FRED is a machine reading tool for converting text into internally well-connected and quality linked-data-ready ontologies in webservice-acceptable time. It implements a novel approach for ontology design from natural language sentences, combining Discourse Representation Theory (DRT), linguistic frame semantics, and Ontology Design Patterns (ODP). The current version of the tool includes Earmark-based markup, and enrichment with word sense disambiguation (WSD) and named entity resolution (NER) o -the-shelf components.
The problem of knowledge extraction (KE) from text is still insu ciently
addressed from a semantic web (SW) perspective. Being able to automatically
produce quality linked data and ontologies from natural language text would be
a breakthrough as it would enable the development of applications that
automatically produce machine-readable information from Web content as soon as
it is edited and published by generic Web users. A rather detailed landscape
analysis of the currently available tools for KE, and their exploitation for SW
basic tasks is presented in [
FRED4 [
Related works and comparison to other tools for knowledge extraction are
detailed in [
In this section we present an overview of the system and a scenario that shows
the output resulting from FRED. [
Boxer frame-based approach supports FRED in automatically design an ontology by following good modeling practices based on ontology patterns. However, Boxer tranforms natural language to a logical form compliant with DRT (substantially a variety of rst-order logic) which di ers a lot from RDF or OWL, and the heuristics that it implements for interpreting a natural language and transforming it to a DRT-based structure can be sometimes awkward when directly translated to ontologies for the SW, because it obeys pure FOL-oriented design style. For this reason, Boxer DRT-based output is transformed by FRED to OWL/RDF ontologies by means of a set of heuristics, some of them are showed in Figure 1. Figure 2 depicts the main components of FRED: Communication: exposes APIs for querying the system; Refactoring : transforms Boxer output5 into a convenient data structure to be passed to the Reenginering component; Reengineering : applies a collection of ad-hoc mapping rules and heuristics for producing logically consistent OWL/RDF like triples.
In addition, FRED architecture is open to be easily integrated with other
components that exploit the text span markup speci cation supported its current
version, i.e. Earmark [
NIF and NERD, makes it trivial to augment FRED graphs with o -the-shelf components for e.g. NER, WSD, etc. Our demo, available online6, allows a user to enter any text (or select one from the list of examples provided), and by simply clicking the Read It! button, to receive a RDF/OWL representation of it7. Some features can be customized, e.g., type of output, NER or WSD activation, tenserelation between events activation, etc. By default, the output is in the form of Graphviz-like graphs (showing only a core subset of triples), to allow human users to quickly check the OWL/RDF representation.
FRED output consists in RDF triples including either xed properties: rdf:type, rdfs:subClassOf, owl:sameAs, dul:associateWith, owl:equivalentTo, Earmark properties, thematic roles for extracted events, etc., or customized properties produced by the automatic (machine) reading of the sentence.
As an example, consider the sentence \The Black Hand assassinated Franz Ferdinand during his visit to Sarajevo." Figure 3 shows FRED output for this sentence: an instance of dul:Event, fred:assassinate 1, is used to represent the assassination of Franz Ferdinand. It is typed as fred:Assassinate, which is disambiguated by the VerbNet frame vn.data:Assassinate 42010000. Such an event involves the individual fred:Black hand as agent, and the individual fred:Franz ferdinand (who is recognized and resolved as the same individual as dbpedia:Archduke Franz Ferdinand of Austria) as patient. Furthermore, the RDF graph expresses that such an event happened fred:during fred:visit 1. FRED heuristically assigns a type fred:Visit to fred:visit 1; such a type is disambiguated by means of alignments to WordNet, which in turn is aligned to other ontologies, so that FRED can infer e.g., that fred:Visit is a d0:Activity. The location of the visit is also identi ed and correctly resolved to dbpedia:Sarajevo. Notice that FRED assigns types to all identi ed individuals either by using classes from existing ontologies e.g., DOLCE, Schema.org, etc., when the entity can be resolved e.g., as a DBpedia entity, or by creating new classes based on the terms used in the input text and disambiguating them on WordNet.
FRED also supports more sophisticated constructs, e.g. propositional referents, called situations, full- edged negation on events or situations, and basic modalities over events. 3 Conclusion We have presented FRED, a machine reader for the Semantic Web, which automatically extracts rich and connected knowledge from text, represents it as OWL/RDF, and links it to other resources: VerbNet, FrameNet, WordNet, DBpedia, foundational ontologies, etc. The current research is on mainly evaluating it on vertical tasks, and extending its internal components for multilinguality.