File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/94/c94-2128_metho.xml
Size: 29,590 bytes
Last Modified: 2025-10-06 14:13:40
<?xml version="1.0" standalone="yes"?> <Paper uid="C94-2128"> <Title>TItE MFRGEI) UPPER MODEL: .A. L\]NGUISTI(; ONTOLOGY FOR (3ERMAN .AND ENGL\]StI Reaiaf,(~ tt(;nsehel aud ,fohn ilat;,enian ~ Projcci, I(OMF, T</Title> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> 1 INTI/.OI)UCTION: MULTII,INGU AL LINCIUISTIC q)NTOI,OCIIES' </SectionTitle> <Paragraph position="0"> Wil,h the need to develop re-usahle rr~un(,works \['or organizinginformalfion (of., e.g., I.he ARI)A Knowl edge Sharing l';ffort \[/)atil cl al., 1992\]), workM)le proposals lbr 'ontologie.q' 1,o Imwiclc sm:h orgaui.</Paragraph> <Paragraph position="1"> Zal,iOlt 32;C iiicreasing 111 ill)\])Ol'Lall(:(L ~llch Oll\[.o\[ogies are now commonly applied in Natural I,anguage Processing systenls sillce t,h(?r(? {,h(? )'(qireSellt.al.ion ()\[' COtlltllOtlS(!ltS(~ ~111(\[ (IOllH/ill klloWl(Xig(' is ess(mtial. Accordingly, a numher or org;a)tizal, ions of knowledge have been dcvdol)Cd some quite cxtensiw:. Th,~sc orgnnizal.ions a.rc (waluntcd hy th(~ (!xt(:lll. l.o whi(:h they I)rOVe re usable ;~cl:oss distinct, domains aim al>plications. The consideration of the reuse of these orga.niza{,ions is, however, co,ul)lical.ed by the rauge of difl'eriHg design criteria I h~Ll. are employed ill their (;o\[Isl;rllC{;ion; all (;xl,eltsiv(; overview of 0,\[)\[)1'O3A:\]1(28 is givcu it, \[llatcnm.n, 1992~\]. One i)articutar apl)roa.ch is ~o define linguistically mogivatcd ontologies, wh(.re l.he criteria \['or organizal;ion rest. on seman{.ic clistiilcl.ions I.h~fl; l;he gr0.ll)lll~ll' ()\[' a \[;Ulgllage 11(2C(\[8 \[.O \]l;I.VO (Ira.wn iu oMer to mol;iwfl.e il.s deployln(ml, of grant.</Paragraph> <Paragraph position="2"> rnalAcal (lisl,incl.ions.</Paragraph> <Paragraph position="3"> Two sizcahle liuguistic ontologies col)sl;ruct.ed in this way are t.he Penman (lpi)(':r Model d(welot)cd for I&quot;,nglish text genel:al;\]Oll within the l)enman i)rojecl.</Paragraph> <Paragraph position="4"> at US(J/ISI \[Prmn,an I'roj(,ct., 1!)8!)\] and tim Upper Model for (\]erimm develol)('d similarly for (,ext g(!ll(!l' ation within the. KOMWr l)rojcct at GM l)/ll'SI \[llatc.</Paragraph> <Paragraph position="5"> ma,it c* al., 1991a\]. The English Ul)pcr Model (l'3/M) is (les(:rib(:d in \[llat(;nlan cl al., 1990\]; th(,. conccl)tS of' Lhe (k.'rrmm Upper Model (('UM) go Imck tlc) \[Stcim!r cl ,'d., 1988\] aml \['l'cich, 1992\]. Ik)th cqltologics arc individuMly examples c,f the most det.a.ih,d sn,:;h on.</Paragraph> <Paragraph position="6"> tologics curreul.ly und(w development., each with over 200 doma.iu and al)l)li(:ation indcl)end(mt (:oncel)ts *Also on indefinite I<tv(~ 1'1&quot;01\[1 the \]'vnma.n I)rojc(:t,</Paragraph> </Section> <Section position="3" start_page="0" end_page="0" type="metho"> <SectionTitle> 1Lq(J/hffornta.tion Sciences Institute, Mariner dcl lie.y, I,os </SectionTitle> <Paragraph position="0"> A ngcles.</Paragraph> <Paragraph position="1"> arranged in a subsumption latl;icc, spraining dis ~iuct t, ypcs of processes (menl;M, communication, re lational, aclions), aud diverse quMitics and object;s. Both onl, ologies ha.vc I~('.cu used in ~t nmnber or (Io. mait,s and show good re-usalfility characteristics mainly due to the l.tct that they :m: liuguistically motiw/.ted. Thus, for exauiple, if language generatioN is required there is tylfically I00% r('. usabili{:y across doniahis it, co,fl.,'ast to the 50% clescrib(~d by \[I)irlein, 1!)ga\] fo,. l;he, largely ,ion-linguistically niotivated, I,II,OG onl, ology, l 'l'here are a mlnilyer of sugg,;stions for the eval ual.iou o\[' ontologies in t(;rlttS of \['orlnal prop(!r ties of consistel,ce and cohorenc, e of the inform;)tioN those outologics contain (e.g., \[lloracek, 1989, (:hmrino, 1994\]). With a restriction 1.o linguisticMly mol iw~l;ed onl.ologies, we can now sg~i.,; rurthcr (1(; sigu principles collcerning whal; is to 1oe represented and how. All;hough Lhese l)rinciplcs were origiually develolmd in oMer to carry out a detailed coull)ar ison of I, he IqUM and th(; GUM, they are g(;nerally apl)lical)l(~ for all liuguistically molJvated ontologi(.s; ewdua.ting t;Im couc(;pt.s l)ropos(;d within su(:h an on tology accordiHg to the principles (les(:l'il)(!(I in t.his paper should iml)rovc {.lie sl.al;us of thud, ontology overall.</Paragraph> <Paragraph position="2"> The main rosull, of the IqUM-C;UM comp~rison is a zIhrged Upp(r Mo(hl presently used within I;h(: KOMI.'/I' pro jecl; as the basis rot multiliugual sentence gel)eration in I'hlglish, (~ermau and I)utch. ~ This l;\]letl also i)rovi(les ~m early answer to a quesi.ion cot, (:eruing a (lilferenl; kind of re-usability of liitgtiisl;i. (:ally motiva, ted ontologics, i.e., l;hc (!xtcnt to which they can be. re-used across distinct languag~.s ,'aLher tha.u across distinct donia.ins.</Paragraph> </Section> <Section position="4" start_page="0" end_page="806" type="metho"> <SectionTitle> 2 TIlE MEllPS~ING MI&quot;,THOI) 2.\] Starting points </SectionTitle> <Paragraph position="0"> Merging dislfiuct ontologies is a I)rot)lcm thaI, will oc-Cllr lllOFO fr(?(lllelll;ly as llCW plX)poS\[-/.lS ~'ul'C ~,o \[)e \['('.('oIloiled. \[tlovy aud Nirenl)m'g, 1992\] i)rol)ose a gelll!ral nlethod for creating a merged ontology out or dif fercnt ontologics where it. does not mattc'r whel, hcr I lun Is smq>ly lmca.usc a linguistically motiwdxM on-I.ology is ImLutd to the. sem~nfl.k:s of a grammar and not to genel'a\[, possibly (\[oinain-l.rattscendenl. knowl.</Paragraph> <Paragraph position="1"> edge. 'l'},e two kinds of onl, ologics should t.hcreR)rc be seen as perk)rnihig dillT~rcnt kinds of work. For ext.en siw~ moi.iwfl.ions for imfinl.ainhlg a linguisti<:Mly moli wired o.tology, see \[llMlida.y a.ml M ~ttl.h\[esseu, t,o ~tpt)ea.r , llatema,t, \[992a\].</Paragraph> <Paragraph position="2"> >l'he comparis(m is based on tim \]';nglish Upper Model ~tiid (',erma.n U pper Model data files from .I uly 1992.</Paragraph> <Paragraph position="3"> Both arl'(\] expressed in the kn<>wledgc representa.tion laat gua.gc t,()oM (\[Mac(Iregor mul lh:ill, 1989\]).</Paragraph> <Paragraph position="4"> differences are language dependent or due to differeat linguistic theories. The commonalities and differences in two ontologies are classified according to Hovy and Nirenburg as follows: ontologies.</Paragraph> <Paragraph position="5"> Extension: There is a concept ill one ontology which is missing in the other, but which specializes the latter ontology fllrther.</Paragraph> <Paragraph position="6"> Cross classification: '\['he partitioning of identified concepts into subeoncepts differs in the considered ontologies.</Paragraph> <Paragraph position="7"> The merging procedure t, hen keeps all concepts of cases (1) and (2) and resolves ease (3) by exhaustive cross classification.</Paragraph> <Paragraph position="8"> A simplified version of this procedure is proposed in \[How and Knight, 1993\]. Here, the cross elassification resulting fi-om nonmatching partitions of identified concepts into snbeoncepts is replaced by parallel subordination of those subconcepts. This results in a substantial reduction in the concepts necessary, but leaves open l;he question of the mutual relation between concepts stemming from different source ontologies. Participating NLI' components can be controlled well by such n shared ontology, but its adequacy as a point of communication in a joint MT systent is less clear.</Paragraph> <Paragraph position="9"> We have found that it is necessary to go beyond the original merging methodology in a number of ways. Nevertheless, as a eonsequence of the criteria for merging that we propose, not all existent concepts of the EUM and GUM need find their representation in the Merged Upper Model and cross classilication is still significantly reduced without impairing inter-translatability across concepts arising fi'oln dif-Drear source ontologies.</Paragraph> <Section position="1" start_page="0" end_page="0" type="sub_section"> <SectionTitle> 2.2 Problems with identity </SectionTitle> <Paragraph position="0"> The crucial point in \[tlovy and Nirenbm'g, 1992\] is the notion of 'identity'. The decision how to deal with different concepts (identification, extension, or cross classification) is based on the possibility of stating an identity between concepts of difl~rent language ontologies. This is somewhat problematic. \]\[21 the comparison between the English and the German upper models, we took as identificatiou criterion the equivalence of the sentences or phrases which can be generated by the concepts. This correspondence relies on the assumption theft German and English sentences have a one-to-one-mapping and that translation is a totally information preserving relation. Although this is not true in general, we based our merging on the assumption that it may be true for simple sentences if we abstract out the textual and interpersonal i.e., non-ezperieniial (examples below) dimensions of utterances, and tim language distance is close. Hence, the whole construction has to be seen in the context of its own relativity.</Paragraph> </Section> <Section position="2" start_page="0" end_page="806" type="sub_section"> <SectionTitle> 2.3 Further Principles Prinniple 1: Removal of non-o.xperiential eon- </SectionTitle> <Paragraph position="0"> nept discrimination DifIicultics can arise when ontologies to be merged are themselves inherently problematic in some way. Internal problems should not be autolnatically transmitted to a merged ontology. Thus, during merging, the distinctions drawn iu individual ontologies need always to be evahlated internally before being admitted. The exclusion of textual and interpersonal inforn'mtion in a merged upper model provides an additional important criterium for an 'extended identilication' of concepts within ontologies to be merged. Two comnlon kinds of non-experiential concept discriminations were found in the GUM. The first introduces distinct upper model concepts in order to motivate lexieogrammatical realization by differing types of grammatical units. The second introduces distinct upper model concepts to motiw~l,e, the selection of semantic roles Dora a given semantic conliguration that are to be lexicogrammatieally expressed. An example of the. first ldnd is offered by the concepts (Ll~elalional and C-l~elalionship :~. These are both responsible for the generation of processes experientially classifiable as relational, but whereas Ggelalionship causes an attributive or adverbial realization, G-Relational causes a. clausal realization.</Paragraph> <Paragraph position="1"> Thus, phrases (1) and (2) can only be generated from dill>rent semantic input expressed here in the form of the typed semantic assertions of the Penman Sentence t)lan Language (sPt,) \[Kasper, 1989\]: 4 (1) Das Mitdchen ist kra.nk. (The girl is sick.) (a / classificatory :attribuant (m / person :lex m/idchen) :classifier (k / quality :lex krank)) (2) das kranke Miklchcn (the sick girl) (a / property-ascription :domain (m / person :lex m~dchen) :range (k / quality :lex krank)) This problenl does not surface so often ill the EUI~\], althougb there are occasional violations e.g., the inchlsion of 'rhetorical relations' that are explicitly textual (see \[Bat(man el al., 1990\] for details).</Paragraph> <Paragraph position="2"> Examples of the second kind a.re offered by seutenees (3a-b).</Paragraph> <Paragraph position="3"> (3)a. l)cr Lehrer antwortcl,, dass (hts Raumschit\[ zuriick gckehrt ist.</Paragraph> <Paragraph position="4"> (The te~tcher answers that the spaceship has term:ned.) h. l)er bchrer a.ntwortet den Schfileru, (lass das l{amnschitf zur/ick gekehrt ist.</Paragraph> <Paragraph position="5"> (The teacher a.nswers the studeats that... ) The differences in (3) arise from differences in the numbe.r of semantic participants in the answeringevent that are made grammatically explicit. Both (3a) and (3b) could be used to deseribe the same experiential event, the selection being made on non-experiential grounds (e.g., lack of relevance of a participant, being known from context, known from preceding text, etc.) specific to the text being created. 3Conccl)ts fi:om the English Upper Model and the German Upper Model will be differentia.ted where rcl-evant by prefixing either 'E-' or 'G-.' as appropriate. 4 Classificatory is it subtype of GUM concept l~c'lational, propert!j-ascr@tion a subtype of relationship. Furthor, in the SPL examples in this paper, lcxic~L1 selection is specitied directly by means ef the keyword : \]ex to ~tvoid complicating the discussion mmeccssa.rily.</Paragraph> <Paragraph position="6"> Both role, however, classilied semantically underneath distinct GUM concepts. These distinct con eepts have differing obligatory role conligurations, which requires that the selection of semantic (ex pcriential) I, ype has go be Inade aceording to the participant, s that ,~re to be expressed a decision that is often made ou textua.l grounds without a change in experiential perspective, lilt the FUM the only semantie distinction in <his area of con> reunite<ion processes is betwe.en 'telling' like ewmts (address<c-orient<d) and 'saying'-like events (nonaddressee-oriented), which is ~ dill<fence in experiential lmrspective. 's In the I';UM, concept discrimination is made with respeet to difIi'.ring possible realizalious of rob:s, not strMghtlbrward absence/presence of roles as in the GUM. 'Missing' surface par<Jell)ants ean be modelled more adequately by an tipper mode.l-grammar interface which Mlows detined sem~mtic roles to ha.ve zero realization. '\['his is an elegant way to deal with optional participants, passive, and iml)ersollal eonstructions.</Paragraph> <Paragraph position="7"> The net effect of both kinds of violations (; of principle I is tha, t the number of concepts is increased and itecessary decisions eone(;rnillg lexi('ogralllmatical realizations arc avoided. 'lPhe. proliferation of concepts if Mlowed would complicate considerably the task of 'identifieation' of similar coneepts in ontologies to be merged.</Paragraph> <Paragraph position="8"> linen \[bllowing extended i(lentifieation of concel)ts , it is uo~ suili(:ient to provide cross products \[br those eoneepts that ~re not identifiable but which classify overlapping sema.,ttie areas. '\['his is clarified l)y the following concrete, al{,hough \[ll/ICtl abbreviated, ex-. ample, of merging in tim al'e&L of material (action) process types. TIll; decisions that ar(; required here are typical of the merging process as ~t whole.</Paragraph> <Paragraph position="9"> The l';-Malerial-Proccss hierarchy distiuguishes processes more or less with regard to transitivity pattcrning. Au l';-Nondirecled-Aclion is a process without external eausation (n,ostly intransitive, although transitive sentences where the objee.t is not alD(;te.d or created by t, he aetion Mso fall into this class), li-Ambienl-Process and l';-Motion-Process are not ex.. haustive subconeepts of L'-Nondirected-Aclion.</Paragraph> <Paragraph position="10"> '\]'he wtse broke. Nondirected-Aetion i pla.y pi~um. Nondircctcd-Action The to,<fist ran. Molion-t'roccss \]t rains here. Ambient-Process An E-Directed-Action in contrast is a proee.ss with an external causer as additionM tmrticipanL. /','\])ireclcd-Actions divide into I':-Crealive-Mahrial-Action and I','-l)isposilive-MateriaI-Aclion. 5'\['he \[()\]?nlel&quot; neeess;trily illvolws a.n addressee n(!lll~l, ll-. ticM\[y, sonleone who in intended to be listeuing, while the latter does l|ol;. This diiR~rence is gramma.ticized tn \[,;nglish in the a.cceptabillty/lmn-a.ceel)tMfi\]ity of '\[ told him thai .... '/'1 s~tid hint tha.t... '. Ill order I,o gratntmaticize a.n *Mdressee iu a. s~tying-like event, it, is necessary to respect its lens centrM role ~tnd 1.o use the form '1 s;tid to hi*n fl~tt... '.</Paragraph> <Paragraph position="11"> ~'l'he,:e ~tr(.' further similar cases; for ex~tlnph~ tit(! GUM also inelmles interpersonally motive.ted eon(:el)t discrimin~Ltions such ;ts negativc@:ctlurc.-ascriplion ~t,td ncgativc-quality. These govern the general, ion of m:ga.tiw~ assertions, Lhus pre erupting at lllOITc *q)propriate speech function co<tirol of ncg~ttion.</Paragraph> <Paragraph position="12"> The chihl broke, the vase.</Paragraph> <Paragraph position="13"> The lion chased the tourist.</Paragraph> <Paragraph position="14"> Mary baked a cake.</Paragraph> <Paragraph position="16"> The GUM differentiates G-Agenl-eenlered, G-Affected-centered, G-Agent-only and G-Affected-only as disjoint (7-Action subtypes. \]lere, we have at first a classifieation with regard to kind and number of partic.ipants. Example.s for the semantic representations of the intransitive process types are given in (4) and (5), again ill Sel, notation: (4) I)er'fimrint ramtte. (The tonrint ran.) (r / action :lex rennen :agent (t / tourist)) (5) I)ie \]'tlanze geht ein. (The plant is dying.)</Paragraph> <Paragraph position="18"> The transitive processes (with t;wo participants) ~rc I'urther broken up into G-Agent-centered and G'-Affected-centered. The G-Affected-centered process type is a. very special case of a transitive process. 'l'he detinition is giwm ill \[Steiner el el., 19881 thus: &quot;X affe.clcd-cenlered-vcrb Y iff X causes that Y a.lfc.cted~ <chief'<d-verb&quot;. I';xtunplcs are: I)as Kind zerbricht (lie Vase.</Paragraph> <Paragraph position="19"> l)as Kind bewirkt, class die Vase zerbricht.</Paragraph> <Paragraph position="20"> The chihl breM(s the vase. ~ * The ehihl brings it about that the vase bremen.</Paragraph> <Paragraph position="21"> Thus a process is called G-Alfected-cenlered if the reMizing verl) is ahh'. to \['orln an ergal:ive pMr. All C-Ajfected-ccnlcred processes have at least two participants, the G-A:lenl and the G-Aft<<led.</Paragraph> <Paragraph position="22"> The (;-Agent-centered process is di!l'crenti~ted with regaM to the different p~rtieipant types for <,he At first, sight, there are few con.nonalities t)etween the.se two onl, ologies. Without deeper hltrospecCion, one can only state an identity E-Ambient- Process =: G- Natural- I'henomcnen mid could me<haul<ally build a cross classification as shown in Figure 1. Sonic ereated concepts should, howew:r, bc omitted from this 'eross produet ontology'. null The \[hrsl, obvkms argumen~ is tile mmdmr of par tieipants. 'Phese are contradictory in the R)llowing cross concepts: li- I)irccted- A <lion/0?- Agcnb. oMy and I'L Direcled-Aetion/ G-Affcctcd-only. A compari-son of the low level eoneepts shows further that the I,'urthermore, it is known fron-i tile definition of E-Nondirecled-Aclion in \[l:~ateman el al., 1990\] that such processes axe either intransitive or they have a second pm:ticipant which is in meaning nothing else ~re rnled out. Fin,nlly, tile exhaustive coverage of the low level subtypes in the I,;UM and GITM supports the tbllowing identities:</Paragraph> <Paragraph position="24"> By these kinds of deta.ilexl considerations, we have filtered an intelligent merge out; of the mechanicM merge. Within the intelligent merge, we omit the German differences concerning tile participant nnmbet (G-dge~tl-o~dy, GGganging) since these violate principle I, and do not estM)lish the very subtle (L AJfccled-cente~vd type. Preferring the I~nglish terminology the result is giwm in l!'ignre 2.</Paragraph> <Paragraph position="25"> This turns ont i~o I)0 mainly tile EITM subhierarchy for material processes. 'lb also cover the German requirements, the Nondirecled-Aclion concept is difDrentiated into Nondirecled-I)oing and Nondircclcd-Happenin 9 according to the distinctiou between Agent-only mM Affecled-only. q'herefbre we do not need to preserw', the Clerman participant types d.qe,tt and A Jr(clod, and can inDr the releva,t inff)rmation from tim new Nondireclcd-Action sub cottcepts. The (.4erman SPL examples (4) and (5) then have the revised semantic form:</Paragraph> <Paragraph position="27"> Because we have \[ixed the semantic differences between the G-Aqcnl and the G~AJ\]?.clcd participant in the process types we do not need this differentialion as partMpant roles again, tlence, we choose the l';nglish p~rticipant types h'-Aclor and I';-dclee, tim correspondence of which to the German G-Agent, G-Affected, G-Eft((ted and G-Process-range ditDrs with the process type (see l&quot;igure 2). For further dermis of l, he merging of all 12 top-level regions of the two ontok)gies, see \[Henschel, 1993\].</Paragraph> <Paragraph position="29"> One peculiarity of the proposed merging is that we do not assume a. strMghfforward correspondence be.tween concepts (especially process types) and sets of surface sentences, fh*tt means, disjoint concepts q'lm two 8(!lltall(,i(; r{2\])l'{;s(2\]ll,a,|,lOllS (;Ol'l'(!s/}on(\] Lo two p~(;lnlilt(!ly aJ(;erllat, iv(! ex\])erien(,ial tmrspecMvcs ol) |;h(~ (2VOIllB) Olle fOCllSillg 11101?{2 oil i|,s ~lc{,iOll like IHttuft, Lhc ot, hcr more on \]l,s r{21ational-\]ik(~ nat\[.'<</Paragraph> </Section> </Section> <Section position="5" start_page="806" end_page="807" type="metho"> <SectionTitle> 3 II,ESULTS AND CONCLUSIONS 3.11 The Merged Upper Model </SectionTitle> <Paragraph position="0"> By N)plying t, he l)rincil)h,.s for merging set (}ut al)ov(!, it was Lmssiblc (,o fully r(H)hc(! both U\])l){!r too(lois r(fiving this intcrfa.<:(,. (,his flexibility is in a.ny (:as0 ~trgued \[or on ol, her grounds ht \[B&tcnta.n, 19921}\].</Paragraph> <Paragraph position="1"> by a shJglc m,.:g('d ultp(w model that diff(.'s v(:ry slighl,ly I'ro.t (;h(~ t'JUM. '\['h(2 Merg{!(I \[/i)l)er Mo(lel C3,II HI f;/,C(, \[}{2 o\])LaJll{~(\] \]'l'Olll l;})(~ \]')~J~/\[ by & s/ll&lI mmtbcr of ad(liLions (8 ucw (X)ll(L(!l)~s (1l\]\[1{\[ \] (2\]1~111<{~(2 o\[ ro\[('. P(2sl;l'i(;~iOltS \[,o a,t/ (!xisl;ing collcel)L). This lac\]C/ of differ(~nc{~ supports tit{2 {:brims (2Ol1(2{2flli11~ mull;Jim gll~t\]il;~,' O~' ('llllCiJiOll&l dcscril)l;io\]|s made i. \[l~tt(miml at M., l{)glb\]. Th(!rc i(, i,~; m'gur.d thai; a, FllllCl;iOllal gI'itlIlIllO+\]' ah'c~My goes beyond slrictly l:tllgllage sit(> cific. (lisi;i.('.tions: the re usa, bilil,y of tim vast ma jority of (;he lqUM orga, niz~{,ion for (a(2rman (lemon sl;ra, t(!s I,\]I;L(, LIlt{I; org~ttli:.';a,l,\[Ol\] is i1()(, Lic(\[ soiely to lBnglish. This is further reinforced by tlm {2XJ)Ol'i(Ul(:(2 {luring (,Ira Ul(~rg(2 |,ha, t wll(2r{*, l,h(! IqU. M e~l,{mdc{I o.</Paragraph> <Paragraph position="2"> disti||cl,ions made i. tim (HIM, (;hose (2:(l, cliMOllS w(erc gmn!rMly equa, lly {qq)licM}le a,u{\] useful I'{}r (;crmall (scc \[\[\]{mschcl, 1993\] for r<dt'.vant examples), 'l'lm result, ()1&quot; our merging i)l'()(:(!(\]lll,(': is; all ollt,oIogy fulfilling the C(}IIS{;PIIC(;iOII i(\[O~tS O\[&quot; \[llovy }tll(\[ Knigh(;, /!).(}8\] in (,hat the I'{~sulting ont,oh)gy c{mtains ~til COll(:c\])ts It(2CO88~l,l!y for (,lie op{'r;:~tk)u of t, lm L'I,;N MAN t\[\]o(hl\](2 a, ll(\[ l,h(! I(()MET module. \[Iowcv0r, i(, (:OllD':-/,(Ii{;Ls t,h(~ lll(!r~ill,~ 10}l{~Ol'y o\]' llovy and t(nigh{, in I;hat. iL sl;ai, es SOlrlC t,heol:(%icaJ priu(:il>l(~s; lbr t, hc )llcl'g(2 COllS(,iqlCl,iol/ which should I)0 mainta.in(~d I)y (,he sou\]'c(2 <)nt, ologics as well ;ts h~ (;he IIIOr~{L</Paragraph> <Section position="1" start_page="806" end_page="807" type="sub_section"> <SectionTitle> 3.2 Merging Statistics </SectionTitle> <Paragraph position="0"> \[~CC~ttlS(! Ot' tJl(;iP qucsl,ionabl(~ st;at;us, wc \[(mvc (,\[~c 'rhetoricM relalJons' out of accoum, in the stal,is (.ical compariso.. WiLhout this l{.Sq'-sul)hi(a'm'chy Dim IqUM includes 252 COllC('.l)l,s. 'l'he (HiM lnakes no i>rccise dis(,iuct, ion between upper and (IotiHtilt |nodel. For (,h('. compm'ison, 235 GUM conc{epgs arc considered. The M{~rged |Jpl)er Mo(hd (:olll.tdns 288 (X) It(;CI){;S.</Paragraph> <Paragraph position="1"> Id(:nf, lty We found I{i7 id(:ntAca.l conccl>l, na.mcs (excluding Chc Ics'r-rclal&ms), from which ouly 87 concepts cm~ ~1 ( ~ally be identiIicd. Identical me,ruing can strongly I>(~ si,al, ed for lOG c.onc(q)i;s (i.e. 19 ha.re distinct Im.lnes). The maiu hlenl;ilh:~tth)u m:cas are the ol)j(:cl; ;t*l(l I;ho (tuMity hiern.rchy as well as I.hc (.etuporM oae. Tin~ precise distribution for strong identical meaning is shown by the numbers on,side of brackets in Figure 3.</Paragraph> <Paragraph position="2"> Union If both considered ontologies are equally weighted as ill \[How and Nirenburg, 1992\], individnal concepts in an ontology must be maintained in any merge.</Paragraph> <Paragraph position="3"> However, in onr approach we have extensively made nse of an ontology:internal concept union. This is a result of the general ontology design principles given in Section 2.3. The ehmse/PP distinction, for example, which is often a concept discrinliliation criterion in the GUM violates Principle 1 and so this <tiscrim: ,nation is not preserved in the Merged Upper Model.</Paragraph> <Paragraph position="4"> Therefore, leaving out of account the clause/PP distinct,on, identical concepts then amount to 163. The number and distribution of concepts identical after union is shown by the numbers in brackets in Figure 3. 106 concepts are strongly identical and 57 merged coneepts are identical with tire unions of dill fcrent GUM con(:epts.</Paragraph> <Paragraph position="5"> Extension Extension can be Pound in both directions. Because of the emphasis we have given to the EUM, most of the extensions are I';UM concepts which extend tile GUM further. These are 60 concepts, 11 for the Mental-Process, ll Participants and 38 others from the Relational-Process hierarchy. On the other hand, only 4 German participant, concepts have found their way into the Merged Upper Model.</Paragraph> <Paragraph position="6"> Cross classification An essentiM field tbr cross classification has been avoided by the relaxation of the upper modelgrammar interface stated in Principle :{ in Section 2.3. For exainple, whereas the cross class,Ifcation (liscussed for the MaleriaI-Proccss/Aclion liierarchy in Section 2.3 would have cross classified 2 Fmglish subconcepts with 5 Gerlnan subconcepts and their subhierarchies respectively, resulting in 42 merged concepts, 9 concepts are suflicient to cover all distinctions expressed in tile EUM and the GUM.</Paragraph> <Paragraph position="7"> Slllillrlary Summarizing the urerging statistics, strong identity can be found for 41%. \[f we allow identification of unified concepts, identity can be stated for 63%. About 25% of' the merged UM are created by extension, and only 3.6% by cross classification. Beside this, there is a small part of t.he Merged Upper Model (8%) where the concepts are not crea.ted by identification, extension and cross classification, but by preferring IgUM concepts over GUM ()lies.</Paragraph> </Section> <Section position="2" start_page="807" end_page="807" type="sub_section"> <SectionTitle> 3.3 Future work </SectionTitle> <Paragraph position="0"> In the current merging proce, ss, we have only h)oked tbr identities and dift~rences between tile given English and German Upper Models. Wc did not try to improve the inherent consistency of both, although it becanie clear during the merge thai, ccrtain distinctions should be removed and others tilrther developed; these local improvements are detailed in \[llensehel, 1993\] and will be incorporated in future vet'sions of tile Merged Upper Model.</Paragraph> <Paragraph position="1"> In addition, one of our tints with the Merged Upper Model is to provide a stable basis for fllrther extension both to include, tim;her linguistic I)he,nolnelaa and to cover further languages. We cx-.</Paragraph> <Paragraph position="2"> pect that an organization of information based on the requirements of natural language gram mars will provide a inert st~d)le and re-ilsable resnlt than of ganizations based on the requirements of individual cornl)utational systems. We are already using the Merged Upper Model as the basis for sentence generation in I)utch and l, here is suggestion here that, again, few additional concepts appear necessary. Of more interest is the extension to rather (lilt>rent lan. guages, some of which has already been begun. Iratailed accounts of this work of extension and comparison are necessary since automatic merging will rarely be possible when these, most, general lewd'Is of information organization are considered.</Paragraph> <Paragraph position="3"> t!'inally, extensions in filtnre may also be made by colrtparison with other ontoh)gies although here it.</Paragraph> <Paragraph position="4"> is necessary to be very (:artful concerning l;he kin(Is of ontoh)gies considered. Since the Merged Upper Model is explicitly a linguisticMly motivated enter ogy, conlparisou witlt outologies with (lifferhig me tivations can be difficult. In considering the ontol.ogy of the IAI,OG l)roject, for example, the nlixt,lre of linguistic and non linguistic information criticized by \[Lang, 1991\] should not be carried over into I;he lnerge.</Paragraph> <Paragraph position="5"> The evaluation of the resulting linguistic ontolo gles as potential semantic type hierarchies for representations in Inachine trailslai:ion, analysis and lnu\] tilingual generation is t\]ieu a eh!ar further step,</Paragraph> </Section> </Section> class="xml-element"></Paper>