pages. We use CSS
selectors [
Albeit the claim of the Knowledge Graph is “things, not strings” [
We have modeled the plaintext Knowledge Graph terms (or predicates) like “Born”, “Full name”, “Height”, “Spouse”, etc. in an informal Knowledge Graph ontology under the namespace okg (for Open Knowledge Graph) with spaces converted to underscores. This ontology has already been partially mapped to common Linked Data vocabularies. One example is okg:Description, which directly maps to dbpprop:shortDescription from DBpedia. Similar to the unknown list of objects in the Knowledge Graph (see Subsection 1.2), there is no known list of Knowledge Graph terms, which makes a complete mapping impossible. We have collected roughly 380 Knowledge Graph terms at time of writing, however, mapping them to other Linked Data vocabularies will be a permanent work in progress. As an example, Listing 1 shows the lifted, meaningful JSON-LD as returned by the extension. { " @id ": " http :// openknowledgegraph . org / data / H4sIAAAAA [...]" , " @context ": { " Name ": " http :// xmlns . com / foaf /0.1/ name ", " Topic_Of ": { " @id ": " http :// xmlns . com / foaf /0.1/ isPrimaryTopicOf ", " type ": " @id " }, " Derived_From ": { " @id ": " http :// www .w3. org /ns/ prov # wasDerivedFrom ", " type ": " @id " }, " Fact ": " http :// openknowledegraph . org / ontology / Fact ", " Query ": " http :// openknowledegraph . org / ontology / Query ", " Full_name ": " http :// xmlns . com / foaf /0.1/ givenName ", " Height ": " http :// dbpedia . org / ontology / height ", " Spouse ": " http :// dbpedia . org / ontology / spouse " }, " Derived_From ": " http :// www . google . com / insidesearch / features / search / knowledge . html ", " Topic_Of ": " http :// en. wikipedia . org / wiki / Chuck_Norris ", " Name ": " Chuck Norris ", " Fact ": [" Chuck Norris can cut thru a knife w/ butter ."] , " Full_name ": [" Carlos Ray Norris "], " Height ": ["5 ' 10\""] , " Spouse ": [ { " @id ": " http :// openknowledgegraph . org / data / H4sIA [...]" , " Query ": " gena o' kelley ", " Name ": " Gena O' Kelley " Listing 1. Subset of the meaningful JSON-LD from the Chuck Norris Knowledge Graph data. The mapping of the Knowledge Graph terms can be seen in the @context.
The facts extracted via the SEKI@home approach are derived from existing
third-party knowledge bases, like the Knowledge Graph. A derivation is a
transformation of an entity into another, a construction of an entity into another,
or an update of an entity, resulting in a new one. In consequence, it is
considered good form to acknowledge the original source, i.e., the Knowledge Graph,
which we have done via the property prov:wasDerivedFrom from the PROV
Ontology [
At time of writing, we have evaluated the SEKI@home approach for the criterium ease of use with a number of 15 users with medium to advanced computer and programming skills who had installed a pre-release version of the browser extension and who simply browsed the Google Knowledge Graph by following links, starting from the URL https://www.google.com/search?q=chuck+ norris, which triggers Knowledge Graph results. One of our design goals when we imagined SEKI@home was to make it as unobtrusive as possible. We asked the extension users to install the extension and tell us if they noticed any difference at all when using Google. None of them noticed any difference, while actually in the background the extension was sending back extracted Knowledge Graph facts to the RDF triple store at full pace. 4.2
On average, the number of 31 triples gets added to the triple store per SERP with Knowledge Graph result. Knowledge Graph results vary in their level of detail. We have calculated an average number of about 5 Knowledge Graph terms (or predicates) per SERP with Knowledge Graph result. While some Knowledge Graph values (or objects) are plaintext strings like the value “Carlos Ray Norris” for okg:Full_name, others are references to other Knowledge Graph entities, like a value for okg:Movies_and_TV_shows. The relation of reference values to plaintext values is about 1.5, which means the Knowledge Graph is well interconnected. 4.3
In its short lifetime from August 11 to September 6, 2012, the extension users have collected exactly 2,850,510 RDF triples. In that period, all in all 39 users had the extension installed in production. 5
A concrete next step for the current application of our approach to the
Knowledge Graph is to provide a more comprehensive mapping of Knowledge Graph
terms to other Linked Data vocabularies, a task whose difficulty was outlined in
Subsection 3.3. At time of writing, we have applied the SEKI@home approach
to a concrete knowledge base, namely the Knowledge Graph. In the future, we
want to apply SEKI@home to similar closed knowledge bases. Videos from video
portals like YouTube or Vimeo can be semantically enriched, as we have shown
in [
In this paper, we have shown a generalizable approach to first open up closed knowledge bases by means of crowdsourcing, and then make the extracted facts universally and openly accessible. As an example knowledge base, we have used the Google Knowledge Graph. The extracted facts can be accessed via the standard SPARQL protocol from the Google-independent Open Knowledge Graph website (http://openknowledgegraph.org/sparql). Just like knowledge bases evolve over time, the Knowledge Graph in concrete, the facts extracted via the SEKI@home approach as well mirror those changes eventually. Granted that provenance of the extracted data is handled appropriately, we hope to have contributed a useful socially enabled chain link to the Linked Data world.
T. Steiner is partially supported by the EC under Grant No. 248296 FP7 (I-SEARCH).