Gathering lexical linked data and knowledge patterns from FrameNet.

Full text

Page 1

Gathering Lexical Linked Data and Knowledge Patterns
from FrameNet
Andrea Nuzzolese
ISTC-CNR, STLab
CS Dep., University of Bologna,
Italy
nuzzoles@cs.unibo.it
Aldo Gangemi
ISTC-CNR, Semantic Technology
Lab, Rome, Italy
aldo.gangemi@cnr.it
Valentina Presutti
ISTC-CNR, Semantic Technology
Lab, Rome, Italy
valentina.presutti@cnr.it
ABSTRACT
FrameNet is an important lexical knowledge base fea-
turing cognitive plausibility, and grounded in a large
corpus. Besides being actively used by the NLP com-
munity, frames are a great source of knowledge pat-
terns once converted into a knowledge representation
language. In this paper we present our experience in
converting the 1.5 XML version of FrameNet into RDF
datasets published on the Linked Open Data cloud,
which are interoperable with WordNet and other re-
sources. In the conversion we have used Semion, a new
tool that allows a rule-based, customized pipeline from
XML to RDF and OWL data. In addition, we introduce
a method to select and refactor part of the information
related to frames as full-fledged OWL knowledge pat-
terns. This last result has required non-trivial assump-
tions on how to interpret FrameNet relations as formal
knowledge.
Categories and Subject Descriptors
I.2.4 [Knowledge Representation Formalisms and
Methods]: Representations (procedural and rule-based),
Frames and scripts; I.2.6 [Learning]: Knowledge ac-
quisition
General Terms
Design, Experimentation, Theory
Keywords
Knowledge Extraction, FrameNet, OWL, Semantic Web,
Knowledge Engeneering
1.
The Web is evolving from a global information space
of linked documents to one where both documents and
INTRODUCTION
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
K-CAP’11, June 26–29, 2011, Banff, Alberta, Canada.
Copyright 2011 ACM 978-1-4503-0396-5/11/06 ...$10.00
data are linked. Underpinning this evolution is a set of
best practices for publishing and connecting structured
data on the Web known as Linked Data [3]. The Linked
Open Data (LOD) project is bootstrapping the Web of
Data by converting into RDF and publishing existing
datasets available under open licenses.
LOD is an ideal platform for empirical knowledge en-
gineering research, since it has the critical amount of
data for empirical research, data that are not necessarily
clean, optimized, or extensively structured. In practice,
it’s a perfect use case for making patterns emerge which
can be studied by knowledge engineering and used for
the design, maintenance, and consumption of data.
In addition, LOD datasets often contain a lot of natural
language text, which is also important in order to make
advanced linking and exploration of data not only in
the broad LOD cloud vision, but also in localized ap-
plications within large organizations that make use of
linked data [2].
Hybridizing natural language processing and semantic
web techniques has therefore become an important re-
search area. Part of the hybridization research, as well
as part of the exploitation of LOD data, is carried out
by means of lexical resources that are represented dir-
ectly as linked data. The major example is the WordNet
RDF dataset [21], which provides concepts (called syn-
sets), each representing the sense of a set of synonymous
words [10].
WordNet RDF has a low level of concept linking, be-
cause synsets are linked mostly by means of taxonomic
(hyponymy) relations, while LOD is mostly linked by
means of domain relations, such as parts of things, ways
of participating in events or socially interacting, top-
ics of documents, temporal and spatial references, etc.
Some lexical resources focus instead on domain relations
as expressed in the lexicon of natural languages.
This paper addresses hybridization research by porting
the largest lexical resource for domain relations, Frame-
Net [1], to the LOD cloud.
FrameNet was previously available only as a lexical data-
base, or as purely semantic web resources [8,20], derived
from the lexical one: previous conversions to RDF are
discussed in Sect. 5. After the release of version 1.5,

Page 2

the Berkeley FrameNet group asked us to produce a new
version of FrameNet in RDF that can be optimized for
use in the growing lexical part of the LOD cloud.
Some lexical resources focus instead on domain rela-
tions as expressed in the lexicon of natural languages.
This paper addresses hybridization research by porting
the largest lexical resource for domain relations, Frame-
Net [1], to the LOD cloud.
FrameNet is based on the notions of Semantic Frame,
Lexical Unit (LU), and Frame Element (FE): for ex-
ample, the Apply heat frame refers to situations in-
volving a Cook using a Heating instrument on some
Food within some Container, etc. These types of in-
volved entities are called FEs, and situations are ex-
pressed by words (lexemes) that manifest a LU, e.g.
fry, cook, roast, etc. All those LUs are lexical coun-
terparts of the semantic frame.
Intuitively, this is a more pragmatic and effective rep-
resentation of lexical meaning, because frames focus on
actual usage of language in real world situations, rather
than on decontextualized terms as in traditional dic-
tionaries (detailed analyses of the cognitive plausibility
of frames as meaning units, besides [1] itself, are [9,12]).
FrameNet was previously available only as a lexical data-
base, or as unlinked OWL resources [8,20], derived from
former versions of the lexical database: such conversions
are discussed in Sect. 5. After the release of version 1.5,
the Berkeley FrameNet group asked us to produce a new
version of FrameNet in RDF that can be optimized for
use in the growing lexical part of the LOD cloud.
Among lexical resources, FrameNet has been success-
fully employed in NLP applications that demonstrate
its potential to improve the quality of question answer-
ing [23] or recognizing textual entailment [6].
Frames as a cognitive, linguistic, or knowledge repres-
entation primitive have been studied many times in the
last century (see [9] for an overview). For example [13]
introduced frames into AI as a hub to factual and pro-
cedural knowledge: systems of interconnected frames
would provide the shifting perspectives or time-dependent
change in a situation. The intended meaning of a frame
across the different theories can be summarized as from [12]:
a (small-sized and richly interconnected) structure, used
to organize our knowledge, as well as to interpret, pro-
cess or anticipate information. In ontology design, frames
are called knowledge patterns [7,12], as a special kind
of design patterns.Following the approach outlined
in [12], we study frames as “units of meaning” for LOD
and semantic web ontologies.
The contribution of this paper is twofold: (i) the pro-
duction and publishing of a LOD dataset for the Fra-
meNet lexical database, and (ii) the description of a
method to produce knowledge patterns out of Frame-
Net frames. For both contributions we use Semion [15]:
a tool for “triplifying” non-RDF data into RDF models,
and for refactoring RDF into other RDF or OWL cus-
tomized models. The transformation process includes
two main steps: (i) a syntactic triplification of the ori-
ginal source and (ii) a rule-based refactoring for adding
semantics to triples.
FrameNet as a LOD dataset provides new blood to
the lexical grounding of semantic knowledge [12], and
boosts the “lexical linked data” section of LOD, by link-
ing FrameNet to other LOD datasets such as WordNet
RDF (section 3.1).
As a further contribution, we introduce a rule-based
method to select and refactor part of FrameNet as full-
fledged OWL knowledge patterns to be used for onto-
logy design and advanced exploration of LOD. We dis-
cuss some non-trivial assumptions on how to interpret
FrameNet relations as formal knowledge (section 4).
The structure of the paper is the following. In section
2 we summarize the conceptual design of FrameNet. In
section 1 we present the production of FrameNet as a
LOD dataset.Section 4 describes experiences in re-
factoring frames as knowledge patterns. Final sections
contain related work and conclusions.
2.
FrameNet [1] is a lexical knowledge base, consisting of
a set of frames, which have proper frame elements and
lexical units, which pair words (lexemes) to frames. As
described in the FrameNet Book [19]:
FRAMENET
a lexical unit (LU) is a pairing of a word with
a meaning. Typically, each sense of a poly-
semous word belongs to a different semantic
frame, a script-like conceptual structure that
describes a particular type of situation, ob-
ject, or event along with its participants and
properties. For example, the Apply Heat
frame describes a common situation involving
a Cook, some Food, and a Heating Instru-
ment, and is evoked by words such as bake,
blanch, boil, broil, brown, simmer, steam, etc.
We call these roles frame elements (FEs) and
the frame-evoking words are LUs in the Ap-
ply heat frame.
FrameNet has a rich internal structure and makes some
cognitive and semantic assumptions that makes it unique
as a lexical resource. Some of them are discussed in
Sect. 4 in view of logical formalization. The basic as-
sumptions are reported here: frame elements are mostly
unique to their frame; a frame usually has only some of
its roles actually lexicalized in texts; frames can be lex-
icalized or not: non-lexicalized ones typically encode
schemata from cognitive linguistics; frames, as well as
frame elements, are related between them, e.g. through
the subframe compositional relation, through inherit-
ance relations, etc.
The semantic part of FrameNet is enriched by semantic
types assigned to frames (e.g.
ments (e.g.
Sentient), and lexical units (e.g.
framal LU).
Artifact), frame ele-
Bi-

Page 3

FrameNet also contains a huge amount of manual an-
notations (annotation sets) of sentences from textual
corpora with frames, frame elements and lexical units,
which make word valences (syntactic and semantic com-
binatory of words) emerge.
3.
FrameNet stores lexical data into an XML database.
We pulled out the semantics of FrameNet and its data
by using Semion [15], a tool grounded on a method
with two main steps: (i) a syntactic and completely
automatic transformation of the data source to RDF
datasets according to an OWL ontology that repres-
ents the data source structure, i.e. the source meta-
model, (ii) a semantic refactoring that allows to express
the RDF triples according to specific domain ontolo-
gies e.g. SKOS, DOLCE, FOAF, the Gene Ontology,
or anything indicated by the user.
results in a RDF dataset, which expresses the know-
ledge stored in the original data source, according to
a set of assumptions on the domain semantics, as se-
lected and customized by the user.
step is the result of a non-trivial knowledge engineer-
ing work that requires a good knowledge of the target
domain semantics. For that reason the refactoring is
semi-automatic as it requires the design of transform-
ation rules by the user. More exhaustively, the refact-
oring is performed by means of transformation rules of
the form “condition → consequent” whose aim is to
apply a tansformation (specified in the consequent) in
the RDF graph only if the condition is satisfied with
respect to the knowledge expressed in the source RDF
graph. A set of rules which co-occur for the finalization
of a transformation process is called a transformation
recipe.
Figure 1 exemplifies the approach followed in this work
for the production of both the LOD dataset and a set
of knowledge patterns from the FrameNet lexical data-
base. That approach grounds completely on the trans-
formation method that the Semion tool implements.
The leftmost part of the figure depicts the logical lay-
ers of the original source, i.e. FrameNet, which con-
tains XML data, whose structure is defined by some
XSD, which has its own standard meta-model [25]. The
second leftmost part contains the result of step i): syn-
tactic transformation of FrameNet to pure RDF triples,
whose aim is twofold: (a) extracting data into RDF
and (b) flattening the distinction between the original
schema and data in order to provide, via the refactor-
ing, a customized, task-oriented way to address domain
semantics.
Those RDF triples are refactored according to trans-
formation recipes, either as a LOD dataset (ABox Re-
factoring, second rightmost column in the figure) or
as Knowledge Patterns (TBox Refactoring, rightmost
column).
BRINGING FRAMENET TO LOD
This last action
The refactoring
3.1FrameNet linked data
Figure 1: Semion tranformation: key concepts.
The approach followed for the creation of a LOD data-
set from FrameNet1is both derived from the transform-
ation method implemented by Semion [15] and based
on an iterative evaluation of the quality of the output
produced with respect to the semantics of FrameNet
formalized into a “gold standard” ontology2that we
have used for the evaluation. Based on that, the trans-
formation of FrameNet v.1.5 from XML to RDF con-
sisted into two steps: (i) the syntactic transformation
of the XML source to RDF according to the OWL meta-
model that describes the structure of the source3, (ii)
the design and the application of a refactoring recipe for
the ABox refactoring on the RDF produced in the first
step. The recipe was derived generalizing and revising
some of the common transformation practices from ex-
isting tools (i.e. XML2OWL [4], TopBraid Composer4,
Rhizomik ReDeFer [18]). For example we used the fol-
lowing mappings: (i) a xsd:ComplexType is mapped
to an owl:Class, (ii) a xsd:SimpleType is mapped
to an owl:DatatypeProperty and (iii) a xsd:Element
is mapped either to an owl:ObjectProperty or to a
owl:DatatypeProperty. Further details can be found
in [4].
As an example, according to the syntax of the rules for
the Semion refactoring, we have that the mapping (i) is
expressed as
is(oxsd:ComplexType, ?type)
->
is(owl:Class, ?classNode)
and maps any individual of the class oxsd:ComplexType
to a owl:Class. We refer to the Semion wiki5for more
information about the tool and the syntax of the rules.
The relavance of a syntanctic transformation and a fol-
lowing refactoring can be clarified saying that it is de-
signed as a semi-automatic approach which allows, via
1The dataset can be accessed through the SPARQL endpoint at
http://bit.ly/fnsparql, as framenet dataset
2http://ontologydesignpatterns.org/cp/owl/fn/framenet.owl
3http://www.ontologydesignpatterns.org/ont/iks/oxml.owl
4http://www.topbraidcomposer.com
5http://stlab.istc.cnr.it/stlab/Semion

Page 4

the refactoring rules, for better addressing the domain
semantics of the original source. As an example, we can
consider a simple frame-to-frame relation like
Inherits from(Abounding with, Locative relation)
which expresses the fact that the frame Abounding with
inherits the schematic representation of a situation in-
volving various participants, properties, and other con-
ceptual roles from the frame Locative relation. This re-
lation is expressed in the XML FrameNet notation as:
<frame name="Abounding_with" ... ID="262">
...
<frameRelation type="Inherits from">
<relatedFrame>
Locative_relation
</relatedFrame>
</frameRelation>
...
</frame>
and, with most of the existing tools, it is transformed
to the RDF schematized in Figure 2. It is easy to notice
how the Inherits from frame-to-frame relation is real-
ized through the reification of the relation RelatedFrame i,
that expresses its type and the related frames, i.e. Inherits
from and Locative relation, as literals.
Figure 2: The “Inherits from” relation mapped
to RDF with a common transformation recipe.
Instead, adopting the syntactic transformation of Se-
mion, we have produced firstly an RDF graph, which is
depicted in Figure 36.
In the figure, fntbox:Frame is no longer an owl:Class,
but an oxsd:Element and fnabox:Abounding with is an
oxml:XMLElement related to fntbox:Frame through
oxsd:hasElementDeclaration.
After having was syntactically converted FrameNet from
XML to RDF, we applied the general recipe with the
Semion Refactorer, in order to derive a LOD dataset
6oxsd and oxml are the default ontologies of Semion for XSD and
XML data structures.
Figure 3: Example of reengineering of the frame
“Abounding with” with its XSD definition.
for FrameNet. As the recipe is based on a general con-
version from XML to OWL, the result was far from
being a good formalization of the semantics of Frame-
Net. For that reason, we have incrementally refined the
recipe in order to fill the gap between the semantics
expressed by the output produced by the refactoring
and the gold standard we had previously defined. We
remark that the aim of the refactoring is to transform
one RDF source to another trying to preserve either ex-
plicit or implicit domain semantics of the original source
without information loss.
For example, the rule which allows to avoid the re-
ification of frame-to-frame relations is shown in Fig-
ure 4. The rule shown in Figure 4 transforms all the
Figure 4: Rule which allows to express frame-
to-frame relation as binary relations.
frame-to-frame relations into binary relations between
frames. The rule extracts the type of the relation from
the nodeValue associated with the type attribute of a
frame. Then it creates a new object property as a sub-
property of hasFrameRelation, and resolves the name
of the related frame that is expressed as a literal in the
relatedFrame element, as shown in the XML code be-
fore. We remark that the model accessed by rules is

Page 5

not anymore the original XML source, but its syntactic
translation to RDF. Figure 5 shows the RDF of the
inherits from relation between the frames Abounding
with and Locative relation obtained by applying the
refactoring recipe with Semion. Figure 6 shows the core
Figure 5: The “Inherits from” frame-to-frame
relation between the frames “Abounding with”
and “Locative relation” after the refactoring.
fragment of the OWL schema of FrameNet used as a
vocabulary for the data from FrameNet.
Figure 6: A fragment of the FrameNet OWL
schema.
The complete refactoring recipe7is composed by 58
transformation rules in forward-chaining inference mode.
An important feature of FrameNet as a dataset in the
LOD cloud , that will be investigated as part of our on-
going work, is the mapping of its frames and frame ele-
ments to other lexical resources, e.g. WordNet. Word-
Net is available as a LOD dataset since 2006 as a result
of the W3C working draft [21].
be obtained from VerbNet [22], a lexical resource that
Such mappings can
7http://stlab.istc.cnr.it/stlab/FrameNetKCAP2011#
tab=ABoxRefactoring
incorporates both semantic and syntactic information
about English verb semantics. The VerbNet 3.1 XML
database provides mappings between VerbNet classes,
FrameNet frames, and WordNet synsets.
For example, from the VerbNet mappings converted to
RDF:8
vnclass:accompany skos:exactMatch
wndata:synset-accompany-verb-2
vnclass:accompany skos:exactMatch
frame:Cotheme
The VerbNet dataset excerpt is intended to demonstrate
linkings between lexical resources. An official release
will be published in the near future.
4. FROMFRAMESTOKNOWLEDGEPAT-
TERNS
In addition to the production of FrameNet as a LOD
lexical dataset that can be accessed and queried over
the Web of Data, our aim is to provide an interpret-
ation of frames as Knowledge Patterns (KP), as they
are defined by [7] and [12]. In other words, following
[12], we promote frames to relevant units of meaning for
knowledge representation.
With reference to Figure 1, we have called this process
TBox refactoring, because a new ontology schema (a
TBox), is obtained starting from data (ABox).
The main problem with TBox refactoring is deciding the
formal semantics to assign to the classes from the Fra-
meNet LOD dataset schema. Since this is a relatively
arbitrary process, SemionRules and recipes are useful
to configure alternative choices or to compare the dif-
ferent assumptions made by knowledge engineers. Here
we present a refactoring experience that exemplifies the
design method behind such process, and how Semion is
useful in supporting it. The recipe exemplified here is
part of a larger project carried out together with Fra-
meNet developers in Berkeley in order to optimize the
refactoring from lexical frames to knowledge patterns:
as such, it certainly bears validity, but it is mainly inten-
ded as a methodological and pragmatical description of
refactoring recipes (also called correspondence patterns
in [24]).
Besides the basic assumptions reported in section 2, this
process is guided by the Book [19], which is quite expli-
cit about possible formal semantic choices:
The most basic summarization of the logic
of FrameNet is that Frames describe classes
of situations, the semantics of LUs are sub-
classes of the Frames, and (...)FEs are
8prefixes:
class:
wndata:
frame:
frame/
skos:
http://www.ontologydesignpatterns.org/ont/vn/class/;
http://www.w3.org/2006/03/wn/wn20/instances/;
http://www.ontologydesignpatterns.org/ont/framenet/
http://www.w3.org/2004/02/skos/core#; vn-

Page 6

classes that are arguments of the Frame
classes. An annotation set for a sentence gen-
erally describes an instance of the subclass
associated with an LU as well as instances of
each of its associated FE classes (...) The
term “Frame Element” has two meanings:
the relation itself, and the filler of the rela-
tion. When we describe the Coreness status
of an FE (...) we are describing the relation;
when we describe the Ontological type on an
FE (...) we mean the type of the filler.
According to these statements, a fragment of the De-
siring frame is transformed into OWL as follows (in
Manchester syntax):
Ontology: odpfn:desiring.owl
Annotations:
cpannoschema:specializes odp:situation.owl
Class: desiring:Desiring
SubClassOf:
desiring:hasEvent some desiring:Event,
desiring:hasExperiencer some desiring:Experiencer,
desiring:hasDegree some desiring:Degree,
desiring:hasReason some desiring:Reason,
Class: desiring:covet.v
SubClassOf: desiring:Desiring
Class: desiring:Event
SubClassOf: semtype:State_of_Affairs
Class: desiring:Experiencer
SubClassOf: semtype:Sentient
Notice that the uniqueness (locality) of frame elements
and lexical units for a frame is obtained simply by
means of a specific namespace (denoted by the desiring
prefix in the example, see below for possible namespace
policies), while a frame is interpreted as an owl:Class,
lexical units as its subclasses, frame elements as both
an owl:Class (e.g. Event) and an owl:ObjectProperty
(e.g. hasEvent), the relation between a frame and a
frame element as a rdfs:subClassOf an owl:Restriction,
and the semantic type assignments to frame elements as
additional subclass axioms. All knowledge patterns de-
rived from frames are considered specialization of the
generic pattern odp:situation.owl9, which general-
izes the situation semantics suggested by Berkeley lin-
guists.
A central role in FrameNet is played by inheritance as-
sumptions. In [19], inheritance is viewed as
the strongest relation between frames, corres-
ponding to is-a in many ontologies. With this
relation, anything which is strictly true about
the semantics of the Parent must correspond
to an equally or more specific fact about the
Child. This includes Frame Element mem-
bership of the frames (except for Extrathem-
9odp:http://www.ontologydesignpatterns.org/cp/owl/,
odpfn:http://www.ontologydesignpatterns.org/cp/owl/fn/
atic FEs), most Semantic Types, frame re-
lations to other frames, relationships among
the Frame Elements, and Semantic Types on
the Frame Elements.
This means that additional axioms must be wrapped
into ontologies derived from frames, e.g.
sample axioms are derived from the inheritsFrom rela-
tion between the Aesthetics and Desirability frames
as well as from the subFE relation between some of their
frame elements:
these two
Ontology: odpfn:aesthetics.owl
Annotations:
cpannoschema:specializes odpfn:desirability.owl
Class: aesthetics:Aesthetics
SubClassOf: desirability:Desirability
Class: aesthetics:Degree
SubClassOf: desirability:Degree
The implementation of TBox refactoring is performed
as a Semion refactoring, where the recipe includes rules
for the mapping between FrameNet LOD dataset and
KPs. Figure 7 shows an overview of TBox refactoring
for deriving KPs from frames. The notation attempts to
Figure 7: Diagram of the transformation recipe
used for the production of knowledge patterns
from FrameNet LOD.
make rules intuitively understandable: arrows between
the clouds represent mappings from entities in the cloud
“FrameNet as LOD” to entities in the cloud “Knowledge
Pattern”, classes are represented as circles, individuals
as triangles, object properties as diamonds, and struc-
tural properties as labeled arcs. Each Frame is mapped
both to an owl:Ontology that identifies the KP and
to an owl:Class. The mapping takes into account the
refactoring of the frame URI intended either as an on-
tology or as a class. Each FrameElement maps both to
an owl:Class and to an owl:ObjectProperty. Again
frame elements follow a renaming policy for the two
interpretations, but in this case the situation is more
complex. In fact, URI policy can follow from different
interpretations:
1. Locality of frame elements within their frames

Page 7

(compatible to locality statements in the Book,
with some exceptions that cannot be discussed
here). E.g. given the frame:
http://someuri/Judgment.owl#Judgment
we obtain the frame element:
http://someuri/Judgment.owl#Cognizer
interpreted as a class and
http://someuri/Judgment.owl#hasCognizer
interpreted as an object property;10
2. Globality of frame elements, abstracted from their
contextual binding to a frame, e.g.
frame:
http://someuri/Judgment.owl#Judgment
we obtain the frame element:
http://someuri/class/Cognizer
interpreted as a class and
http://someuri/property/hasCognizer
interpreted as an object property.
given the
Lexical units are refactored as subclasses of the classes
derived from the frames they are lexicalizations of, e.g.
lexunit:cool.a SubClassOf:desirability:Desirability
Lexemesarerefactored
class
semantics:Expression;
is relatedtoalexeme
semantics:isExpressedBy.
Finally, each frame has owl:someValuesFrom restric-
tions accounting for the semantic roles implicit in
frame elements (see example above).
Locality and globality alternatives required two re-
factoring recipies each of one composed by 4 rules in
forward-chaining inference mode. The complete TBox
refactoring recipe can be found in the wiki page11.
asindividuals
each
through
ofthe
unitlexical
the property
5.
The literature in reusing FrameNet for NLP tasks such
as question-answering is too large to be covered here,
and not central to the work described (see e.g. [23] [6]).
Work in using FrameNet jointly with other lexical re-
sources, although not in the LOD way, include at least
[5], which creates a linking from WordNet to FrameNet
in a purely NLP context.
Previous FrameNet conversions to RDF include [8,14,
20]. [14] proposes a partial translation of FrameNet ver-
sion 1.2 to RDF, and uses DAML both for the FrameNet
meta-model, and the conceptual elements (frames, ele-
ments, etc.). They developed an automatic translator
specific to that purpose. In 2003, the mixing of meta-
model and FrameNet data made it difficult to be pro-
cessed by reasoners for OWL (but it’d be acceptable in
OWL2). For that reason, [20] applied an ad-hoc XSLT
to move part of the FrameNet version 1.3 XML database
to OWL. While the quality of the partial transformation
RELATED WORK
10An OWL2 alternative is also possible, with multiple interpreta-
tions for the same constant.
11http://stlab.istc.cnr.it/stlab/FrameNetKCAP2011#
tab=TBoxRefactoring
is high, the process is not easily customizable. [20] also
proposes a solution to deductive reasoning with natural
language based on combining lexical resources with the
world knowledge provided by ontologies.
After the release of version 1.5, the Berkeley FrameNet
group asked us to produce a new version of FrameNet
in RDF, optimized for publishing in the growing lexical
part of the LOD cloud. This is what we describe in this
paper.
From the viewpoint of formal semantic interpretation of
FrameNet, [8] uses both ABox and TBox conversions to
perform automatic enrichment of FrameNet with refer-
ence to a large corpus where frames are detected, new
frames and elements are discovered and typed with a
WordNet Supersense learner, and finally reengineered
through a previous alignment to the LMM semiotic on-
tology [17] (used in this paper with the odp:semiotics.owl
knowledge pattern.
[9] is a deeper analysis of the semiotic relations be-
hind FrameNet, VerbNet and WordNet, and proposes
a method to formalize their mappings. The semantics
of the frames is put in perspective with the Descrip-
tions and Situations knowledge pattern, partly reused
in this paper to represent the situation-based semantics
declared by [19]. The article also proposes to represent
the full semantics of frames as n-ary polymorphic rela-
tions in FOL. This proposal is not directly implement-
able in OWL, but provides a useful abstraction across
the different notions of a frame in cognitive science, AI,
linguistics, knowledge engineering, etc.
[16] is an attempt to formalize and “clean” the semantics
of FrameNet version 1.3.
cleansing need by performing “ontological analysis”: e.g.
they claim that frames do not always refer to situation
classes because some of them actually represent ab-
stract relations such as part of : since abstract entities
should be assumed as non-localized, non-temporal entit-
ies, while situations should be interpreted as events oc-
curring in time, frames should be formalized differently
according to their ontological type. Frame to frame re-
lations are also suggested an extensive revision on sim-
ilar grounds. This work, besides the problem of shar-
ing agreement on the general principles adopted for the
analysis, could benefit from a customized refactoring of
FrameNet, in order to perform their analyses directly
on formal ontologies.
The authors motivate the
6.
We have presented a conversion of FrameNet to RDF,
published a dataset in the LOD cloud, linked to Word-
Net and other lexical datasets. We have also presented
a method to convert FrameNet data into knowledge pat-
terns. For both projects, we have employed the Semion
tool with SemionRules, which allows a customized and
explicit transformation from RDB or XML to RDF and
OWL.
The intricate semantics of FrameNet, only partly de-
scribed in this paper, gets to grips with the expressive
DISCUSSION

Page 8

power of natural language. A fixed, ad-hoc transform-
ation would be best for one, arbitrary for another, bad
for a third. Customization is key with lexical data be-
cause there are use cases for maintaining the semantics
of the original resource, often a purely intensional one
(similar to the practice of using SKOS with thesauri),
as well as for morphing the original semantics to some-
thing closer to the extensional formal semantics of web
ontologies. In between these two ends, there are sev-
eral intermediate cases and exceptions, which make the
case for tools that minimize hard-coding of the trans-
formation semantics, and preserve the opportunity to
learn and share good practices for transforming lexical
resources to linked data and domain knowledge.
Current ongoing work concentrates on refinement of the
RDF dataset with the Berkeley FrameNet group, the
generation of new links to lexical datasets as well as
other relevant LOD datasets (e.g. DBpedia), the cre-
ation of the FrameNet valence dataset, which will be a
substantial (about 35 million triples) resource for hy-
bridizing semantic web and linked data, and the refine-
ment of a recipe to produce and automatically pub-
lish FrameNet-based knowledge patterns on the ODP
portal12. These knowledge patterns implement a large
section of the rich knowledge pattern structure envis-
aged by [12], with formal axioms, lexically motivated
vocabulary, textual corpus grounding, and data ground-
ing.
Acknowledgements
This work has been part-funded by the European Commission
under grant agreement FP7-ICT-2007-3/ No. 231527 (IKS - In-
teractive Knowledge Stack).
7.
[1] C. F. Baker, C. J. Fillmore, and J. B. Lowe. The berkeley
framenet project. In Proceedings of the 17th international
conference on Computational linguistics, pages 86–90,
Morristown, NJ, USA, 1998.
REFERENCES
[2] C. Baldassarre, E. Daga, A. Gangemi, A. M. Gliozzo,
A. Salvati, and G. Troiani. Semantic scout: Making sense
of organizational knowledge. In P. Cimiano and H. S.
Pinto, editors, EKAW, volume 6317 of Lecture Notes in
Computer Science, pages 272–286. Springer, 2010.
[3] C. Bizer, T. Heath, and T. Berners-Lee. Linked Data - The
Story So Far. International Journal on Semantic Web and
Information Systems, 5(3):1–22, 2009.
[4] H. Bohring and S. Auer. Mapping xml to owl ontologies. In
Proceedings of 13. Leipziger Informatik-Tage (LIT 2005),
Sep. 21-23, Lecture Notes in Informatics (LNI), September
2005.
[5] A. Burchardt, K. Erk, and A. Frank. A WordNet Detour to
FrameNet. In Proceedings of the GLDV 2005 GermaNet II
Workshop, pages 408–421, Bonn, Germany, 2005.
[6] A. Burchardt and M. Pennacchiotti. FATE: a
FrameNet-Annotated Corpus for Textual Entailment. In
LREC. European Language Resources Association, 2008.
[7] P. Clark, J. Thompson, and B. Porter. Knowledge
Patterns. In A. G. Cohn, F. Giunchiglia, and B. Selman,
editors, KR2000: Principles of Knowledge Representation
12http://www.ontologydesignpatterns.org, cf. [11]
and Reasoning, pages 591–600, San Francisco, 2000.
Morgan Kaufmann.
[8] B. Coppola, A. Gangemi, A. M. Gliozzo, D. Picca, and
V. Presutti. Frame detection over the Semantic Web.
Paper, ESWC,, 2009.
[9] A. Gangemi. What’s in a Schema? In C. Huang,
N. Calzolari, A. Gangemi, A. Lenci, A. Oltramari, and
L. Prevot, editors, Ontology and the Lexicon. Cambridge
University Press, 2010.
[10] A. Gangemi, R. Navigli, and P. Velardi. The OntoWordNet
Project: Extension and Axiomatization of Conceptual
Relations in WordNet. In R. Meersman and Z. Tari,
editors, Proc. of On the Move to Meaningful Internet
Systems (OTM2003) (Catania, Italy), pages 820–838.
Springer-Verlag, 2003.
[11] A. Gangemi and V. Presutti. Ontology Design Patterns. In
S. Staab and R. Studer, editors, Handbook on Ontologies,
2nd Edition. Springer Verlag, 2009.
[12] A. Gangemi and V. Presutti. Towards a Pattern Science for
the Semantic Web. Semantic Web, 1(1-2):61–68, 2010.
[13] M. Minsky. A Framework for Representing Knowledge. In
P. Winston, editor, The Psychology of Computer Vision.
McGraw-Hill, 1975.
[14] S. Narayanan, M. R. L. Petruck, C. F. Baker, and C. J.
Fillmore. Putting framenet data into the iso linguistic
annotation framework. In Proceedings of the ACL 2003
workshop on Linguistic annotation, LingAnnot 03, pages
22–29, Stroudsburg, PA, USA, 2003. ACL.
[15] A. G. Nuzzolese, A. Gangemi, V. Presutti, and
P. Ciancarini. Fine-tuning triplification with Semion. In
V. Presutti, V. Svatek, and F. Sharffe, editors, Wks. on
Knowledge Injection into and Extraction from Linked Data
(KIELD2010), pages 2–14, Lisbon, Portugal, October 2010.
[16] E. Ovchinnikova, L. Vieu, A. Oltramari, S. Borgo, and
T. Alexandrov. Data-driven and ontological analysis of
framenet for natural language reasoning. In Proceedings of
LREC’10, Valletta, Malta, may 2010. ELRA.
[17] D. Picca, A. M. Gliozzo, and A. Gangemi. LMM: an
OWL-DL MetaModel to Represent Heterogeneous Lexical
Knowledge. In LREC. European Language Resources
Association, 2008.
[18] Rhizomik. ReDeFer. http://rhizomik.net/html/redefer,
2011. (accessed 15-02-2011).
[19] J. Ruppenhofer, M. Ellsworth, M. R. L. Petruck, C. R.
Johnson, and J. Scheffczyk. FrameNet II: Extended Theory
and Practice.
http://framenet.icsi.berkeley.edu/book/book.html, 2006.
[20] J. Scheffczyk, C. F. Baker, and S. Narayanan. Reasoning
over Natural Language Text by means of FrameNet and
Ontologies. In C. Huang, N. Calzolari, A. Gangemi,
A. Lenci, A. Oltramari, and L. Prevot, editors, Ontology
and the Lexicon. Cambridge University Press, 2010.
[21] G. Schreiber, M. van Assem, and A. Gangemi. RDF/OWL
Representation of WordNet. W3C Working Draft, W3C,
June 2006.
http://www.w3.org/TR/2006/WD-wordnet-rdf-20060619/.
[22] K. K. Schuler. VerbNet: A Broad-Coverage, Comprehensive
Verb Lexicon. PhD thesis, University of Pennsylvania, 2006.
[23] D. Shen and M. Lapata. Using Semantic Roles to Improve
Question Answering. In EMNLP-CoNLL, pages 12–21.
ACL, 2007.
[24] O. Sv´ ab-Zamazal, V. Sv´ atek, and F. Scharffe.
Pattern-based ontology transformation service. In J. L. G.
Dietz, editor, KEOD, pages 42–47. INSTICC Press, 2009.
[25] P. Walmsley and D. C. Fallside. XML Schema Part 0:
Primer Second Edition. W3C recommendation, W3C,
October 2004. http://www.w3.org/TR/2004/REC-
xmlschema-0-20041028/.