from Dependency Structures
Udo Hahn Martin Romacker
Text UnderstandingLab, Group,
FreiburgUniversity, D-79085Freiburg,Germany
http://www.coling.uni-freibu rg.de /
Abstract
We propose atwo-layeredmodel forcomputing
se-manticand conceptual interpretations from
depen-dencystructures. Abstractinterpretationschemata
generate semantic interpretations of `minimal'
de-pendency subgraphs, while production rules whose
specication is rooted in ontological categories
de-rivea canonical conceptual interpretation from
se-mantic interpretation structures. Congurational
descriptionsof dependencygraphsincreasethe
lin-guistic generalityof interpretation schemata, while
interfacingschemataandproductionstolexicaland
conceptualclasshierarchiesreducestheamountand
complexityofsemanticspecications.
1 Introduction
The syntax/semantics interface has always been a
matter of concern for constituency-based feature
grammar theories (cf., e.g., Creary and Pollard
(1985),Moore(1989),Dalrymple(1992),Wedekind
andKaplan(1993)). Withinthedependency
gram-marcommunity,farlessattentionhasbeenpaidto
this topic. As aconsequence,thereis noconsensus
how syntactic dependency structures might be
ad-equately transformed into semanticinterpretations
(cf., Hajicova(1987),Milward(1992),Lombardoet
al. (1998)foralternativeproposals).
In this paper, we introduce a two-layered
inter-pretationmodel. Inarst pass,dependencygraph
structureswhichresultfromincrementalparsingare
immediately submitted to a semantic
interpreta-tion process. Such a process is triggered by
gen-eralschematawheneverasemanticallyinterpretable
subgraphofasyntacticdependencygraphbecomes
available (cf. Section 3). As a result, lexical items
andthedependencyrelationsholdingbetweenthem
are directly mapped to associated conceptual
enti-tiesand relationsat thelevelof semantic
represen-tation(cf. Sections 4and5). Inasubsequentstep,
the(quasi-inferential)implicationsoftheknowledge
representationstructuresemergingfromthe
seman-tic interpretation step are accounted for by a
pro-cess we here refer to as conceptual interpretation.
tual representation level only and are triggeredby
a variety of production rules rooted in ontological
categories in order to generate acanonical
concep-tualrepresentation of theparsed sentence (cf.
Sec-tion6). Thissecondlevelofinterpretationisusually
nottakenintoconsiderationbycomputational
mod-elsofsemanticinterpretation,neither
constituency-basednordependency-basedones,althoughitturns
outtocrucial fornaturallanguageunderstanding.
2 Grammar and Concept Knowledge
Grammatical knowledge for syntactic analysis is
based on a fully lexicalized dependency grammar
(Hahnetal.,1994). Ourpreference fordependency
structuresismotivated,amongotherthings,bythe
observationthat thecorrespondenceof dependency
relations(holding betweenlexical items)to
concep-tual relations (holding between the concepts they
denote)is much closer than for constituency-based
grammars(Hajicova, 1987). Hence, a
dependency-based approach eases inherently the description of
theregularitiesunderlyingsemanticinterpretation.
Inthislexicalizeddependencyframework,lexeme
specicationsform theleafnodesofalexiconDAG,
which are further abstracted in terms of lexeme
class specications at dierent levels of generality
(cf. Figure 1). This leads to a lexeme class
hier-archy, which consists of lexeme class names W :=
fverbal,verbintrans, nominal,noun, ...gand
a subsumption relation isa
W
= f(verbintrans,
(with)
genatt
genatt
(memory)
(deliver)
subj
subj
ppadj
ppadj
dirobj
dirobj
werden_passive
R := {patient co-patient}
+
mit
+
R := {has-part instrument co-patient ..}
+
D := { }
-D := { }
Speicher
Verbal
...
VerbIntrans
VerbTrans
liefern
+
-+
D := { }
D := {subj, dirobj, indirobj}
-+
-D := {subj}
D := { }
D := {dirobj}
D := { }
Lexeme
Auxiliary
Preposition
Nominal
Pronoun Noun
...
spec:
genatt:
spec:
ppatt:
Die
Festplatte
des
Computers
mit
350Mhz-CPU
kann
von
geliefert
werden
.
propo:
Transtec
The hard disk of the computer with 350Mhz-CPU can - by Transtec- be delivered.
pobject:
ppsubject:
verbpart:
verbpart:
subject:
pobject:
Figure2: ASample DependencyGraph
verbal),(noun,nominal),...gWW.
Inher-itance of grammar knowledge is based on the idea
that constraints are attached to the most general
lexemeclassestowhichtheyapply,leavingroomfor
moreandmorespecic(possibly,evenidiosyncratic)
specicationswhenonedescendsthishierarchy.
A dependency grammar captures binary
con-straintsbetweenasyntactichead(e.g.,anoun)and
one of its possible modiers (e.g., a determiner or
an adjective). In order to establish a dependency
relationÆ 2D :=fspecier, subject,dirobject, ...g
betweenaheadandamodier,lexeme-class-specic
constraints on word order, compatibility of
mor-phosyntactic features and semantic integrity must
befullled. Figure2depictsadependencygraphin
whichwordnodesaregivenin boldfaceand
depen-dencyrelationsare indicatedbylabelededges.
Conceptualknowledgeoftheunderlyingdomainis
expressedintermsofaKl-One-likeknowledge
rep-resentation language(Woods and Schmolze, 1992).
The domain ontology consists of a set of concept
namesF:=fCompany,Hard-Disk,...ganda
sub-sumptionrelationisa
F
=f(Hard-Disk,
Storage-Device), (Transtec, Company), ...g F
F. The set of relation names R := fhas-part,
deliver-agent, ...g denotes conceptual relations
whicharealsoorganizedinasubsumptionhierarchy
isa
R
= f(has-hard-disk, has-physical-part),
(has-physical-part, has-part), ...g. 1
Examples
ofemergingconceptandrelationhierarchiesare
de-pictedinFigure3(rightbox).
Inourapproach,therepresentationlanguagesfor
semanticsanddomainknowledgecoincide(for
argu-mentssupportingthisview,cf.Allen(1993)).
Link-ing lexical items and conceptual entities proceeds
asfollows: Uponenteringtheparsingprocess, each
lexical itemw thathas aconceptualcorrelateC in
thedomainknowledgebase,w:C2F (mostlyverbs,
nounsandadjectives),getsimmediatelyinstantiated
intheknowledgebase,suchthatforanyinstanceI
w ,
initially, 2
type(I
w
) = w:C holds (e.g., w =
\Fest-platte", I
w
= Hard-Disk.2, w:C =
type(Hard-Disk.2) =Hard-Disk). If severalconceptual
cor-relatesexist,either dueto homonymyor polysemy,
1
All subsumption relations, isa
W , isa
F
, and isa
R , are
consideredtobetransitiveandreexive.
2
Forinstance,anaphoramightnecessitatechangesofthis
initialreferenceassignment,cf.StrubeandHahn(1999).
Figure3: RelatingGrammatical(leftbox)and
Con-ceptualKnowledge(rightbox)
each lexical ambiguity is processed independently
withinseparatecontextpartitions oftheknowledge
base(RomackerandHahn, 2000a).
3 Interpretable Subgraphs
Intheparse treefrom Figure 2,wecandistinguish
lexicalnodesthat haveaconceptualcorrelate(e.g.,
\Festplatte"relating to Hard-Disk, \geliefert"
re-lating to Delivery) from others that do nothave
suchacorrelate(e.g.,\mit"(with),\von"(by)).
Se-manticinterpretation capitalizeson thisdistinction
in order to nd adequate conceptual relations
be-tweenthecorrespondingconceptinstances:
Direct Linkage.Iftwowordnodeswith
concep-tualcorrelatesarelinkedbyasingledependency
re-lation, a direct linkage is given. Such a subgraph
canimmediately be interpreted in terms of a
con-ceptual relation licensed by the corresponding
de-pendencyrelation. ThisisillustratedinFigure2by
the direct linkage between\Festplatte" (hard disk)
and \Computers" via the gen[itive]att[ribute]
rela-tion,which getsmapped to thehard-disk-ofrole
linkingthecorrespondingconceptualcorrelates,viz.
Hard-Disk.2 and Computer-System.4,
respec-tively(see Figure 4). This interpretation usesonly
knowledge aboutthe conceptualcorrelatesand the
linkingdependencyrelation.
Indirect Linkage.If twoword nodes with
con-ceptual correlates are linked via aseries of
Depen-have a conceptual correlate, an indirect linkage is
given. Forsucha\minimal"subgraph,semantic
in-terpretation is made dependent onlexical
informa-tionfromtheinterveningnodes,aswellasknowledge
abouttheconceptualcorrelatesanddependency
re-lations. Figure 2 illustrates such a conguration
bythelinkagebetween\Computers"and
\350Mhz-CPU" via the intervening node \mit" (with) and
theppatt[ribute]andpobjectrelations,theresultof
whichis aconceptuallinkage between
Computer-System.4and350Mhz-Cpu.6viatherelation
has-cpuin Figure4.
Inordertoincreasethegeneralityandtopreserve
the simplicity of semantic interpretation we
intro-duce a generalization of the notion of dependency
relation such that it incorporates direct aswell as
indirect linkage: Twocontentwords(nouns,
adjec-tives,adverbsorfullverbs)standinamediated
syn-tactic relation, if one can pass from one word to
theother along theconnecting edges ofthe
depen-dency graph without traversing word nodes other
thanprepositions, modalorauxiliaryverbs(i.e.,
el-ements of closed word classes). In Figure 2, e.g.,
the tuples (\Festplatte", \Computers") or
(\Com-puters",\350Mhz-CPU")standinmediated
syntac-tic relations, whereas,e.g., the tuple (\Festplatte",
\Transtec")doesnot,sincetheconnectingpath
con-tains\geliefert"(delivered),acontentword.
This leadsto the followingdenition: Let wand
w 0
betwocontentwordsin asentence S. In
addi-tion,let w
2 ;:::;w
n 1
2S (n2) be prepositions,
auxiliaryor modal verbs, and w
1
standinamediated
syntactic relation, i there exists an index l 2 f1,
:::, ngso thatthefollowingtwoconditionshold:
1. w
Wecallasubgraphidentiedbysuchaseriesw
1 ,
:::,w
n
asemantically interpretablesubgraphofthe
dependency graphof S. The denition of a
medi-ated syntacticrelationencompasses thenotionofa
direct linkage (n := 2, sothat an empty set of
in-terveningnodesemerges). Thespecialcasesl := 1
andl := nyieldanascendinganddescendingseries
ofhead-modierrelations,respectively.
4 Semantic Interpretation Model
The model of semantic interpretation we propose
comprises two constraint layers. First, static
con-straintsforsemanticinterpretationderivedfrom
di-rectly mappingdependencyrelations to conceptual
roles, and, second, asearch of the knowledge base
whichdynamically takesthesestaticconstraintsinto
account. The translationfrom the syntacticto the
semanticlevelisachievedinastrictlycompositional
resentationsofsemanticallyinterpretablesubgraphs
untiltheentiredependencygraphisprocessed.
Static Constraints. Interpretation procedures
operating on semantically interpretable subgraphs
mayinheritrestrictionsfromthetypeofdependency
relations or from the lexical material they
incor-porate. Constraint knowledge from the grammar
levelcomes in twovarieties, viz. via apositive list,
D lexval
+
,andanegativelist,D lexval
,ofdependency
relations, from which admitted as well asexcluded
conceptualrelations, R
+
and R , respectively, are
derivedbyasimplestaticsymbolmapping.
KnowledgeaboutD lexval
+
andD lexval
ispartof
thevalencyspecications. Itisencodedatthelevel
oflexeme classesW,suchthatlexval2WD. By
wayofpropertyinheritancethisknowledgeispassed
on to all subsumed lexical classes and instances.
For instance (cf. Figure 1), the lexeme class of
in-transitiveverbs,verbintrans 2W,denes forits
subject valency D
hverbintrans; subjecti
+
:= fsubjectg
andD
hverbintrans;subjecti
:=;,whereasfor
preposi-tionaladjunctswerequireD
hverbintrans;ppadji
+
:=;
andD
hverbintrans;ppadji
:= fsubject,dirobject,
in-dirobjectg. All these constraints are inherited by
thelexeme class verbtrans. Wethen distinguish
threebasiccaseshowcorrespondingconstraintsmay
aectsemanticinterpretationprocesses:
1. Knowledge available from syntax determines
thesemanticinterpretation,ifD lexval
2. Knowledge available from syntax restricts the
semantic interpretation, if D lexval
+
= ; and
D lexval
6=;(e.g.,forprepositionaladjuncts).
3. If D
constraints apply and semantic interpretation
proceeds entirely concept-driven, i.e., it relies
ondomainknowledgeonly(e.g.,forgenitives). 3
In order to transfer syntactic constraints to the
conceptuallevel, we dene i: D ! 2 R
, amapping
from dependency relations onto sets of conceptual
relations. Some of these mappings are already
de-pictedin Figure3(e.g., i(subject):= fagent,
pa-tientg). FordependencyrelationsÆ2Dthat
can-notbelinkedaprioritoaconceptualrelation(e.g.,
gen[itive]att[ribute]),werequirei(Æ):=;.
Theconceptualrestrictions,R
+
andR ,mustbe
computed from D lexval
+
and D lexval
, respectively,
byapplyingtheinterpretationfunction itoeach
el-ement of the corresponding sets. This leads us to
R
Wehavecurrentlyno empirical evidencefor the fourth
possiblecase,whereD lexval
6=;andD lexval
interpretationimpliesasearchintheknowledgebase
which takes the constraints into account that
de-rivefrom aparticular dependency parse tree. Two
sortsofknowledgethenhavetobecombined|rst,
a pair of concepts for which a connecting relation
pathhastobedetermined; second,conceptual
con-straintsonpermittedandexcludedconceptual
rela-tionswhenconnectedrelationsarebeingcomputed.
The rst constraint type incorporates the content
wordslinkedbythesemanticallyinterpretable
sub-graph,thelatteraccountsfor theparticular
depen-dencyrelation(s)holdingbetweenthem. Schema(1)
describes the most general mapping from the
con-ceptual correlates,h:C
from
andm:C
to
, in F of the
twosyntacticallylinkedlexical items, hand m,
re-spectively,toconnectedrelationpathsR
con
Aconnectedrelationpathrel
con
Arelationpathiscalled connected,ifforallitsn
constituent, noncomposite relations r
i
the concept
type of the domain of the relation r
i+1
subsumes
theconcepttypeoftherangeoftherelationr
i .
Tocomputeasemanticinterpretation,sitriggers
asearchthroughthe knowledge base and identies
allconnectedrelationpathsfromC
from toC
to . Due
to potential conceptual ambiguities in interpreting
syntactic relations, more than one such path may
exist (hence, we mapto thepowerset of R
con ). In
order to constrain connectivity, si takes into
con-siderationall conceptualrelationsR
+
Rapriori
permittedfor semanticinterpretation,aswellasall
relationsR Raprioriexcluded. Bothofthem
re-ecttheconstraintssetupbyparticulardependency
relationsornon-contentwordsguringaslexical
re-latorsof contentwords. Thus, rel 2 g
R
con
holds, if
relisaconnectedrelationpathfromC
from to C
to ,
obeyingtherestrictionsimposedbyR
+
and R .
If the function si returnsthe empty set (i.e., no
valid interpretation can be computed), no
depen-dency relation will be established. Otherwise, for
all resulting relation paths rel
i
tional axiom of the form (h:C
from
added to the knowledge base, where rel
i
denotes
thei th
reading. Ifi>1,conceptualambiguities
oc-cur,resolution strategiesforwhicharedescribedin
RomackerandHahn(2000a).
TomatchaconceptdenitionC againstthe
con-straintsimposedbyR
+
andR ,wedenethe
func-tionget-roles(C)=: CR ,whereCRdenotestheset
andgeneralityofspecication,R
+
andR consistof
themost generalconceptual relationsonly. Hence,
the concrete conceptual roles CR and the general
onesin R
+
and R maynotalwaysbecompatible.
Sopriorto semanticinterpretation, weexpand R
+
andR intotheirtransitiveclosures,incorporating
alltheirsubrelationsintherelationhierarchy. Thus,
R
correspondinglydened. R
+
restrictsthesearch to
relationscontainedinCR\R
+ ,iR
+
isnotempty
(otherwise,allelementsofCRareallowed),whereas
R allowsonlyforrelationsinCRnR
.
5 A Sample Semantic Interpretation
Whenever asemantically interpretable subgraph is
complete, semantic interpretation gets started
im-mediately. Asanexample,wewillconsideracaseof
indirectlinkage,asillustrated bythe occurrenceof
auxiliaryandmodalverbswithin apassiveclause.
Wheninterpretingindirectsyntacticrelations,
in-formation not only about content word nodes but
also about intervening noncontent word nodes
be-comesavailable. Thisway,furtherstaticconstraints
areimposedonR
+
(andR )intermsofalistR
lex
Rof permitted conceptualrelations. This
infor-mationisalwaysspeciedatthelexemelevel. Since
R
lex
relatestoclosed-classitems only, therequired
numberofspecicationsiseasytosurvey.
In ourexample(cf. Figure 2), thecontentwords
\Festplatte" (hard disk) and \geliefert" (delivered)
arelinkedbyamediatingmodalverb(\kann"(can))
andapassiveauxiliary(\werden"(be
passive
)). The
semantic interpretation schema for passive
auxil-iaries(2)addressestheconcepttypeoftheinstance
for their syntactic subject, C
subj
= type(I
subj ) =
Hard-Disk, and that for their verbpart,C
verbpart
= type(I
verbpart
) = Delivery. The relation
be-tweenthesetwo,however,isdeterminedbyR
passaux
:= fpatient, co-patientg, constraint knowledge
which residesin the lexeme specication for
\wer-den"aspassiveauxiliary(cf.Figure1).
si
(Delivery, fpatient, co-patientg,
;, Hard-Disk), we get the conceptual relation
deliver-patient(cf. Figure 3), since Hard-Disk
issubsumedbyProductand,thus,alegalllerof
deliver-patient2R
passaux .
6 Conceptual Interpretation
Conceptualinterpretationusesaproductionrule
sys-tem (Yen et al., 1991) which accounts for
charac-teristic patterns of assertions that result from the
tualinterpretationabstractsawayfromthesesurface
phenomena and creates a `normalized', canonical
conceptual representation of the input, as needed,
e.g.,foruniformly queryingtheknowledgebase.
AsanexampleofsuchinferencesconsiderFigure
5,withthedeliversrelationlinkingTranstec.9,
a hardware supplier, and Hard-Disk.2. By
com-puting a conceptual relation representing the
un-derlying Action Transtec.9 and Hard-Disk.2
areintegratedin anormalizedconceptgraph. Note
thatthecorrespondinglexicalitems,\Transtec"and
\Festplatte" (hard disk), are not linked via a
me-diated syntactic relation in Figure 2. Hence, we
may clearly discern semantic interpretation, which
operates on single semantically interpretable
sub-graphsonly, from conceptual interpretation, where
the inference-based interpretation of relationships
amongdierentsubgraphscomesintoplay.
An independent level for conceptual
interpreta-tionalsobecameanecessityduetoanalytic
consid-erations. Oftenthelocal constraintsfor conceptual
rolesofAction,State,orEventconceptscannot
be formulated restrictive enough for the semantic
interpretationprocess. For example,theconceptual
correlateoftheverb\possess"doesnotimposeany
restrictiononitspatientrole(linkedtothesubject
dependencyrelationin asemantically interpretable
subgraph). Rather, restrictions apply to properly
relating the ller of the patient slot with that of
the co-patient slot (dirobject at the dependency
level). Conceptual interpretationrulesare ameans
tofurtherconstrainthese`context-sensitive'aspects
oftheinterpretationprocess.
Since verbs play aprominentrole in dependency
grammars, the production rule system for
concep-tualinterpretationisbasedupontheconceptual
cor-relates of verbs (henceforth verb concepts) in the
knowledgebase. Dierentviewsaredenedforverb
conceptsbyusing threeabstractiondimensions.
First, verb concepts are classied, according to
the set of thematic roles they supply, as Action,
State orProcess. Delivery, e.g.,is assignedto
Action,sincebothagent andpatient formpart
oftheconceptdenition(cf.Figure3,rightbox).
Thesecondlevelofabstractionconsistsof
catego-rizationswhichreectacommoncoremeaning. The
upmost conceptualnode in thishierarchyis
Cate-gory. Delivery, e.g., is considered as a concept
whichrepresentstheActionoftransferingaGood
Figure5:ASampleConceptualInterpretationofthe
category are subsumed by the corresponding
con-cept CAT-Transfer-Good. (We here make use
ofmultipleinheritancemechanisms.)
Finally, every verb concept is linked to some
Verb-Model.Deliveryoranyotherverbconcept
of the CAT-Transfer-Good category is a
con-stituent phase of the Buy-and-Sell-Model. To
generalizeappropriatelyfromindividualverbs,verb
categorieswereextractedfromourtextcorporathat
further rene a large-scale taxonomy for German
verbs (Ballmer and Brennenstuhl, 1986). In this
work,atotalof about20,000verbswere subsumed
by700categoriestoreectasemanticgeneralization
intermsofahierarchyofverbcategories.
Theproductionrulesforconceptualinterpretation
operateonthiscategorialhierarchy. Everyverb
con-ceptinthehierarchyisasubconceptofexactlyone
categoryintheknowledgebase. Wheneverthe
pre-conditions of an interpretation rule are fullled, a
conceptualinterpretationiscomputed.
Conceptual and semantic interpretation depend
oneachother,sincethebasicinterpretationschema
(cf.expression(1)inSection 4)issuppliedwith
ac-tual parameters from production rules. We
there-foremay dene another specialization of the basic
interpretationschema forconceptual interpretation
si
conc
. In particular, path searches are triggered
thatarerestrictedbyapositivelistrenderedbythe
applicableproductionrule.
Foroursamplesentence(cf.Figures2and4),the
conceptual correlate for the verb \delivers"
(De-livery) is a subconcept of Action.
Addition-ally, Delivery is a subconcept of the category
CAT-Transfer-Good (cf. Figure 3). The
corre-sponding conceptual interpretation rule is given in
Figure 6. Whenever an instance of the category
CAT-Transfer-Goodisencounteredandbothits
agent and patient roles arelled, relation paths
are computed from the types of the two instances
involved, a and p, respectively. For each relation
found by the search algorithm (R EL in Figure 6),
acorrespondingassertionisaddedtotheknowledge
base (TELLin Figure 6). In the example, the
in-terpretationschemaisinstantiatedwiththe4-tuple
(Company, ftransfers-goodg, fg, Hard-Disk)
resultingin the computation of fdeliversg asthe
properrelationlink (cf.Figure 5),sinceitis a
sub-relationoftransfers-good.
EXISTSv;a;p:
v : CAT-Transfer-Goodu
vagentau vpatientp=)
IF siconc (type(a), ftransfers-goodg, fg, type(p)) 6= ;
THEN
REL:=si
conc
(type(a),ftransfers-goodg,fg,type(p))
Weevaluatedthis approachto semantic
interpreta-tion ona random selectionof 54 texts (comprising
18,500words)from twotextcorpora,viz.consumer
product test reports and medical nding reports.
Forevaluationpurposes,weconcentratedonthe
in-terpretation of genitives (as an instance of direct
linkage) and on the interpretation of periphrastic
verbalcomplexes, i.e.,passive,temporalandmodal
constructions(asinstancesofindirectlinkage).
The underlying ontology consists of an upper
generic part (containing about 1,500 concepts and
relations)anddomain-specicextensionsrelatingto
informationtechnology(IT)and(partsof)
anatomi-calmedicine(MED).Eachofthesetwodomain
mod-els adds about 1,400 concepts and relations to the
upper model. Corresponding lexeme entries in the
lexiconprovidelinkagesto theentireontology.
Weconsideredatotalof247genitivesinthe
sam-ple. Recallwas higherformedicaltexts(57%)than
forIT documents(31%),though, ingeneral, rather
low. However,precisionpeakedat97%and94%for
medical and IT texts, respectively. Thenumberof
syntactic constructions with modal verbs or
auxil-iaries amout to 292 examples. Compared to
geni-tives, we obtained a slightly more favorable recall
forbothdomains |66%forMED, 40%for IT|,
whileprecisiondroppedslightlyto95%and85%for
medicalandITdocuments,respectively. 4
Aswithanysuchevaluation,idiosyncrasiesofthe
coverage oftheknowledgebases areinevitably tied
with the results and, thus, put limits on too
far-reaching generalizations. However,our data reect
the intention to submit a knowledge-intensivetext
understander to a realistic, i.e., conceptually
un-constrainedandtherefore\unfriendly"test
environ-ment. Judged from the gures of our recall data,
thereisnodoubt, whatsoever,thatconceptual
cov-erage of the domain constitutes the bottleneck for
any knowledge-based approach to NLP. 5
Sublan-guagedierencesarealsomirroredsystematicallyin
these data, sincemedicaltexts adhere moreclosely
to well-established concepttaxonomiesand writing
standardsthanmagazinearticlesintheITdomain,
whoserhetoricalstylesvary toalargerdegree.
8 Related Work
The standard way of deriving a semantic
interpre-tationforconstituency-basedgrammarsistoassign
eachsyntacticruleoneormoresemantic
interpreta-tionrules(e.g.,vanEijckandMoore(1992)),andto
4
Amoredetailedpresentationofthisevaluationstudyis
giveninRomackerandHahn(2000b).
5
Forthemedicaldomainatleast,wearecurrentlyactively
pursuingresearchonthesemiautomaticcreationoflarge-scale
ontologiesfromweakknowledgesources, viz.medical
termi-constituents. This approach hasalso been adopted
inthefewexplicitattemptsatincorporating
seman-ticinterpretationintoadependencygrammar
frame-work(Milward,1992;Lombardoetal.,1998). There
arenoconstraintsonhowtodesignandorganizethis
rulesetdespitethosethatareimpliedbythechoice
of the semantic theory. In particular, abstraction
mechanisms (going beyond the level of sortal
tax-onomies for semantic labels, cf., e.g., Creary and
Pollard (1985)), such as property inheritance,
de-faults,arelacking. Accordingly,thenumberofrules
increases rapidly and easily reaches orders of
sev-eral hundreds in a real-world setting (Bean et al.,
1998). As an alternative, we provide a small set
ofgenericsemanticinterpretationschemata(bythe
order of10)andconceptualinterpretationrules(by
the order of 30 for 200 verb concepts) instead of
assigning specicinterpretationrulesto each
gram-mar item (in our case, single lexemes), and
incor-porate inheritance-based abstraction in the use of
these schemataduring theinterpretationprocess in
the knowledge base. We clearly want to pointout
that while this rule system covers a wide variety
of standard syntactic constructions (such as
geni-tives,prepositionalphrases,varioustenseandmodal
forms), itcurrentlydoesnotaccountfor
quantica-tional issues (like scope ambiguities) for which
en-tirelylogic-basedapproach(CharniakandGoldman,
1988; Moore, 1989; PereiraandPollack,1991)
pro-vide quitesophisticatedsolutions.
Sondheimer et al. (1984) and Hirst (1988)treat
semantic interpretation as a direct mapping from
syntactic to conceptual representations. They also
share with us the representation of domain
knowl-edgeusing Kl-One-style terminologicallanguages,
and,hence,theymakeheavyuseofproperty
inher-itance(ortyping)mechanisms. Themaindierence
to our approach lies in the status of the semantic
rules. Sondheimer et al. (1984) attach single
in-terpretation rules to each role (ller) and, hence,
have to provide utterly detailed specications
re-ectingtheidiosyncrasiesofeachsemantically
rele-vant(role)attachment. Propertyinheritancecomes
onlyinto play when theselection of alternative
se-mantic rules is constrained to the one(s) inherited
fromthemostspeciccaseframe. Inasimilarway,
Hirst (1988) uses strong typing at the conceptual
object level only, while we use it simultaneously at
thegrammarandthedomainknowledgelevelforthe
processingofsemanticschemata.
9 Conclusions
We introduced an approach to the design of
com-pact, yet highly expressivesemantic interpretation
anisms, are based on inheritance principles.
Sec-ond,interpretationschemataabstractfrom
particu-larlinguisticphenomena(specic lexicalitems,
lex-emeclassesordependencyrelations)intermsof
gen-eralcongurationpatternsindependencygraphs.
Underlying these design decisionsis a strict
sep-aration of linguistic and conceptual knowledge. A
clearly dened interface is provided which allows
thesespecicationstomakereferencetone-grained
hierarchical knowledge, no matter whether it is of
grammaticalor conceptual origin. The interfaceis
divided into two levels. One makes use of static,
high-level constraints supplied by the mapping of
syntactic to conceptual roles or supplied as the
meaningofclosedwordclasses. Theotherusesthese
constraints in a dynamic search through a
knowl-edge base, that is parametrized by few and simple
schemata. Finally, at the level of conceptual
in-terpretationinferencesemergingfromsemantic
rep-resentations are computed by a set of productions
whichmakereferencetoaverbcategorialhierarchy.
Alsosincethenumberofschemataatthesemantic
descriptionlayerremainsrather small, their
execu-tion is easy to trace and thus supports the
main-tenance of large-scale NLP systems. The high
ab-stractionlevelprovidedbyinheritance-based
seman-tic specications allows easy porting across
dier-ent application domains. Our experience rests on
reusingthesetofsemanticschemataoncedeveloped
fortheinformationtechnologydomain inthe
medi-caldomainwithoutfurtherchanges.
Acknowledgments. Wewanttothankthemembers
ofthe groupforclosecooperation. M.Romackerwas
supportedbyagrantfromDFG(Ha2097/5-1).
References
J. Allen. 1993. Natural language, knowledge
rep-resentation, and logical form. In M. Bates and
R.Weischedel,editors,ChallengesinNatural
Lan-guageProcessing,pages146{175.Cambridge
Uni-versityPress.
T. Ballmer and W. Brennenstuhl. 1986. Deutsche
Verben.Einesprachanalytische Untersuchungdes
deutschen Verbwortschatzes. Tubingen: G.Narr.
C. Bean, T. Rindesch, and C. Sneiderman. 1998.
Automatic semantic interpretation of anatomic
spatialrelationshipsinclinicaltext. InProc.1998
AMIA AnnualFallSymposium,pages897{901.
E.Charniak andR.Goldman. 1988. Alogicfor
se-manticinterpretation. InProc.ofthe26thAnnual
Meetingof the ACL,pages87{94.
L. Creary and C. Pollard. 1985. A computational
semantics for natural language. In Proc. of the
23rdAnnualMeetingofthe ACL,pages172{179.
M.Dalrymple. 1992. CategorialsemanticsforLFG.
In COLING'92 { Proceedings of the 15th [sic!
current,object-orientednaturallanguageparsing:
the ParseTalkmodel. International Journal of
Human-Computer Studies,41(1/2):179{222.
E.Hajicova. 1987. Linguisticmeaningasrelatedto
syntaxandtosemanticinterpretation. InM.
Na-gao, editor, Language and Articial Intelligence,
pages327{351.North-Holland.
G.Hirst. 1988. Semanticinterpretation and
ambi-guity. ArticialIntelligence,34(2):131{177.
V. Lombardo, L. Lesmo, L. Ferraris, and C.
Sei-denari. 1998. Incrementalinterpretationand
lex-icalized grammar. In CogSci'98 { Proceedings of
the 20th Annual Conference,pages621{626.
D.Milward. 1992. Dynamics,dependencygrammar
and incremental interpretation. InCOLING'92{
Proceedings of the 15th [sic! 14th] International
Conference, pages1095{1099.
R. Moore. 1989. Unication-based semantic
in-terpretation. In Proceedings of the 27th Annual
Meetingof theACL,pages33{41.
F.PereiraandM.Pollack. 1991. Incremental
inter-pretation. ArticialIntelligence, 50(1):37{82.
M.RomackerandU.Hahn. 2000a. Copingwith
dif-ferenttypesofambiguityusingauniformcontext
handling mechanism. In Applications of Natural
Language toInformation Systems.Proceedings of
the 5th NLDB Conference.
M. Romacker and U. Hahn. 2000b. An empirical
assessmentofsemanticinterpretation. InProc. of
the6thAppliedNaturalLanguageProcessing
Con-ference &1st Conference of the North American
Chapter of the ACL,pages327{334.
S. SchulzandU. Hahn. 2000. Knowledge
engineer-ing by large-scale knowledge reuse: experience
fromthemedicaldomain. InKR'2000{Proc.7th
International Conference,pages601{610.
N. Sondheimer, R. Weischedel, and R. Bobrow.
1984. Semanticinterpretationusing Kl-One. In
COLING'84{Proc.10thInl.Conference&22nd
Annual Meetingof the ACL,pages101{107.
M.StrubeandU.Hahn. 1999. Functionalcentering:
grounding referential coherence in information
structure. Computational Linguistics, 25(3):309{
344.
J. van Eijck and R. Moore. 1992. Semantic rules
forEnglish. InH.Alshawi,editor,TheCore
Lan-guage Engine, pages83{115.MITPress.
J.WedekindandR.Kaplan. 1993. [Type-driven
se-mantic interpretation of f-structures << J;W >
;<R ;K>>]. InEACL'93{Proc. 6th Conf.
Eu-ropeanChapter of theACL,pages404{411.
W. Woods and J. Schmolze. 1992. The Kl-One
family. Computers & Mathematics with
Applica-tions,23(2/5):133{177.
J. Yen, R. Neches, and R. MacGregor. 1991.
CLASP: integrating term subsumption systems
and production systems. IEEE Transactions on
