=Paper=
{{Paper
|id=Vol-1123/3
|storemode=property
|title=FRED as an Event Extraction Tool
|pdfUrl=https://ceur-ws.org/Vol-1123/paper3.pdf
|volume=Vol-1123
|dblpUrl=https://dblp.org/rec/conf/semweb/GangemiHPR13
}}
==FRED as an Event Extraction Tool==
https://ceur-ws.org/Vol-1123/paper3.pdf
FRED as an Event Extraction Tool
Aldo Gangemi1,2 , Ehab Hassan1 , Valentina Presutti2 , Diego Reforgiato
Recupero2
1
LIPN, Université Paris13-CNRS-SorbonneCité, France
2
STLab, ISTC-CNR, Rome-Catania, Italy.
Events are elusive entities; as the authors of [7] argue, even human annotators
do not agree on what is an event and what is its boundary in terms of the
extension of its participants, temporal and geospatial extent, etc.
More aspects of events appear when trying to recognize or extract them
from text: polarity of speaker’s judgment on events, negation, modality, relations
(temporal, causal, declarative, etc.) to other events, etc.
For example, the text:
The Black Hand might not have decided to barbarously assassinate Franz Ferdinand
after he arrived in Sarajevo on June 28th, 1914.
expresses three events (decide, assassinate, arrive), with Black Hand being a par-
ticipant in two of them, Franz Ferdinand in the third (arrive), a temporal extent for
the third (June 28th, 1914 ), and a relative temporal extent for the other two (given
the third’s extent and the past tense suffixes in the first and third), a geo-spatial ex-
tent (Sarajevo), a judgment with negative polarity on the second event (barbarously),
a negation (not) over the modality (might) modifying the first event, and an explicit
temporal relation between the second and third event (after ).
Extracting, logically representing, and connecting elements from a sentence is cru-
cial to create semantic applications that are event-aware. In addition, it’s important
to disambiguate as much as possible the entities and concepts expressed, in order to
make the extracted model linked, and to exploit the full power of the Semantic Web
and Linked Data.
FRED1 [5] is a tool to automatically transform knowledge extracted from text
into RDF and OWL, i.e. it is a machine reader [2] for the Semantic Web. It is event-
centric, therefore it natively supports event extraction. In a recent landscape analysis of
knowledge extraction tools [3], FRED has got .73 precision, .93 recall, and .87 accuracy,
largely better than the other tools attempting event extraction.
FRED is available as a RESTful API and as a web application. In its current form,
it relies upon several NLP components: Boxer2 for the extraction of the basic logical
form of text and for disambiguation of events to VerbNet, UKB3 or IMS4 or BabelNet
API5 for word sense disambiguation, and Apache Stanbol6 for named entity resolution.
1
http://wit.istc.cnr.it/stlab-tools/fred
2
http://svn.ask.it.usyd.edu.au/trac/candc/wiki/boxer
3
http://ixa2.si.ehu.es/ukb/
4
http://www.comp.nus.edu.sg/˜nlp/sw/
5
http://lcl.uniroma1.it/babelnet/
6
http://stanbol.apache.org
Fig. 1: A diagram showing the FRED graph for the Black Hand sentence.
FRED contains several functionalities for event extraction, which can be summa-
rized according to typical subtasks:
– Event identity: FRED focuses on events expressed by verbs, propositions, common
nouns, and named entities (typically proper nouns).
– Event classification: FRED uses Linked Data-oriented induction of types for the
identified events, reusing e.g. VerbNet7 , WordNet8 , DBpedia9 , schema.org, and
DOLCE10 as reference ontologies.
– Event unity: FRED applies semantic role labeling [4] to verbs and propositions in
order to detect event boundaries, and frame detection [1] for resolving roles against
a shared event ontology.
– Event modifiers: FRED extracts logical negation, basic modalities, and adverbial
qualities, applied to verbs and propositions, which can then be used as event judg-
ment indicators.
– Event relations: FRED relates events via the role structure of verbs and proposi-
tions, and extracts tense relations between them.
The beginning and the following sentences are used as a lead example for showing
FRED’s functionalities:
The Renaissance was a cultural movement that spanned in Italy from the 14th to
the 17th century. Some sources report that the Renaissance might have been started by
Greek scholars from Constantinople.
In the diagram from Figure 2, the following events are recognized, extracted, clas-
sified, and aligned to WordNet, VerbNet, and/or DOLCE: Renaissance (classified as a
Movement, and aligned to the WordNet Motion synset, and to the DOLCE Situation
class), span 1, report 1, and start 1 (classified as occurrences of the Span, Report
and Start frames respectively, and aligned to VerbNet).
Furthermore, the events have participants (e.g. Italy, scholar 1, source 1, etc.,
also classified and linked appropriately) through some roles labelled with properties
derived from VerbNet(e.g. vn:Agent), or from the lexicon used in the sentence (e.g.
ren:from) In one case, a modal modifier (Possible) to the event start 1 is added.
7
http://verbs.colorado.edu/˜mpalmer/projects/verbnet.html?
8
http://wordnet.princeton.edu
9
http://dbpedia.org
10
http://www.ontologydesignpatterns.org/ont/dul/DUL.owl
Finally, some relations between events are detected: report 1 vn:Theme start 1,
and span 1 before report 1 (through the now 1 interval).
See also Figure 1 for the graph obtained from the beginning sentence.
Fig. 2: A FRED graph depicting the core subset of triples representing event-related
knowledge.
The triples given as output by FRED are more than those visualized, for example
they include text spans and their reference to the semantic annotations, through the
Earmark vocabulary [6].
FRED is therefore an intermediate component for event extraction and representa-
tion, which can be augmented with background knowledge, and whose graphs can be
combined e.g. in time series for historical tasks.
FRED will be demoed as an event extractor by showing event-intensive sentences,
and examples of views that focus on relevant event knowledge. RDF models can be
morphed to concentrate on specific features. For example, Figure 3 semantically sum-
marizes the model from the Black Hand sentence by only showing events with their
relations, and their main participant, obtained by means of the following SPARQL
query:
PREFIX dul:
PREFIX vnrole:
PREFIX boxing:
PREFIX boxer:
PREFIX :
CONSTRUCT {?e :agent ?x . ?e ?r ?e1}
WHERE {
{{?e a boxing:Situation} UNION {?e a ?class . ?class rdfs:subClassOf+ dul:Event}}
?e ?p ?x
FILTER (?p = vnrole:Agent || ?p = boxer:agent || ?p = vnrole:Experiencer || ?p = vnrole:Actor
|| ?p = vnrole:Actor1 || ?p = vnrole:Actor2 || ?p = vnrole:Theme)
FILTER NOT EXISTS {?e vnrole:Theme ?x . ?e vnrole:Agent ?y
FILTER (?x != ?y)}
OPTIONAL {{{?e ?r ?e1} UNION {?e ?s ?z . ?z ?t ?e1}} {{?e1 a boxing:Situation} UNION
{?e1 a ?class1 . ?class1 rdfs:subClassOf+ dul:Event}} FILTER (?e != ?e1)}}
Fig. 3: A summarized FRED graph showing only event relations and agentive partici-
pants for the Black Hand sentence.
References
1. Bonaventura Coppola, Aldo Gangemi, Alfio Massimiliano Gliozzo, Davide Picca,
and Valentina Presutti. Frame detection over the semantic web. In Lora Aroyo et
al., editor, ESWC, volume 5554 of LNCS, pages 126–142. Springer, 2009.
2. Oren Etzioni, Michele Banko, and Michael Cafarella. Machine reading. In Proceed-
ings of the 21st National Conference on Artificial Intelligence (AAAI), 2006.
3. Aldo Gangemi. A comparison of knowledge extraction tools for the semantic web.
In Proceedings of ESWC2013. Springer, 2013.
4. A. Moschitti, D. Pighin, and R. Basili. Tree kernels for semantic role labeling.
Computational Linguistics, 34(2):193224, 2008.
5. Valentina Presutti, Francesco Draicchio, and Aldo Gangemi. Knowledge extraction
based on discourse representation theory and linguistic frames. In EKAW: Knowl-
edge Engineering and Knowledge Management that matters. Springer, 2012.
6. Peroni S., Gangemi A., and Vitali F. Dealing with markup semantics. In Pro-
ceedings of the 7th International Conference on Semantic Systems, Graz, Austria
(i-Semantics2011). ACM, 2011.
7. Chris Welty and Lora Aroyo. Harnessing disagreement for event semantics. In Proc.
of DERIVE Wks at ISWC2012, 2012.