An Integrated Model of Semantic and Conceptual Interpretation 
fl'Oln Dependency Structures 
Udo Hahn Martin Romacker 
I~,: .... Groul), Text Understanding Lab, 
Fre, iburg University, D-79085 Freilmrg, Germany 
http ://www. coling, uni-freiburg, de/ 
Abstract 
\Ve propose a two-layered model for computing se- 
mantic and (:onceI)tual interpretations from del)en- 
dency struetm'es. Abstract interl)retatio,t s(:hemata 
generate semantic interpretations of 'minimal' de- 
pendency sul)gral)hs , while production rules whose 
sl)eeitieation is rooted ill ontologi(:al categories de- 
rive a canonical con(:eptual il~terl)retation from se- 
lna.ntic int;ert)retal;ion sl;ruel;ures. Configm'ational 
descriptions of del)endeney gral)hs increase the lin- 
guistic generality of interl)rel,ation s(:hemata, while 
interfimillg sehemata and t)ro(htcl;ions t() lexi(:al and 
COll(:el)l:tl~ll (;lass hierarchies re(ht(:es the amount and 
complexity of semantic sl)e(:iti(:atJons. 
1 Introduction 
The syntax/semanti(:s interface has always t)een a 
matter of com:ern for (:OllStit;llell(:y-l)ased feal;llre 
grammar theories (of., e.g., Creary an(1 Pollard 
(\]9S5), ~.'~ooIx, (i\[989), \])ah'yllll)le (1.992), Wedekind 
and Kaplan (1993)). Within the dependen(:y gram- 
lllal' COllllllllllil;y: f3r loss at, te, ntion has 1)een l)ai(t t() 
l;his tel)it:. As a (:onse(tuen(:e, ther(~ is no consensus 
how syntactic del)en(lency structures might l)e a(t- 
e(tuately transl'()rm(~d into semanti(: interl)rel;aiions 
(el., naji~:ova (:t987), Milward (1992), Lombardo et 
al. (1998) for alt;ernative proposals). 
In this paper, we introduce, a two-layered inter- 
pretation model. In a first; pass, dependency graph 
structures which result fl'om in(:remental parsing are 
immediately submitted to a semantic intcrl)reta- 
tion process. Such a process is triggered by gen- 
eral schemata whenever a semantically interl)retable 
subgraph of a syntactic dependency gral)h t)ecomes 
ava.ilable (el. Section 3). As a result, lexical items 
and the del)endency relations holding 1)etwe.en them 
are directly mat)ped to associated concel)tual enti- 
ties mid relations at; the level of semantic represen- 
tation (cf. Sections 4 mtd 5). In a subsequent steI), 
the (quasi-inferential) iml)lications of the knowledge 
representation structures emerging from the seman- 
tic interpretation stet) are accounted for l)y a pro- 
eess we here refer to as concepl,'aal interpretation. 
The (:orresl)onding ot)erations relate to the (:(mcel)- 
tua\] representation level only and are triggered by 
a variety of production rules rooted in ontological 
categories in order to generate a canonical concep- 
tual representation of the parsed sentence (el. Sec- 
tion 6). This second level of interpretation is usually 
not taken into consideration by computational mod- 
els of semantic interpretation, neither constituency- 
based nor (lei)endency-based ones, although it turns 
out to crucial for natural  ',,ndeTwtandin.q. 
2 Grammar and Concept Knowledge 
Grammatical knowledge for syntactic analysis is 
t)ased on a fully texicalized dependency gl'alllltla\] 
(Itahn et al., 1994). ()ur preference for dependency 
structures is motivated, among other things, t)y the 
observation that the corresl)ondence of det)endency 
relations (holding between lexical items) to con(:ep- 
tual relations (holding between the, concepts they 
denote) is much closer than t.'or constituency-based 
grammars (Ilajicova, 1987). Ilence, a dependency- 
based al)i)roach cases inherently the descrit)tion of 
the regularities mMerlying selnantic inl;erl)retation. 
In this lexicalized del)endeney framework, lexeme 
sl)eeitications form the leaf nodes of a lexicon I)AG, 
which are further al)sl:racted in terms of lexeme 
class specilical;ions at different levels of generalit:y 
(of. Figure 1), This lea& to a lexeme (:lass hier- 
archy, which consists of lexeme (:lass names 142 := 
{VE1HIAL, VEllBINTII.ANS, NOMINAL~ NOUN~ ...} alld 
a subsuinption relation i.saw = {(VEI{IIINTI{ANS, 
gexenle 
Verbal ... Preposilion ... Nominal 
Verbhltrans Auxiliary Pronoun Noun Df'J. '= 
1,,.101 , Dg'""'":= \[ \] 
D"'I~J:= { \] i D- v''n'm .= \[ \] 
I I)PP'O:= \[sul o, dirobj, imlirobj} 
VelbTrallS u,erde*t passive Sl~eicherOncmory) 
::lglir°bj:- \[di,'ol,j} R+:= \[patten! CO-l,ali,'nt } 
:. qdm,l,j:= \[ } mit (with) 
)i~:lt!rll (deliver) R+ := \[has-part instrmm,nt co-patient ..} 
Figure l: Fragnmnt of the Lexeme Class Hierarchy 
271 
propo: ____~ 
ka~n sublCCt: ~-- ~_~ovorbpart 
Fesl latte Sp~tt: werden 
v~orDpa rt. 
Die Compellers pps~ferl spec:~. ~ ~..~.att 
des mit yon pobjo~,:'?"~ "-~o~jo~t 
350Mhz-OPU Transtec 
7he hard disk of flw ¢vmlputer with 3.$OMhz-CPU call - b)' 7ra/t.~lrt:- be delivrred. 
Figure 2: A Sample Dependen(:y Grat)h 
VERBAL), (NOUN, NOMINAL), ...} C "l/V X \]/V. Inher- 
itance of grammar knowledge is based on the idea 
that constraints are attached to the most general 
lexeme classes to which they apply, leaving room for 
more and more specific (possibly, even idiosyncratic) 
specifications when one descends this hierarchy. 
A dependency grammar captures binary con- 
straints between a syntactic head (e.g., a noun) and 
one of its possible modifiers (e.g., a determiner or 
an adjective). In order to establish a dependency 
relation ~ C :D := {specifier, s~fl~iect, dirobject, ...} 
between a head and a Inodifier, lexenle-class-specific 
constraints oil word order, comt)atit)ility of Inor- 
phosyntactic features and senmntic integrity must 
be fiflfilled. Figure 2 depicts a dependency graph in 
which word nodes are given in bold face and depen- 
dency relations are indicated I)y labeled edges. 
Conceptual knowledge of the underlying domain is 
expressed in terms of a KL-ONE-like knowledge rep- 
resentation  (Woods and Schmolze, 1992). 
The domain ontology consists of a set of concei)t 
names 9 r := {COMPANY, tIAI~I)-.DI.g\[<, ...} and a sub- 
sunlI)tion relation isaf = {(HAIiD-DISK, NTOHACF- 
DFVICF), (TRANS'rI.:(~', COMPANY), ...} C ~P x 
Y. The set of relation nantes ?g :-- {IIAS-t'All;F~ 
I)EI, IVEI~-AGENT, ...} denotes con(:eptual relations 
which are also organized in a subsmnption hierarchy 
isa?z = {(ttAS-IIARI)-DISK, ItAS-PItYSICAL-PAVI'), 
(IlAS-PHYSICAL-PAR.T, ItA.q-PAltSI'), ...}.1 Examples 
of emerging concept mid relation hierarchies are de- 
picted in Figure 3 (right box). 
Ill our at)proach, the representation s for 
semantics and domain knowledge coincide (for argu- 
ments supporting this view, el. Allen (1993)). Link- 
ing lexical items and concet)tual entities proceeds 
as follows: Upon entering the parsing process, each 
lexical itern w that has a (:onceptual correlate C in 
the domain knowledge base, w.C E Y (mostly verbs, 
Ilouns and adjectives), gets immediately instantiated 
in the knowledge base, such that for any instance Iw, 
initially, 2 type(I,,) = w.C holds (e.g., w = :'Fest- 
platte", I,,, = HAI{D-Dlst,:.2, w.C = type(HaI~.D- 
DISK.2) = HARD-DISK). If several conceptual cor- 
relates exist, either due to honmnymy or polysemy, 
I Atl sllbstllnpl;ioIl ,elations, isaw, isa:v, and isw~, arc 
considered to be transitive and reflexive. 2For 
instance, anaphora might necessitate changes of this 
initial reference assignment, el. Strube and Ilahn (1999). 
i Syntactic Level -> Conceptual Level 
I 10op-'o,~): !subl\[ecl\]'~-------------t-~aecnt Action C,'dcgory 
I)l.'.~',d?lfe.c't\]~paticnt~,s.a..< is-at: i~-~¢: dirobjfect\] ~ 
'\\ \[CAT-Vrans fct'-no,)d 
: I1" oo-pa,i . 
has-part \ ~k ..... :C' -- Persal gen\[itive\]att\[ribute\] I " \\l, divery \l delivcr.aealt 
vcrbpart II instrumcnt "},,a;,,2J',,:ti2 '~"""~' 
ppatt\[ributc\] II destination ~ ¢--! ..... P ...... -- I deliver-recipient 
spcc\[ificr\] /I l, : 
.... ~ L 
Figure 3: Relating Grammatical (left box) and Con- 
ceptual I(nowledge (right box) 
each lexical ambiguity is processed independent.ly 
within set)arate (:ontext partitions of the km)wledge 
base (Romacker and Hahn, 2000a). 
3 Interpretable Subgraphs 
II1 the parse tree from Figure 2, we can distinguish 
lexical nodes that have a conceptual correlate (e.g., 
"Fcstplatte" relating to HAl{D-DIsK, "gelicfert " re- 
lating to DEIAVh;RY) from others that do not have 
Such a correlate (e.g., "mit" (with), "yon" (by)). Se- 
mantic interpretation capitalizes oil this distinction 
in order to tind adequate conceptual relations be- 
tween the corresl)onding concept, insta.nces: 
Direct Linkage. If two word no(tes with (:oncet)- 
tual correlates are linked by a single depen(ten(:y re- 
lation, a direct linkage is given. Such a subgraph 
can immediately be interpreted in ternis of a (:on- 
ceptual relation licensed by the correspondiitg de- 
l)endency relation. This is illustrated in Figure 2 by 
the direct linkage between "Festplatte" (hard disk) 
and "Computers" via the gen\[itive\]att\[ribut(~\] rela- 
tion, which gets nmpped to tile IIARD-DISK-OF role 
linking the corresponding con(:eptual correlates, viz. 
HARD-DISK.2 alld C()MPUTER-,~YSTFM.4, respec- 
tively (see Figure 4). This interpretation uses only 
knowledge about the concet)tual correlates and the 
linking dependency relation. 
Indirect Linkage. If two word nodes with con- 
ceptual correlates are linked via a series of depen- 
Delivery.10 
...... c~. ..... Transtec.9 
de|lver -agent !- 
.... ---41~--- Hm'd-Disk.2 ! 
deliver -pattellt \[ --(~}-- COlllpllie\[ Syst('lll.4 
hard-disk-of \[ (Ji 350Mhz 
CPU.6 hai-qm 
Figure 4: Semantic Interpretation of the Depen- 
dency Graph fl'om Figure '2 
272 
dency relations and none of tile illtervening no(les 
have a conceptual correlate, an indirect linkage is 
given. For such a "nfinimal" sut)grat)h, semantic in- 
tert)retal;ion is made dependent on lexical informa- 
tion from the, intervening nodes, as well as knowledge 
aboul; the conceptual correlates and del)cn(lency i'e- 
lations. Figure 2 illustrates such a eontiguration 
lay tile linkage between "Com,putcr.s" and "350Mhz- 
CPU" via the intervening node "mit" (will O and 
the pI)att\[ributc\] mid l>objc, ct relations, the re, suit <)f 
which is a conceptual linkage between COMI'UTFI/.- 
SYSTI,'M.4 and 350N/IIIZ-Cl'U.6 via the relation HAS- 
CI'U in Figure 4. 
In oMer to increase tile generality and to t)reserve 
the simi)licity of semantic interpretation we intro- 
duce a generalization of the notion of del)emlency 
relation such that it incorporates direct as well as 
indirect linkage: Two content words (nouns, adjec- 
tives, adverbs or flfll verl)s) stand in a 'me.dialed syn- 
tactic relation, if one can t)ass from one word to 
the other along the connecting edges of the (tcl)en- 
(tency gral)h wit;hour traversing word nodes other 
than t)repositions, modal or auxiliary verl)s (i.e., el- 
ements of eh)sed woM classes). In Figure 2, e.g., 
the tuples ( "Fcstplattc.", "Co'm,p',,ter.s ") or ( "6'om,- 
putcrs", "350Mh, z-CPU") stand in mediated syntac- 
tit relations, whereas, e.g,, the tuple ("Fest, plattc", 
"Tra'nstcc") does not, since the comiecting 1)ath COlt- 
rains "gelicfc.rt" (dclivcrcd), a (;Ollt;elll; word. 
This leads to the following detinition: Let w and 
'w' be l, wo (:onl;enI; words in a senten(;(; ,5'. 1111 addi- 
l;i(-)n, lel; 'w2,... ,'wn-1 E S ('11, ~ 2) l)e l/rel)osil;ions , 
auxiliary or modal verl)s, and wl := 'w an(1 w, := 
'u/. Then we say thaJ; 'w and 'w' st:and in a 'nt(:diat('d 
syntactic r(:lat, ion, iff there exists an inde.x I C {1, 
..., n} so that the following two (:on(litions hold: 
1. 'wi is modifier of'wi+l for i C {1, ..., l-1}; 
2. wi is head of wi+\] for i E {I,..., n-l}. 
We call a subgral)h identitie(1 t)y su(:h a s(!rics 'w j, 
... , w, a, sc'm, anl, ically intcrFrc, tabh', .s"u, bgraph, of 1;11(,' 
dependency graph of S. The, detinMon of a medi- 
ated syntactic relation encoml)asses the notion of a 
direct linkage (n := 2, so t;hat an empty set of in- 
tervening nodes emerges). The special eases 1 := 1 
and I := n yield an ascending and descending series 
of head-modifier relations, resl)ectively. 
4 Semantic Interpretation Model 
The model of semantic intert)re.tatiol~ we l)rOl)OSe 
(:Oral)rises two (:onsl,raint layers. First, sl, atic (:on- 
straints fl)r semmltic intert)retation derived from di- 
rectly mapping dependency relations to conceptual 
roles, and, second, a search of the knowledge t)ase 
which dynamically takes these static constraints into 
account. The translation from the syntacti(" to the 
semantic level is achieved in a strictly COml)ositional 
way l)y incrementally c()mbining the conc(;ptual l"(;I)- 
resentations of semantically interl)retat)le sul)gral)hs 
until the entire del)endeney graph ix processed. 
Static Constraints. Intert)retation 1)rocedures 
operating on semantically interi)retable, subgrai)hs 
may inherit restrictions from the tyl)e of dependency 
relations or from the lexical material they incor- 
1)orate. Constraint knowledge from the g;ranmmr 
level comes in two varieties, viz. via a positive list,, 
Icx v.l D4 .... , and a negative list,, D\] (''~'~'°l, of det)endcncy 
relations, tronl which adlnitted as well as excluded 
concel)tual relations, //,~ and /~,_, resl)ectively, are 
derived 1)y a silnple static symbol mal)l)ing. 
Knowledge at)out D_~ .......... I and Dj c*val is part of 
the valen(:y sl)e(;itications. It is encoded at the level 
of lexeme classes l/V, such that lcxval C )IV x "D. By 
way of 1)rol)erty inheritance this knowledge is passed 
on to :ill subsumed lexical classes and instances. 
For insta.nce (of. Figure l), the lexeme class of in- 
transitive verbs, VEIIBINTI/ANS G ~/~), d()\[iiles for igs 
sul).icct valency D ~! v''W"rr' .......... kiccO := {sul).icct} 
~lll(l \]) {r, crbinlrans, sul, jccl.) _ ' := (/), whereas for 1)ret)osi- 
• I) (vcrbinlrans, plmdj) tiolml a(liuncts we i'equire ~.~_ := 0 
anti D (,,~,.I,h,,. ....... m,a4i) := {sul)ject, dirob.ject, in- 
dirol)ject}. All these constl'ainl;s arc inherited t)y 
the lexenm class VIgtlBTI/AN,S. \¥e thell distinguish 
tin'co t)asic cases how (:orrespmMing constraints may 
alib, ct semantic interl)rctation 1)roc(!sses: 
I. Knowledge availalfle flom ,qy'lfl;ax deter'minas 
tim st;mind;it interl)retation , if Dq{ c~:v"! ~- 0 and 
D\] ......... t = ~ (e.g., the subject of a verb). 
2. Knowledge availal)le from syntax vt',stricLs' t.he 
semanl;i(-intcrpr(;l.a.tion~ if' Dq{ (':''''''t = ~ mM 
1)\] .......... t ¢_ 0 (e.g., for 1)reposit.ional adjuncts). 
3. if' Dq{ ....... I = 0 all(1 D / ........ 1 = (/), lie syntacti(: 
(:(mStl'aints ap1)ly and semanl.ic interl)re, l;ation 
pr()(-eeds e, ntire, ly concept-driven; i.e., it relies 
on domain knowledge only (e.g., for genitives).a 
hl order to transfer syntacl;ic constraints to l;}le 
eoncel)tual level, we define i: 7) --+ 2 "n, a mal)t)ing 
fi'onl det)endency relations onto sets of concel)tual 
relations. Some of these mal)l)ings are already de- 
I)icted in Figure 3 (e.g., i(subjecO := (aC~NT, PA- 
TIENT}). For del)en(lency relations 5 E "D that can- 
not })e linke, d ;I priori to a concel)tual relation (e.g., 
g(,n\[itiv@,tt\[ributeJ), we require i(5) := 0. 
The (:oncel)tual restrictions, R+ and/~,_, must be 
(:Oinlmt.ezl from Dq{ <':vat and D_ h,xvaz, respectively, 
1)y al)l)lying the interi)retation flmction i to each el- 
ement of the corresponding sets. This leads us to 
1~+ := {, I :': e V,! ........... z/~ y C i(:,,)} ~uld \]~ := 
{V I :" e n2'*'" a , e i(:,')}. 
a\Ve hawe currently ilo eml)irical evidence for the fourth 
possible (:as('., where 1) lca'wd ~k 0 and 1) lcxval ~ O. 
273 
Dynamic Constraint Processing. Semantic 
interpretation implies a search in the knowledge base 
which takes the constraints into account that de- 
rive fl'om a particular dependency parse tree. Two 
sorts of knowledge then lmve to be combined -- first, 
a pair of concepts for which a connecting relation 
path has to be determined; second, conceptual COl> 
straints on permitted and excluded conceptual rela- 
tions when connected relations are being computed. 
The first constraint type ineorporates the content 
words linked by the semantically interpretable sub- 
graph, the latter accounts for the pm'ticular depen- 
dency relation(s) holding between them. Schema (1) 
describes the most general mapping fl'om the coil- 
ceptual correlates, h.Cfrom and 11t.Cto~ ill ~ of tile 
two syntactically linked lexical items, h and m, re- 
spectively, to connected relation paths 12~o,,.. 
)c x 2 ~ x 2 v¢ x .7 -~ 2/~ .... 
";: (CI,.o,,~ , 12+, 12-, C,o) ~ 12~o~,, (1) 
A connected relation path rel~o,, C R .... is defined 
by: 
relcon((rt,..., rn)) :¢:;' Vi C {1, ..., n - 1} : 
A relation path is called connected, if for all its n 
constituent, noncomposite relations ri the concept 
type of the domain of the relation ri+l subsumes 
the concept type of the range of the relation ri. 
To compute a semantic interpretatiol~, si triggers 
a search through the knowledge base and identifies 
all commeted relation paths from Cf~.o,, to Cto. Due 
to potential conceptual ambiguities in interpreting 
syntactic relations, more than one such path may 
exist (hence, we mat) to the power set of 12~o,~). In 
order to constrain connectivity, si takes into con- 
sideration all conceptual relations 12+ C Td a priori 
permitted for semantic interpretation, as well as all 
relations R_ C 7~ a priori excluded. Both of them re- 
flect the constraints set 11 t) by particular (lel)endency 
relations or non-content words figuring as lexical re- 
lators of content words. Thus, tel G R~o,, holds, if 
tel is a connected relation path from Cf,.o,,~ to Cto, 
obeying the restrictions imposed by 12+ and 12_. 
If the fllnction si returns the empty set (i.e., no 
valid interpretation can be comtmted), no depen- 
dency relation will be estat)lished. Otherwise, tbr 
all resulting relation paths RELi E I~con an asser- 
tional axiom of tile form (h.Cfro,n I~.ELi m.Cto) is 
added to the knowledge base, where RELi denotes 
tile i th reading. If i > 1, conceptual ambiguities oc- 
cur, resolution strategies for which m'e described ill 
Romaclmr and Hahn (2000a). 
To match a concept definition C against tile con- 
straints imposed by 12+ and 12_, we define the func- 
tion get-roles(C) =: CR, where C12 denotes the set 
of conceptual roles associated with C, which are then 
used as starting points for the path search. For ease 
and generality of st)ecification , R+ and li~_ consist of 
the most general conceptual relations only. Hence, 
the concrete conceptual roles CR and the general 
ones in R+ and R_ may not always be compatible. 
So prior to semantic interpretation, we expand I2+ 
and R_ into their transitive closures, incort)orating 
all their subrelations in the relation hierarchy. Thus, 
12; :-- { ,.* e I ,' 12+ : is,  ,. }. R*_ is 
correspondingly defined. 12+ restricts the search to 
relations contained in C12 rq 12"4-, iffR+ is not empty 
(otherwise, all elements of C12 are allowed), whereas 
12_ allows only for relations in C12 \ 12"_. 
5 A Sample Semantic Interpretation 
Whenever a semantically interpretable subgraph is 
complete, selnantic interpretation gets started inl- 
mediately. As an example, we will consider a case of 
indirect linkage, as illustrated by the occurrence of 
auxiliary and modal verbs within a passive clause. 
When interpreting indirect syntactic relations, in- 
fl)rmation not only about content word nodes but 
also about intervening noncontent word nodes be- 
comes available. This way, further static constraints 
are imposed on R+ (and 12_) in terms of a list RI~.~. 
C T~. of permitted conceptual relations. This infor- 
mation is always specified at the lexcme level. Since 
12t~:~ relates to closed-class i|;ems only, the required 
nmnber of specifications is easy to survey. 
In our example (ef. Figure 2), the content words 
"Festplatte" (h, ard disk) and "flcliefcrt" (dclivered) 
are linked by a mediating modal verb ("lvann" (can)) 
and a passive auxiliary (%;erden" (bq ..... i,,c)). The 
semantic interI)retation schema tbr passive auxil- 
im'ies (2) addresses the concept type of the instance 
fbr their syntactic subject, Cs,o,j = t!/pe(I.~ul, j) = 
HARI)-DISK, and that tbr their verbpart, C,,e,,1,pa,.t 
= type(Iv~:,.bv~rt) = DELIVERY. The relation be- 
tween these two, however, is determined by Rva**a,~:~. 
:= {PATIENT, CO-PATIENT}, constraint knowledge 
which resides in the lexeme specification for %mr- 
den" as passive auxiliary (of. Figure 1). 
: , , 12 o,,, (2) 
With sia,,x(DELIVEI{Y, {PATIENT, CO-PATIENT}, 
~, HAIl.D-DISK), we get the concet)tual relation 
DEH\ Ea-IATmNT (of. Figure 3), since HM{D-DISK 
is subsumed by PI~ODUCT and, thus, a legal filler of 
DELIVER-PATIENT C 12passaux" 
6 Conceptual Interpretation 
Conceptual interpretation uses a production rule sys- 
tem (Yen et al., 1991.) which accounts for charac- 
teristic patterns of assertions that result froth the 
semantic interpretation process. While the outcome 
of semantic intert)retation (cf. Figure 4) still adheres 
274 
to the surfa(:e fOl'Ill ()t' Ill(} parse(l sent(m('('. (:(m(:el)- 
tual i(ltt?l'pl(~tali()ll abstla(:ts away' fr()ill 1 }IOS(~ Slll'faC(} 
1)\]1(UH)I11(~I1~1 ~11((\[ CI'O,~/L(~a ~l ;lIOl'IlI~diz(~d'~ (:altOllical 
(:OllCel)tual ret)resentatiolt of lhe inlmt, as need(~(t, 
e.g., for mfiformly queryilig th(} kl(owlr;dge lmse. 
As an exanq)le of such inferences consider Figure 
5. with the I)ELI\:I't/S relation linking TI/ANS'II.'(;.9, 
a haMware supplier, aud H..\IiI)-I)I.sI,:.2. By (:om- 
imting a (:on(:eptual relation representing the m> 
de(lying A(:TION TI/AN.qTE(:.!) and tIAItl)-I)I.qK.2 
are integrated in a n(n'malized (:(m(:ept graph. Note 
that the corl'eSlmnding lexi(:al it(!ins, "Tran.s'tec" and 
'q+stplatte" (hard disk), are not linked via a me- 
diated synta.cti(: relation in Figure 2. tlence, we 
lnay (:learly discern semanti(: interl)retation , which 
operates on sinqh, semantically interl)retabl(~' sub- 
graphs only, from concel)tual interi)retation , where 
the inferenc(>llased interl)retati(m of relationshil)s 
among different sift)graphs (:onles int(~ I)lay. 
An ind(~t}(uldent level for (:(m(:ei)tual imerpreta- 
(tOll also bo.came a necessity due to analyti(: consid- 
erations. Often the to(:al constraints for (:onc(~t)tual 
roles of ACTION, ~T..VI'E, or EVENT concepts (:armor 
be formulated restrictive enough for the semantic 
interpretation process. For exanlple, the. (xmceptual 
correlate of the verb "possess" does not impose any 
restri(:tion on its I'ATIENT role (linked to the siltlj(}(:t 
(lotmn(tency relation in a ,~enlantically into.rpretal)lc' 
sul)grat)\[ O. I{ather, restric.tions apply to l(roperly 
relatin9 the filM of the. PATIENT slot with that of 
the CO-PATIENT slot (dirobjoct, at lh0. do.1)enden(:y 
h.'vel). C, oncet)tual interl)retation rules are a means 
to further constrain these :cont(.~xt--sensitive ' aslm('ts 
of the interpretation I)rocess. 
Sin(:(} verbs play a prominent role in dcI)emlency 
gratmnars, the production rule system for con(:el> 
tual interpretation is ba.qed ut)on the conceptual COl'- 
relat, es of verbs (h/}ncetiwth verb concepts) in the 
k(lowl(.,(lg(.' base. Ditfer(}nt views are defined for ve,b 
con(:epts t) 3' using three ~d)st;racti(m dinmnsi(ms. 
First, verb concel)tS are classified, ac(:()rding to 
the set of thematic roles riley supply, as ACTION, 
~TA'I'E or Pt(OCES$. I)EIAVH{Y, e.g., is assigned to 
:\(TI'ION, Sill(;(} both A(;ENT and I'ATIENT form part 
of the concept detinition (of. Figure 3, right box). 
The second level of al~straction (:ol).sists (~f (:at:ego- 
rizations which reflect a ('omnmn core meaning. The 
upnlost conceptual node in this hierarchy is CATI.;- 
(;()I/Y. \]-)HAVH/Y, e.g., is considered as a C(}IIcopi 
whi(:h repres(}l~tS the ACTION of transfering a GOOl) 
"l'l'al IS;t ec.9 
i ¢:~:, Ilal'd-i)isk.2 
deliverl \[ ..... I.l .... ('OIllplll el" Sysielln.4 
hard di,k-ur \[ ~I~:: , 35(Mhz-CPU 6 
has-cpu 
Figure 5: A Sample Concel~tual Interllretati(m of the 
1)op(m(lency Graph fl'om Figure 2 
t() a customer. All verb (:(mcet)ts belonging t() this 
Ci(l;(~rol'y HI'O Slll)S(llllOd })y the correst)onding COl> 
COl)t. (2AT--TIIAN.qFI.;II..G()OD. (\Ve here make llse 
of nnfll, ilfle inheritance ln(!(:hanisnls.) 
Finally, every verb con(:et)t is linked to s(-)Ill(~ 
\-Iqt/I/-~IOI)F.I, \])I'LI\;EI/X/or ally other \,el'l) concept 
()f the C_-\T--TII.\NSI.'EI/-(~OOl) category is a con-. 
stiLu(,.nt i)hase of the BI;Y-ANI)-,~I.;IdJ-~\[OI)EL. "l\]() 
gem'.ralize appr()t)riately fr(nn individual verbs, vml) 
cat.egorie:; were extracted from ore text corpora that 
further refin(~ a large-scale taxononly for Gernmn 
verbs (Balhner and Brelmenstuhl, 1986). hi this 
\v(M(, a total of about 20,000 verbs \\;ere subsumed 
by 700 categories to reflect a semantic generalization 
ill l:erms of a hie.rarchy of verb categories. 
The production rules for cont:el)tual interim;Cation 
operate on this calegorial hierarchy. Every verb con- 
Cel)t in the hierarchy is a sul)con(:ept of exactly one 
(:ateg()ry in the knowledge base. Whe,mver the pre- 
c(mditions of ;111 imert)retati(m rule are fulfilled, a 
concel)tual interl)l'(.'tat, iolx ix con(puled. 
Coil(:el)tual and semanti(: int(U'l)retation d('.pend 
on each other, since the bast(: interpretation schema 
(of. expression (1) in Section 4) is supplied with ac- 
tual t)aramete.rs frolll t)rodu(:tion rules. We there- 
fore may de.line another specialization of the basic 
int(.wl)retation schenla for (;on(:el)tllal interpretation 
sico,,,,. In particular, path searches are triggere(l 
that are re.~;ia'i(:ted by a positive list l'end(~red l)y the 
apl)licable production rule. 
For our ,~ample sent(}n(:e (of. Figures 2 and 4), the 
/:oncel)tual (:orr(,latc' for the verb "delivers" (DI,:- 
IAVEH.Y) is a sul)concept of A(:TIt)N. Addition- 
ally, l)l.'l.IVl.:l/Y is a mlbconcei)t of the (:ategory 
CAT--Tt/:\';SFI.'I~-Go()I) ((:f. Figure. 3). The. corre- 
sponding (:on(:el)tual interl)retation rule is given in 
Fi{{(lro, (J. Wht}llever }111 instance of /,he category 
(~A'I'-TItAN,qFEIt-.(~OOI) is encounter(,.d and \[)oth its 
:\(IENT and I'ATIF.NT r(}los ale. filled, relation paths 
are (:omput;ed from th(; types of the two instances 
involved, a and p, reslm(:tively. For each relation 
found by the search algorithnl (l~\]'2L in Figure 6), 
a correspondil~g assertion is added to the knowledg(.' 
base (TELL in Figure 6). In the examt>le, the in- 
terl)retatioll s(:hmna is inst;mMated with the 4-tuple 
((:OMI':\NY, {TI~:\NSI"I':I~S-(;O()D}, {}, H.M~I/-I)IsK) 
resulting in t, he comfmtation of {I)I:.LIVFA/S} as the 
in'Olw.r re.lation link (of. Figure 5), since it is a sub- 
ro.lation of TI/AN.qFEI/S-(;OOI). 
EXISTS v, a, p: 
t; : CAT-TRANSH.:R-(~OOI)F1 
t' A(IENT ,'1 M o PATII,'NT p 
IF .~i ........ (tUp,'(,z), {T,..~NSF,.:RS-GOO,)}, {}, tVp,~(p)) ¢ 
TItEN 
I{EL := .,;i ........ (t!jpc(~), {TII~.NSI.'I.:ItS-C;OOl)}, {}, type(p)) 
TELl, o, Itl,;l~ p FOR,AId; Itl,:l, ~ HEL 
Figure 6: Samt~le (~onceptual Imo.rpr0.tati(m I/,ule 
275 
7 Evaluation 
We evaluated this approach to senmntie intert)reta- 
tion on a random selection of 54 texts (comprising 
18,500 words) from two text corpora, viz. consumer 
product test reports and medical finding reports. 
For evaluation purposes, we concentrated on the in- 
terpretation of genitives (as an instance of direct 
linkage) and on the interpreta.tion of t)eriphrastic 
verbal complexes, i.e., passive, temporal and modal 
constructions (as instances of indirect linkage). 
The underlying ontology consists of an upper 
generic part (containing about 1,500 concepts and 
relations) and domain-specific extensions relating to 
information technology (IT) and (parts of) mmtomi- 
cal medicine (MED). Each of these two dolnain mod- 
els adds about 1,400 concepts and relations to the 
upper model. Corresponding lexeme entries in the 
lexicon provide linkages to the entire ontology. 
We considered a total of 247 genitives in the smn- 
ple. Recall was higher for medical texts (57%) than 
for IT documents (31%), though, in general, rather 
low. However, precision peaked at 97% and 94% fo," 
medical and IT texts, respectively. The mnnt)er of 
syntactic constructions with modal verbs or auxil- 
iaries amour to 292 exmnples. Compared to geni- 
tives, we obtained a slightly more favorable recall 
for both doinains 66% tbr MED, 40% for IT --, 
while precision dropped slightly to 95% and 85% for 
nmdical and IT documents, respectively. 4 
As with any such evaluation, idiosyncrasies of the 
(;overage of the knowledge bases are inevitably tied 
with the results and, thus, put limits on too far- 
reaching generalizations. However, our data reflect 
the intention to submit a knowledge-intensive text 
un(lerstmlder to a realistic, i.e., conceptually un- 
constrained and therefore "unfriendly" test environ- 
ment. Judged from tile figures of our recall data, 
there is no doubt, whatsoever, that conceptual cov- 
erage of the domain constitutes the bottleneck for 
any knowledge-based approach to NLP. s Sublan- 
guage differences are also mirrored systematically in 
these data, since medical texts adhere more closely 
to well-established concept taxonomies and writing 
standards than magazine articles ill the IT domain, 
whose rhetorical styles vary to a larger degree. 
8 Related Work 
The standard way of deriving a semantic interpre- 
tation for constituency-based grammars is to assign 
each syntactic rule one or more semantic interpreta- 
tion rules (e.g., van Eijck and Moore (1992)), and to 
4A more detailed presentation of this evaluation study is 
given in Romacker and Itahn (20001)). 
5For the medical domain at least, we are currently actively 
pursuing research on the semiautomatic creation of large-scale 
ontologies from weak knowledge sources, viz. medical termi- 
nologies; cf. Schulz and Hahn (2000). 
determine the meaning of the syntactic head fl'om its 
constituents. This approach has also been adopted 
in the few explicit attempts at incorporating seman- 
tic interpretation into a dependency grmninar fi'mne- 
work (Milward, 1992; Lombardo et al., 1998). There 
are no constraints on how to design and orgmfize this 
rule set despite those that are imI)lied by the choice 
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, are lacking. Accordingly, the number of rules 
increases rapidly and easily reaches orders of sev- 
eral lmndreds in a real-world setting (Bean ctal., 
1998). As an alternative, we provide a small set 
of gencric semantic interpretation schemata (by the 
order of 10) and conceptual interpretation rules (by 
the order of 30 for 200 verb concepts) instead of 
assigning specific interpretation rules to each gram- 
mar item (in our case, single lexemes), and incor- 
porate inheritance-based abstraction in the use of 
these schemata during the intert)retation process in 
the knowledge base. We clearly want to point out 
that while this rule system covers a wide variety 
of standard syntactic constructions (such as gent- 
tives, prepositional phrases, various tense and modal 
forms), it currently does not account fbr quantifica- 
tional issues (like scope ambiguities) for which en- 
tirely logic-based approach (Charniak and Goldman, 
1988; Moore, 1989; Pereira and Pollack, 1991) pro- 
vide quite sophisticated solutions. 
Sondheilner et al. (1984) and Itirst (1988) treat 
semantic interpretation as a direct mapt)ing front 
syntactic to conceptual rel)resentations. They also 
shm:e with us tim representation of doinain knowl- 
edge using Kl,-ONE-style terminological s, 
and, hence, they nmke heavy use of property inher- 
itance (or typing) inechanisms. The main diflbrence 
to our approach lies in the status of the semantic 
rules. Sondheimer et al. (1984) attach single in- 
terpretatiotl rules to each r'olc (filler) and, hence, 
have to provide utterly detailed specifications re- 
flecting the idiosyncrasies of each semantically rele- 
vant (role) attachment. Property inheritmme comes 
only into play when the selection of alternative se- 
mantic rules is constrained to the one(s) inherited 
from the most specific case frame. In a similar way, 
Hirst (1988) uses strong typing at the coueeptual 
object level only, while we use it simultaneously at 
the grmnmar and the domain knowledge level for the 
processing of semantic schemata. 
9 Conclusions 
We introduced an al)proach to the design of com- 
pact, yet highly expressive senmntic interpretation 
sdmmata. They derive their power from two sources. 
First;, the organization of grammar and domain 
276 
knowledge, its well as semantic interpretation mcch- 
alliSllIS: al'e lmsed on inheritance principles. Soc- 
ou(1, interpretation schemata al)stract from 1)articu- 
lar linguistic phenomena (spe('ilic lexical items, lex- 
eme classes or dependency relations) ill terms of gen- 
eral contiguration l)atterns in (tepo.nden(-y gral)hs. 
Underlying these design decisions is a strict sep- 
aration of linguistic and eouceptual knowledge. A 
clearly defined interface is provided which a.llows 
these st)ecitications to nmke reference to line-grained 
hierarchical knowledge, no ma.tter whether it is of 
gramnmtical or conceptual origin. The interface is 
divided into two levels. O11o nmkes use of static, 
high-level (:onstraints sut)l)lied l)y the nlal)l)ing of 
syntactic to conceptual roles or sut/1)lied its the 
meaning of closed word classes. The other uses these 
constraints in a dynanfic search through a knowl- 
edge base, that is l/arametrized by few and simI)le 
schenmta. Finally, at the level of conceptual in- 
terprotatiou inferences emerging fl'om senmntic rq)- 
resentations are COml)uted by a s(.'l; of t)roductious 
which make reference to a verbcategorial hierarchy. 
Also since the numl)er of s(:hentata at the semantic 
description layer remains ratho.r sulall, their o.x('.cu- 
tion is easy to tra(:e and thus SUl)l)orts the main- 
tenanco of largo-scale NLP systenls. The high ab- 
straction level 1)rovided by inheritance-based seman- 
tic sl)ecilications allows easy 1)orting across (liflhr- 
ent al)l)lication domains. Our exl)orienco resls on 
reusing the set of s(;mald;ie sehenmta once deveh)t/ed 
for the information technoh)gy domain in the nm(ti- 
ca\] domain without further (:hallges. 
A(-knowledgments. Wc want to thank the members 
of the l~{group for close (;OOl)eration. M. Romacker was 
SUl)t)ortcd by a grant fi'om DFC (Ita 2097/5-11. 

References 

J. Allen. 1993. Natural , knowledge rep- 
resentation, and logical tbrm. In M. Bates and 
R. Weischedel, editors, Challc'nges in Natural Lan- 
guage Proccssinfl, pages 146 175. Cambridge Uni- 
versity Press. 

T. Balhn(?r and W. Brenncnstuhl. 1986. Deutsche 
Vcrbcn. Einc ,sprach, analytisch, c Untcrsuchunfl des 
deutschcn Verbwortschatzcs. Tiil)ingen: G. Narr. 

C. Bean, T. Rindtlesch, and C. Sneidernlan. 1998. 
Automatic semantic intert)retation of anatonfic 
spatial relationshil)s in clinical text. Ill Prec. 1998 
AMIA Annual Fall Symposium, pages 897 901. 

E. Charniak and R. Goldman. 1988. A logic for semantic interpretation. In Proc. of the 261h Annual Meeting of the ACL, pages 87-94. 

L. Creary and C. Pollard. 1985. A computational 
semantics for natural . In Proc. of the 
23rd Annual Meeting of the ACL, pages 172-179. 

M. Dalrymple. 1992. Categorial semantics for LFG. 
In COLING '92 Proceedings of the I5th, International Conference, pages 212-218. 

U. Itahn, S. Sohacht, and N. BrSker. 1994. Con- 
current, object-oriented natural  t)arsing: 
the PAI{SE'I'AIA{ model. International Journal of 
II'aman-Computcr Studies, 41(1/2):179 222. 

E. Hajieova. 1987. Linguistic meaning as related to 
syntax and to senmntic interpretation. Ii1 M. Na- 
gao, editor, Language and Artificial Intelligence, 
pages 327-351. North-Holland. 

G. Hirst. 1988. Senlantic intert/retation and ambi- 
guity. Artificial Intelligence, 34(21:131-177. 

V. Lombardo, L. Lesmo, L. Perraris, and C. Sei- 
donari. 1998. Incremental interpretation and lex- 
icalized grmnmar. In CogSci'98 - Proceedings of 
the 20th Annual Conference, pages 621-626. 

D. Milward. 1992. Dynamics, dependency grammar 
and incremental interpretation, In COLING '92 
Proceedings of the 15th, International 
Conference, pages 1095-1099. 

R. Moore. 1989. Unification-based semantic interpretation. In Proceedings of the 27th Annual Meeting of the ACL, pages 33--41. 

F. Pereira and M. Pollack. 1991. Incremental inter- 
pretation. Artificial Intelligence, 50(11:37-82. 

M. \]~,mmck(?r and U. tIahu. 2000a. Coping with dif- 
ti?rent types of anlbiguity using a uniform context 
haudling nmchanism. In Applications of Natural 
Lang'aaflc to h~formation Systcm.s. PTvccedinfls of 
the 5th, NLDB Confc~w~,cc. 

M. Romacker and U. Haln. 2000b. An empirical 
assessment of semantic interpretation. In Proc. of 
the 6th Applied Natural Language Processing Conference, and 1st Conference of the North American Chapter of the ACL, pages 327-334. 

S. $chlflZ and U. lIahn. 2000. Knowledge engineer- 
ing I)y large-scale knowledge rollS(?: OXl)eriem:e 
from the inedical donm.in. In KR '200(\] Prec. ~\[t\[t 
Inter'national Conference., l)ages 601-610. 

N. Sondheimer, R. Weischedel, and R. Bobrow. 
1984. Semantic interpretation using KL-ONE. In 
COLING '84 Proc. 1Oth International Conference and 22nd 
Annual Meeting of the ACL, pages 101-107. 

M. Strube and U. Hahn. 1999. Functional centering: 
grounding referential coherence in information 
structure. Computational Linguistics, 25(3):309-344. 

a. van Eijck and R. Moore. 1992. Semantic rules 
for English. In H. Alshawi, editor, The Core Lan- 
flua.qc Engine, pages 83-115. MIT Press. 

J. Wedekind and R. Kaplan. 1993. Type-driven semantic interpretation of f-structures. In EACL '93 Proc. 6th, Conference European Chapter of the ACL, pages 404-d11. 

W. Woods and J. Schmolze. 1992. The KL-ONE 
fiunily. Computers "U Mathematics with Applica- 
tions, 2312/51:133 177. 

J. Yen, R. Neches, and R. MacGregor. 1991. 
CLASP: integrating term subsuml/tion systems 
and l)roduction systems. IEEE Transactions on 
Knowlcd.qc and Data Engineering, 3(1):25--32. 
