A ~,"e'aa~)lework ~bv Lexkat Setecth~n h~ Natm°a~ 
ge~;vei Nire~tburg 
Ca, r~egie,.Mel#m Unive~'i~y 
/~'".~;". Nirenbu~ " 
t :acm~gL~ Grotq, lnc 
A~M~'g~ef~ This t~i~ dc:~cdbcs a proeedmc Ru' lexical se.. 
le=,:!\[~ia~ of OlW, n-claSS k:xicN items in a natmal langp.age 
gca~:~atkm sysRm. A,l optimum R;xical seleetiol; ua}t~nig 
~mlst \[i~.: able tO Ira|nice. rt~.!lizati,'m decisions mldc; vary-. 
i~g rcugextna~ cit'cmastances. First, it must be N)le to 
iy;;,?-,i~ ~, withont the illlhlCUCe of coiitext~ based ou |llCT, ll. 
lug cm~e~poadcuces I~elwcen elemeuls of couccptual ill" 
imt a,,d ~hu k:xic.~l i~e~'emo~y of thc talget language. S(m- 
O~ld, it must t~.~ ~,bh~ to use contextl~d constlaims, as sup° 
F;~f~ by coll~,.:adoaM iaf(~nuatio~ ii~ tfie gcneralion lexb 
c(m. '~ hi~d, ihcru lilllSt \])!) illt oplion O~ realizing input tep- 
resenh liolts laonomiaally or th~ongh detinite dc.scripliuns. 
t,'il!;iii3, the~e nntst alsu b(3 au option of tLsing elliplical 
co~tstruciions. The ~tttum of 1)ackgtouud kuowledgc and 
th{; all o~ithu~ v/c sugge.~t lbr this task are. dcscrilav& The 
icxical xetcctiou prt~dttre iS a part of a couq)fehgnsive 
j~eltcra{ton Systeub D\[OGI,2qE% 
Natnrai iaa;;~tago y, en~:laiion i~, haditk!nafiy divided into two stages: file 
nlt(;ml,'ee plmmiug ('what to. :lay') stal;~: aul! li~e lexical arm ,';ynhactic 
;¢mlizafion (how to say it') stage. 'llle latte_., ~;tage consists, essentially, 
o\[ ;~ larf~,e set ot re::diz:ltion choices fill the vmious nteunings of the iu- 
put, "../!dub th~ morphologk~d, synku;fic sad lexicai tne.'.uls el extm.ssion 
in the bq~gct lal~guage (TL). Research ~c,l×)rted here dc~tls with ~le pro- 
tess (d" lexk~l selection dmiug this s(~ol|d singe of generation. Many 
o'~ the exis|ing generatinu systems ha~,e bees conceived as cornlRlUenls 
cJ" natfl\[~l ~Ilngn~lgc intcrraces to datatmse sysJ~')lnS. In SutAh generators 
the lexical invettto~y can be strongly coustraiued withralt jeopardizing 
the quality of the iatcmctiun (eL, e,g., McKeown, 1985). Such systems 
~meessarily cunccntrate on choosing aptnz)priate "l'l, syutax .- indeed, 
gellCl'a~)rs Roe exp~xiled to produce atkqqllate sy~ltaclic stractqre.s, l,cx ~ 
ieal selcctiovi I~morne* ladle important when it is difficult R) constrain 
Ihe tytw'S of (mtpnt in genea~iiJon, aud, COllseqneutly, wheli the lexicon 
i;*x:t)mes large,. Mactfine hauslalion atul automatit: text summarization 
a~ among q~plicalifms thai by nature, require a wide range of outpuls 
amt iw.ve h) use a sizeable lexicon. Note that t)f these two the \[()i'nlcf 
d(~3S lint involve n|tcratlce plaualng Rlld concentrates on lexical and 
symactic x~:.~lization. 
la ~t~e n Ittmd hmguage genemtiou commmdty the task of lexical 
selcctio~t h;~; uot ye! alh'a,;ied a sufficient atnuunt of attentiua, though 
it was v~ddtesse.~t in a we\]l-know~l early gcncratitnl project (Goldman, 
1975) aud i~ wide\]y nx;og~fized as an imlumant woblem (el. Dan- 
!us, 1994; 3~.~'ot:,S, 1995; Bienkowsld, 19fi6; attd the survcy dimming, 
/986). Oac motivation for lhis ~esenrclt was that we agree with Mat'-. 
cn~ (Iqff/~ p, 211) that 'most gencxaiion systems don't use words at 
alL' a~td we belkwe ~hal the quality of generation ouqmt will improve 
~ignilicauO f on¢c an ad~luate lexical sclccti(a~ contponeat becom,.~ a 
:'taud:-a'd t~ {ff a N\]2.G .';ystcm. 
~'escmch J~c'poi-~eA in tills paper was lXMO~mcd within tile DIOGENES 
"pmj(~l (l':\[h~ ubuig, 19~{7), whose o~jeetivc is to l:,~ovide a high-quality 
gcllcra~:~r %: a knowledg~ol~qs,&l intedingnal machi.c translation sys~, 
tom, 'file in!rot *o this gcut;mtor is a set of a) world concept instance* 
that ce, pa'e,",:>t lhc lhoi)0silional contenl of the original texL attd b) sets 
of mat pa~'ametc; valses \[:hat repre.'mnt its pragmatic content, (Ihese 
cortcepts are repre.seuted in a frame.-orietlWd formalism aw.l are intcr- 
connecteA according to the rules of a special g~'ammar -- sea; NireuOmg 
ct M., 1986 for a detailed dcscdption.) In this paper we deal with a 
subset of the gencralion task, :lamely, the selection of Ol_W,a--class lexical 
items Io realize the meanings of ohject, event aud property tokens in 
tl~; iupuL Thus, ihe output of file geueratiou module described here is 
a lexical unit or a pronram in the target laugnage. 
Ore' appruach (and especially the expected input) to text genaratio. 
is similar to that of the SEMSYN project (e.g. R6snea', 1986). l.exical 
selection is uot, however, an immediate concern of and is not tiiset~ssed 
at auy leugth ia SEMSYN descriptions (see, for iust-mcc,/,aubsch cl aL 
(1984, p. 492), aud a published analysis of pr¢~',tica/ difticnhic~; c> 
cnuntered by the project (tlattakata et al., 1996) d(x:s not addm!'.s 9~i:.i 
issue at all. Furthexnlore, since until very rt~:eutly that prqje~;t had m 
genelate sentencc-leugth texts (article titles), Ihc problem of dc.iiaite 
descriptions, prouominalizalion arid ellipsis did not ~wx:oiae acutely im- 
l×l~mt. 
3 Why is it a difficult task? 
Dextral choice is not a stl~ightlb~ward task, Snppase we hay.': to express 
ill F, nglJsh the meaning 'a person whose sex is lUnle aud wh(}t;c affc is 
between 13 and 15 years.' What knowledge do people use in order to 
come up with an appropriate choice out of such candidate realizations 
as those listed in (1). 
(1) boy, kid, teenager, youth, child, young man, schoolboy, adoles- 
cent, man, 
Without a sentential context file choice, based on eloseuess of i|t~ 
nmauing match and genendity of meaning, should t~e boy. I;or a coul- 
puter program to tm capable of making choices like this, ii Ilas to 
possess a preferenee-assigding capability on the nmtches betweetl tht: 
meaniugs of tile cattdidate lexical l~aliTatkm on the one hand aud tilt', 
input meaning unit (see the discussion of thc matching lUCtlic below). 
3A Collocations 
Lcxical choices am, however, tylfically made in centext. Conlextual 
rclations anloug lcxical units reflect lueauiflg.-iudilced consiraillts ori 
eamccarfence (selectional ~estrictions: admire utkes a human subject). 
Sometimes, however, it is difficult to formulate a cot~:cnrrencc C()ll- 
straint ill terms of selectional rcslriclions alone. Ttnls, 10r cxanqfie, 
tile causative construction wilh the English inJhzence requires eae, t; ils 
Russian equivalent vlijanie requires okazyvat', aed the \]alter is not a 
Rnssian correlate of exert other Itmn ill the above and very R'.w similar 
syntagmatic conslructions. Why do wc use, in Euglish, shed with iea,'s 
or leaves but don't usuaUy say shed water out era bucket or they drop 
tear,*" every time when <...>? Snch properties of the lexical sleek of 
a natural language ate called eolloeational. We will now ilhlsi~tte, Ihe 
c~meept of collouation through several examples, 
Cousidcr thc conceptual Old,rater of a large quantity of, a (relative) 
value liar measuring quantities (of materMs, lknccs, qualities, properties, 
etc.). It is realized in Euglish in sorer&race with collocational propm'ties 
of the lexical units that arc ascxl its its ope.lands. Not every quantity goes 
with every realization of the above opeaattm Members of ~hc set <big, 
enormous, great, high, large, strong, wide> of potential realb,~ltions 
of a large quantity of call reoccur with every of file members of the 
set <amount, difficulty, expanse, selection, voltage> of quantities. We 
say high voltage but a large amount. It wouhl lx~ inapprollriate liar 
a generation system to produce someflfing like high selection or large 
dif3eulty. (Note that in parsing the problem of assigning a similar 
senatntk: marker to all file various expressions from tile e.x~allpl(,; C~tU~ 
in priuciple, be. tackled through a mechanism of metaplt(w pnmcssing 
(e.g., Cat'boneR, 1987), whereby a geueral heuristic rule is dcw;luped 
for processing metaphorical input belouging a single class, such as, for 
instance, a large quantity of... - see Lakoff and Johnson, 1980, 10r an 
extensive listing of potential metaphor classes; in generalion, however, 
the task is the opposite - to produce llucnt nmtaphorical lauguuge. 
Since this depends not on regularities of meaning, but mthcJ on the 
idiosyncrasies of nteaniug realization in the varlons natural langtmges, 
the general rules will be more difficalt to come by and furumlate,) 
An additional class of collocations are the paradigmalic collocao 
lions. "these am best exemplified by the 'set-complement' collocaliuns 
such as the English left and right or parents and chiMren. 'l~le knowi. 
edge of these collocations, for instauee, simplifies the proccss of lcxical 
selection of conjoined constructions, such as ladies and gemlemen. 
Colloeatioual relatious are defined on lexical ualls, not meaning 
~'epresentadons. The study of collocations asceuds to Firdi (195 It; it i:, 
4 711 
a central part of the Meaning - Text school of linguistics --- cf. Mel'~uk, 
19'14; 1981. 'll,e importance of collocational propelzies in generations 
has been recognized (cf. Cummiug, 1986), bat relatively few systems 
actually include collocational information in their decision processes. 
3.2 ElLipsis and Anaphora 
Certain contexts completely alleviate the problem of open-class lexical 
selection. Consider the following (gloss of an) input segment: 
(2) Clause1:BuyOohn3 book7), time|, focus: book7 
Clause2:Bring(John3 book7 office~), belong-to(office~ John3), 
time2: time~ > timeh focus: office| 
Clause3:Read(John3 hook7), aspect: inchoative, time: after(timez) 
One of the adequate ways of realizing it is: 
(3) John bought a book. He brought this book to his office and 
started to read it. 
Ttiefe are seven instances of the three object-type concepts in the 
case-role slots of the input propositions above. Each of the three con- 
cepts is realized lexically only once. In two cases these meanings 
were realized through prononlinalization and in one each through def- 
inite description and ar| elliptic'/ construction. This example shows 
that non-lexical realization is an integral part of the process of lexical 
selection m generation. 
In what fellows we briefly describe the system architecture,, the 
knowledge structures and the algorithm we use for selecting open-class 
lexical items in generation. 
4 The System and the Knowledge 
DIOGF.NI~S is a dislributed natural language generation system featuring 
a blackboard-type control structure. The processing in it is concerto 
teated in the knowledge sources which are triggered by the state of 
the various blackboards. The latter contain the input to generation as 
well as all intermediate and final results of DIOGEN&S operation, rep- 
resented uniformly in a frame-o|icoted knowledge representation lan- 
guage. Background knowledge in 9\[OG~qES includes the following 
components relevant to the task of lexical selection: 
a concept lexicon, a set of knowledge structures that describe ob- 
ject ,qnd eveut-types in the (sub)world of the texts to be generated 
(the fn'st application of DIOG~'qES is; for example, in the domain 
of computer hardware maan~ds) 
o a generation lexicon that F,~ks (sub)world concepts (or, more 
accurately, their instances) with particular lexical units of the 
target Danguage. 
The above description is necessarily incomplete. See Nkanhurg, 
1997 for an extensive specification of all the facets of DIOGrlNES. 
'll,e implementation vehicles for DIOGENRS ate, the Fraraekit lmowl- 
'edge representation hmgaage (Carbonell and Joseph, 1985) and CMU 
CommonLisp running on an IBM PC RT. 
Sample eoueept lexicon entries are illustrated in Figure 1. The figure 
shows a screen of the knowledge acquisition and main~nance system, 
called ONTOS (Nireuburg et al. 1988), which we use for a~quiring 
and mah~taining the lexicons. The figure shows a partial view of the 
concept network and three concept lexicon frame, s cow,responding to the 
concepts of research-workstation, memory and disk. 
qbe following is a sample input tlmt will allow D\[OG~qI~S tO produce 
the sentence 
The basic IBM personal computer XT consists of a system unit and a 
keyboard 
((iI) clause\].) 
(PROPOSITION part-ofl) 
(MODALITY real) 
(SUBWORLO computer-world) 
((SPEECH-ACT definition) 
(DIRECT? us) 
(SPEAKER author) 
(HEARER reader)) 
(EOC\[JS 
(GIVEN role\].) 
(NEW (and role2 role3)))) 
((PROPOSITION part-ofl) 
(IS--TOKEN-OF part-of) 
(PATIENT role\].) 
(COUNTERPATIENT ((and role2 role3)) 
(ASPECT 
(P~ASE begin-end) 
(DURATION always))) 
Graph ~ro~aer: DEVICE elol~m 
r~e Edit: RESE(~RCH-IdORKSTDTIOH|illglKt C| t~o" CIRa#: Lg~e Edlt.I RES\[ ................................................ ~ I~V 
i IS-41 ( CCHPtD'ER- CO~P~WIEHT) INkiEr-oF" (COHpIff£R) 
iZS-O (PEROOHOL) ~,\[EICLASSE$ (DISK) iSUBCLASSES (IBH-PFJRT) DEFINITION ("WORKIHG STORAGE FOR~ 
~DEVIHXTION ("hl=h p~=rmd") THE CPU") !I.EHGTH (0.5 TO 1.5) Ill|= ~z~to4 I~= ~lcJ~-o~: 
\[UIOTH (0.1 TO 0,5) LEla£,TO 0.1 "re 1000000) \[HEIGHT (0.1 TO 0.5) WIDTH (o.1 TO 1000000) 
OPERATIHG-SYST£M (UHIX V~4S 05 HSDOS ÷ HEIG~Ff (O.i TO 1000000) ! 1= ~CPH HGCW) MASS (O.t TO 1000050) 
! • t~ Ix~=Jltu~ f~ COW~%hr~: ~HAPE IL 
JlIAS-SS-PNRT (DISK CPU HEMORY) COLOR on. i ¢\]=t= l~LI~OI£taa era= I'II3~IP~--J~L-OL~'~1~: ; ASK o,1 TO 1000000) 
iCS~L~PR E HH~L L ,I Cc4~annd) 
(o.: TO I000000) j 
IPQRT-OF" ODJECT) I iI~ELOHGS-TO (CREATURE DRGAHIZATIOH) I 
igo~and> Top-: e~ 
i co.~.";d7 d 
,teme c;a;~: DI~( 
iPARY~OF (COHPUTER) !IS-A ( C f~4PUTE R-PERI PIIERAL 5¢ 
MEMORY) IIEFIHITIOR ("Secondaru wtoras~") 
alota ~,~ltel f~w ¢~\]NtE~-~IF~| ~: 
~UBCLASSES (DISK I/O-DEVI CE) list| ~i~tea from ~IC~-OlJ~: 
.EUGTH (0,I TO 1o00OO0) WIDTH (0.1 TO 1000000; 
~I:FN" (0,I TO I000000) H05S (0.1 TO 1000000) 
SHfmPE NIL COLOR HIL 
AGE (0 i TO I000000) 
Top-level c~and dlm10g 
=r.eh devloe 
Figure I. Concept Lexicon Entries 
472 
((II) role\]) 
(EXTENS IOJ~ IBM-PCI) 
;an object instance to which all the various 
;descriptions (inters\]ors} of it refer; 
;for example, "John Sril\]th" can be intensionally 
;represented as "John," "Mr. Smith," or "Jim's 
; father" ;-- but it will refer to the same 
;extension 
(INTENSION 
(IS-TOKEN-OF IBM-PC) 
(QUANTIFIER un\] versal) 
(SUBWORLO computer-world) 
(MODEL XT) 
(CONFIGURATION minimal) ) ) 
;"basic" r~Leans "minimal" ,~et of components 
;that ca~\] be called a PC; the best way of 
;treating this is to define, in the onto\]ogy, an 
;attrlbut6 "configuration" whose domain will be 
; (car house computer ...) -- anything that has a 
;basic pzJce and extras, and whose range will be, 
;fo~ the time being, (minimal regular extra) 
( (ID role2) 
(F.XTENS ION comput er- syst era-unit I ) 
( \]: NTENS \]:(IN 
(IS-TOKEN-OF computer-system-unit) 
(QUANTIFIER unlversal) 
(PART-OV IBM-PC model - XT) 
;one needs \[:his tautology; otherwise, system 
;units have to be concept NAMES in the 
;ontology. note the binding for the "mode\]." 
;which does the same compositionally, without 
;pro\] ifcrating names 
(SUBWOR\]JD computer-world) ) ) 
((ID role I) 
(EXTENSION computer-keyboard\]) 
( INTF.NS II)N 
(I S-TOKI£N -OF computer-keyboard) 
(QUANT I},'I ER universal) 
{PART-Oi '~ IBM-PC model - XT} 
(!;UBWOR%D computer-world} ) ) 
40~ The (~e~cration Lexicon 
The main static knowledge source for generating of open-class items 
is a specialized generation lexicon (GL). The structure of an entry i. 
the g~neral\]ou lexicon in DIOGENES is shown in Figure 2 (the BNF is 
incomplete wherever obvions): 
The importance value serves to ~stingulsh the sal~ney of the var- 
ious relations for the idenfl~ of the entry head. Thus, f~ iustance; 
generating youth instead of boy seems to be less a deviation than gen- 
erating gld. Th~ is why the importance of the sex slot in the example 
below is grea~r than that of the age s~t 
The sample GL entries below do not contain a full complement of 
collocation re~fions. 
(make-frame toss 
(is-token-of (value throw)) 
(direction (value up) 
(importance 3)) 
(altitude (value high) 
(importance 3)) 
(velocity (value low) 
(importance 9)) 
(object (value coin) 
(lexeme (value "toss")) 
(syntactic-info 
(lexical-class verb) 
(verb-type transitive) 
(morph regular) 
(para-collocation 
(antonym catch) 
(synonym cast propel toss 
fling hurl pitch pass))) 
(make-frame new 
(is-token-of (value age.CL)) 
(age (percent-of-range (0 25))) 
(domain (value non-living.CL)) 
(lexeme (value "new")) 
(syntactic-info (lexical-class adjective)) 
(morphologieal-info (comparative regular) 
(superlative regular)) 
(para-~collocation (antonym old))) 
(make-frame boy 
(is-token-of (value person. CL)) 
(sex (value male) 
(importance i0)) 
(age (value (2 15)) 
(importance 4)) 
(lexeme (value "boy")) 
(para-collocation (synonym lad kid child) 
(antonym girl adult) 
(hypernym person)) 
(syn-collocations-in (value boy.syn))) 
(make-frame boy.syn 
(agent-of (value play throw run jump) 
(strengh 0)) 
(place (value school playground ballfield) 
(strength 0))) 
GL-entry ::= ( <meaning-pattern><TL-pattern>* ) 
<meaning-pattern> ::= ( (token-of (value <CL-concept>)) 
\[( <relation> (value <value>*) 
(importance <importance-value>))\]* ) 
<CL-concept> ::= {any concept in concept lexicon} 
<relation> ::~ {any relation from Concept Lexicon} 
<value> ::= (any concept or attribute (scale) 
value in Concept Lexicon} 
<importance-value>::~ 1 l 2 l .-. l I0 
<TL-patte~n> ::= (<TL-lexeme><lex-info><collocation>) 
<TL-lexeme> ::= (<language>TL-lexical-unit I (synonym TL-lexical-unit*)) 
<language> ::= english I spanish \[ Japanese I ... 
<lex-info> ::~ ((<syntactic-info>) (morph <inflection-type>)) 
<syntactic-info> ::= {the contents of a syntactic dictionary 
(cf. e.g. Ingria, 1987)} 
<inflection-type> ::= (an indication of irregularities in forming word forms, 
e.g., @i\[goose\] - pl. @i\[geese\]} 
<collocation> ::= ( (<dimension> <dimension-value>*}* ) 
<dimension> ::= {the name of a (syntagmatic or paradigmatic) 
collocation relation based on the CL slot names 
for the concept in question} 
<dimension-value> ::= {a TL lexical unit (word or expression) 
that can ordinarily collocate with the 
TL lexical unit in <TL-lexeme> above 
and connected to the TL unit on a 
specified dimension; een be recursive} 
Figure 2. ~llle Structure of the Generation Lexicon. 
473 
5 The Algorithm 
in the DIOGENES generator art inslar~tiation of a head-selecfing knowl.- 
edge source is triggered simultaneously for every event ,and role instance 
in the input representation. The results of their operafiou are posted to a 
public blackboard, so that ail knowledge source instances can draw on 
Ihis knowledge in their own decisiou processes. The knowledge sources 
responsible for selecting modifiers are triggered when the heads of their 
phrases have already been selected. 
Figut~ 3 illustrates dm algorithm tora single lexical selection (head 
or modifier) knowledge source. If an input frame was already menlo 
tonext in the input, the question arises whether it should be realized 
non-lcxically, that is, using deictic means (this is the case with the seco 
ond appeanmce of John in (2)). If so, a proper realization must be, 
found and posted on the corresponding blackboard. If this process fails 
at arty point, We revert to the 'regular' case of lexical realization. This 
latter consists, first of all, in scanning the generation lexicon in semeh 
of a set of candidate realizations for the input frame. (1) is an example 
of such a set. When such a set is produced, we attempt to filter it 
by removhlg those candidates that are not compatible with realizations 
already decided upon for other input fi'ames in the same sentence. This 
processing is based on comparing the collocation infonnation in the 
lexicon entries for the members of various candidate realization seLq. 
F~r example, if a neighbor frame has already been realizexl as demon- 
strator, then the collocational intbrmation will filter out all mcmlxa's 
of (1) but youth, teenager, man. If the residual set has cardinality one, 
we post the result. Otherwise -- as in the case when no collocationnl 
intbrmatkm cnn be used -- we pmcexxl to select the ~v~aliz~tion based 
solely on the entries in the cnndidate realization set (ttmt is, without the 
ILT / First ~. No j No- ~c 'ca ~-~. 
.......... ~" reference to ~'~--- ~.~"~ realization ~-~ frame 
~..concept ~ ..~. ..... )nd,cated../- 
--Choose -- I / 
Produce \[ ..:J non-lexical 
candidate 1~ realization 
realization \] l 
set \] ~..~ .... 
+o 
Process 
collocational 
constraints 
element 
filtered 
in choice 
above 
threshold? 
Yes 
context-independent of match Select 
meaning matching higher than best 
metric threshold? . match 
Post 
result 
Figure 3. A procedure for selecting open-class lexicaI items during 
text generation. Incorporates the capability to intrcxhice anaphora and 
ellipsis. "rakes into account collocational knowledge for pimlucing con- 
textually appropriate realizatious. 
Augment 
Generation 
Lexicon 
474 
heaefit of a context). This routine uses a well-defined inexact'matching 
metric that e~deulates distances between the meaning of the input franle 
and the raea,fings of the lexieai units in the candidate realizution set. 
The closest meaning is then selected and posted. 
6 Stattas and Future Work 
The blackboard architecture and the inexact meaning matching mod- 
ule luts been implemented; the collocation treatment module has also 
been implemented, but extensive testing has not been performed due 
to the lack of a lm.ge-seale lexicon. The anaphora treatment module 
has been implemented for pronominalization only, and the nnmber of 
pronominali,~ation rules employed has to be and will be increased. 
It is chair that the acquisition of the generation lexicon is a major 
and extremely labor-intensive task in natural language genelution. The 
acquisition of this dictionary, especially of the colloeafional information 
cannot at present be done automatically. But the efficiency of the team 
of human lcxicograplters working on this problem can be increased 
dramatically through the use of specialized intelligent interactive aids. 
We have developed one such Knowledge Base Maintenance System 
(el. Nirenbtwg et al., 1987) for rite acquisition of concept lexicons and 
will extend it so that it becomes applicable to the task of acquiring 
genernlJon ~exicons as well. 
Acknowledgements 
The authors would like to thank Victor Raskin, James Pestejovsky, 
Rita McCardcll, Cad Pollard, Eric Nyberg, Scott Hufflnann, and Ed 
Kensclmft lbr fruitful discussions of the topic. 

References

Bienkowski, M.A. 1986. A Computational Model for Extemporaneous 
Elaboratkms. CSL Report 1, Cognitive Science Laboratory, Princeton 
University. 

Carbonell, J. and R. Josepb. 1985. FRAMEKIT, A Reference Manual. 
Department of Computer Science. Carnegie-Mellon University. 

Cmnmmg, S. 1986. The Lexicon in Text Genelution. Paper pre° 
sent¢~l at ~.be LSA Linguistics Institute Workshop on Lexicon, New 
York. 

Dmdos, L. 1984. Conceptual and Linguistic Decisions in Gene~a- 
lion. In: Proceedings of COLING-84, pp. 501-504. 

Firth, J.R. 1951. Modes of Meaning. In: J.R. Firth, Papers ill 
Linguistics, London. 

Goldman, N. 1975. Conceptual Generation. In: R. Schank (ed.), 
Conceptual Information Processing. Amsterdam North Holland, pp. 289-3~2. 

I~anakata, K., A. Lesniewski, S. Yokoyama. 1986. Semantic-Based 
Generation of Japanese-Germml Translation System. Proceedings of 
COLING-86. Bonn, pp. 560-562. 

Hovy, E. "Integrating Text Planning and Production in Generation," 
in Proceedings of IJCAI-85, l.os Angeles, 1985. 

Ingria, R. 1987. Lexical Information for Parsing Systems: t'oints 
of Convergence ,and Divergence. hu D. W~ker, A. Zaml~olii and N. 
Carzolari (errs.) Automating the Lexicon: Research and Pr.'~ctice in 
a MultiHngaal EnvirmLment. (in print). 

Jacobs, P. 1985. A knowlexlge-based approach to language produc- 
tion. Ph.D. dissertation, University of California at Berkeley. 

Laubsch, J., D. RiSsner, K. l-lanakata, and A. Lesniewski. 1984. 
Language Generation from Conceptual Structure: Syuthesis of Ger- 
man in a Japanese/Genuan MT Project. Proceedings of COLtNG-84. 
Stanford. pp. 491-494. 

McKeown, K. 1985. Text Generatiml. Canlbridge University 
Press. 

Mel'~uk, I.A. 1974. Towards a Theory of Linguistic Models of 
the Memllng-Text 'l~,pe. Moscow: Nauka. 

Mel'~uk, I.A. 1981. Meaning - Text Models: A Recent Trend iu 
Soviet Linguistics. Annual Review of Anthropology. VoL 10, pp. 
27-62. 

Nirenburg, S., V. Raskin and A. Tucker. 1986, Ou Knowledge- 
Based Machuie Translation. Proceediugs of COLING-86. Bonn, pp. 
627-632. 

Nironburg, S. 1987. A Distributed System for Language Genera- 
tion. Technical Report CMU-CMT-86-102. Carnegie-Mellon Univer- 
sity. May. 

Ni~'enburg, S., I. Monarch, M. Calvin and T. Kauimaan. 1988. 
ONTOS: A Knowledge Acquisition and Maintenance System. CMU- 
CMT Internal Memo. 

R6sner, D. 1986. When Mariko Talks to Siegfried. Proceedings of 
COLING-86. Bonn, pp. 652-654. 
