Semantic web applications are integrating data from more and more di erent types of sources about events. However, most data annotation frameworks do not translate well to semantic web. We describe the grounded annotation framework (GAF), a two-layered framework that aims to build a bridge between mentions of events in a data source such as a text document and their formal representation as instances. By choosing a two-layered approach, neither the mention layer, nor the semantic layer needs to compromise on what can be represented. We demonstrate the strengths of GAF in exibility and reasoning through a use case on earthquakes in Southeast Asia.
Semantic web applications are ingesting data from more and more di erent sources such as output from natural language processing applications, sensor data, videos or nancial transactions. Each of these domains has their own data annotation practices which rst need to be reconciled with semantic web standards. One issue with integrating information from di erent sources is that representation formats tend to look at their domain in isolation, making it di cult to integrate information that comes from other domains.
The Grounded Annotation Framework (GAF) [
We will describe GAF in Section 2, present an example in Section 3 and conclude with pointers for future work in Section 4.
The Grounded Annotation Framework
The main property of GAF is that it distinguishes instances from instance
mentions. A mention is the act of referring to an object where an instance
is the object itself. The relation between instances and mentions is de ned by
gaf:denotedBy, which is the only new predicate GAF introduces. Di erent
resources (or even the same resource) may refer to an instance in di erent ways
and each of these references may have properties of its own. This is quite
common in natural language, where authors tend to alternate terms to refer to the
same object for stylistic reasons, but it can also play a role in other sources of
information. If, for instance, a sensor displays a measured temperature, this
displayed value has properties of its own that are clearly not properties of the value
that was measured, such as the instrument that was used to measure it and its
error rate. In the remainder of this contribution, we will illustrate GAF through
the example of presenting instances in the Simple Event Model (SEM) [
SEM is a model to express who did what, where, and when. It is not the
only RDF model to describe events but as SEM is not tied to a any domain
and is among the most exible, we chose this model as the core of our semantic
layer. It should be noted however that, in principle any RDF schema can be
integrated into GAF. TAF is designed to annotate coreference relations between
event mentions as well as participants, locations and temporal expressions, which
covers the kind of information also represented in SEM. TAF has the additional
advantage that it already distinguishes between instances and instance mentions
for participants and locations. We use a slightly adapted variant of TAF that
extends this distinction to events and temporal expressions as described in [
The gaf:denotedBy relations is used to link events represented in SEM to
speci c mentions represented in TAF. If a linguistic analysis identi es a syntactic
relation between an event mention and the mention of a person, we can derive
that this person is an Actor of the event in SEM according to the analysis of
a speci c text. Mentions thus play an important role in modelling provenance
of information. To model provenance we use the PROV-O ontology [
GAF annotations can be created both by starting from the linguistic layer and
the semantic layer. When starting from text for the mention layer, rst TAF
annotations are added to the text using the Celct Annotation Tool [
When starting from the semantic layer, events and event properties are linked
to textual mentions. We are currently working towards an annotation
environment based on CROMER [
Examples The example sentences shown in Figure 1 both contain information about the 2004 Indian Ocean Earthquake and Tsunami. The articles disagree on the cause of the earthquake; where Bloomberg ascribes it to moving tectonic plates, Veteran's Today sees a stealth attack submarine as the likely cause. Figure 2 shows that these two declarations can co-exist within the GAF representation of the earthquake. It is up to the application or user accessing the information to interpret the fact that there is a contradiction and for example select only particular sources for further processing. GAF provides the glue to connect non-semantic web data to semantic web representation formats. The rdfs:isDe nedBy relation at the top of Figure 2 shows how RDF predicates can be used to link GAF representations to external resources such as the Linked Open Data cloud.1 "Indonesia lies in a zone where the Indo-Australian, Eurasian, Philippine and Pacific plates meet and occasionally shift, causing earthquakes and sometimes generating tsunamis. There have been hundreds of earthquakes in Indonesia since a 9.1 temblor in 2004 caused a tsunami that swept across the Indian Ocean, devastating coastal communities and leaving more than 220,000 people dead in Indonesia, Sri Lanka, India, Thailand and other countries." (Bloomberg, 2009-01-07 01:55 EST) "...were most concerned about the cause, scope, and consequences of the December 26, 2004 Indian Ocean tsunamis because they were far bigger and more destructive than they had anticipated. More important, it had no clear alibi that their most likely source of the disaster, the Multi-Mission Platform of the new stealth attack submarine, the USS Jimmy Carter, had not been the culprit." (Veteran's Today, 2011-10-02) We have presented GAF, a grounded annotation framework for integrating information from various sources. We have shown its exibility in representing contradicting information from di erent textual sources.
We are currently developing an annotation tool that allows users to easily switch between linguistic and semantic annotation layers. After which we plan to develop tools supporting easy integration of other types of information, such as data from the Linked Open Data cloud, video metadata or sensor data. 1 http://groundedannotationframework.org/ provides full examples and the GAF definition.
This research is supported by the European Unions 7th Framework Programme via the NewsReader Project (ICT-316404) and by the BiographyNet project, funded by the Netherlands eScience Center (http://esciencecenter.nl/). sem:Place
dbpedia:Bloomberg dbpedia:Veterans_Today rdf:type prov:at ributedTo
skos:exactMatch gaf:G2 dbpedTirau:nScuhndra_ skos:exact
Match gaf:INSTANCE_200 owl:objectProperty sem+:causes wn30:synset-stableadjective-1 gaf:INSTANCE_201 skos:exact
Match dbpedia:USS_Jimmy_Carter _(SSN_23) gaf:INSTANCE_202 sem:Event rdf:type
rdf:type rdf:type rdf:type rdf:type dbepaerdthiaq:u2a0k0e4__aInnddi_ants_uOncaemain_ wn30:synnosuent--1tsunami
sem:EventType sem:subEventOf sem:EventTypesem:subEventOf
gaf:INSTANCE_188 sem+:causes gaf:INSTANCE_186 taf:INSTANCE_MENTION_118 taf:causal_c taf:INSTANCE_MENTION_120 prov:at ributedTo rdfs:isDefinedBy
gaf:G3 dbpedia:Tectonic_Plate gaf:INSTANCE_197 sem:has Actor wn30:synset-shiftverb-4
rdf:type sem:EventType sem:EventType gaf:INSTANCE_181
gaf:causes sem:subEventOf gaf:INSTANCE_179 skos:exactMatch
gaf:denotedBy gaf:G4
gaf:denotedBy taf:INSTANCE_MENTION_112 taf:hasParticipant
_nsubj taf:INSTANCE_MENTION_40 str:anchorOf str:anchorOf "plates"@en "shift"@en