We showcase the KnowledgeStore (KS), a scalable, fault-tolerant, and Semantic Web grounded storage system for interlinking unstructured and structured contents. The KS contributes to bridge the unstructured (e.g., textual document, web pages) and structured (e.g., RDF, LOD) worlds, enabling to jointly store, manage, retrieve, and query, both typologies of contents.
Despite the widespread diffusion of structured data sources and the public acclaim of the Linked Open Data (LOD) initiative, a preponderant amount of information remains nowadays available only in unstructured form, both on the Web and within organizations. While different in form, structured and unstructured contents are often related in content, as they speak about the very same entities of the world (e.g., persons, organizations, locations, events), their properties, and relations among them. Despite the last decades achievements in Natural Language Processing (NLP), now supporting large scale extraction of knowledge about entities of the world from unstructured text, frameworks enabling the seamless integration and linking of knowledge coming both from structured and unstructured contents are still lacking.1
In this demo we showcase the KnowledgeStore (KS), a scalable, fault-tolerant, and
Semantic Web grounded storage system to jointly store, manage, retrieve, and query,
both structured and unstructured data. Fig. 1a shows schematically how the KS
manages unstructured and structured contents in its three representation layers. On the one
hand (and similarly to a file system) the resource layer stores unstructured content in
the form of resources (e.g., news articles), each having a textual representation and
some descriptive metadata. On the other hand, the entity layer is the home of structured
content, that, based on Knowledge Representation and Semantic Web best practices,
consists of axioms (a set of hsubject, predicate, objecti triples), which describe the
entities of the world (e.g., persons, locations, events), and for which additional metadata are
kept to track their provenance and to denote the formal contexts where they hold (e.g.,
point of view, attribution). Between the aforementioned two layers there is the mention
layer, which indexes mentions, i.e., snippets of resources (e.g., some characters in a text
document) that denote something of interest, such as an entity or an axiom of the entity
layer. Mentions can be automatically extracted by NLP tools, that can enrich them with
additional attributes about how they denote their referent (e.g., with which name,
qualifiers, “sentiment”). Far from being simple pointers, mentions present both unstructured
1 See [
has mention Indonesia Hit By Earthquake A United Nations assessment team was dispatched to the province after two quakes, measuring 7.6 and 7.4, struck west of Manokwari Jan. 4. At least five people were killed, 250 others injured and more than 800 homes destroyed by those temblors, according to the UN.
Mention Layer expressedby
Mention refers to REenlattitioynMMeenntitoionn tsaorugrecte . . .
Entity describedby
Axiom
holds in
Context dbpedia:United_Nations dbpedia:United_Nations rdf:type yago:PoliticalSystems dbpedia:United_Nations rdfs:label "UnitedNations"@en dbpedia:United_Nations foaf:homepage <htp:/www.un.org/> (a) (b) (c) and structured facets (respectively snippet and attributes) not available in the resource and entity layers alone, and are thus a valuable source of information on their own.
Thanks to the explicit representation and alignment of information at different levels, from unstructured to structured knowledge, the KS supports a number of usage scenarios. It enables the development of enhanced applications, such as effective decision support systems that exploit the possibility to semantically query the content of the KS with requests combining structured and unstructured content, such as “retrieve all the documents mentioning that person Barack Obama participated to a sport event”. Then, it favours the design and empirical investigation of information processing tasks otherwise difficult to experiment with, such as cross-document coreference resolution (i.e., identifying that two mentions refer to the same entity of the world) exploiting the availability of interlinked structured knowledge. Finally, the joint storage of (i) extracted knowledge, (ii) the resources it derives from, and (iii) extracted metadata provides an ideal scenario for developing, training, and evaluating ontology population techniques. 2
An overview of the KnowledgeStore
In this section we briefly outline the main characteristics of the KS. For a more
exhaustive presentation of the KS design, we point the reader to [
Architecture At its core, the KS is a storage server whose services are utilized by external clients to store and retrieve the contents they process. From a functional point of view, we identify three main typologies of clients (see Fig. 1b): (i) populators, whose purpose is to feed the KS with basic contents needed by other applications (e.g., documents, background knowledge from LOD sources); (ii) linguistic processors, that read input data from the KS and write back the results of their computation; and, (iii) applications, that mainly read data from the KS (e.g., decision support systems). Internally, the KS consists of a number of software components (see Fig. 1c) distributed on a cluster of machines: (i) the Hadoop HDFS filesystem provides a reliable and scalable storage for the physical files holding the representations of resources (e.g., texts and linguistic annotations of news articles); (ii) the HBase column-oriented store builds on Hadoop to provide database services for storing and retrieving semi-structured information about resources and mentions; (iii) the Virtuoso triple-store stores axioms to provide services supporting reasoning and online SPARQL query answering; and, (iv) the Frontend Server has been specifically developed to implement the operations of the CRUD and SPARQL endpoints on top of the components listed above, handling global issues such as access control, data validation and operation transactionality. User Interface (UI) The KS UI (see Fig. 2) enables human users to access and inspect the content of the KS via two core operations: (i) the SPARQL query operation, with which arbitrary SPARQL queries can be run against the KS SPARQL endpoint, obtaining the results directly in the browser or as a downloadable file (in various file formats, including the recently standardized JSON-LD); and, (ii) the lookup operation, which given the URI of an object (i.e., resource, mention, entity), retrieves all the KS content about that object. These two operations are seamlessly integrated in the UI, to offer a smooth browsing experience to the users. 3
Showcasing the KnowledgeStore and concluding remarks
During the Posters and Demos session, we will demonstrate live how to access the KS
content via the UI (similarly to the detailed demo preview available at [
In the last few months, several running instances of the KS were set-up (on a cluster
of 5 average specs servers) and populated using the NewsReader Processing Pipeline [
Acknowledgements The research leading to this paper was supported by the European Union’s 7th Framework Programme via the NewsReader Project (ICT-316404).