American Journal of Computational Linguistics 
PROCEEDINGS 
13TH AWNUAL MEETING 
ASSOCIATION FOR COMPUTATIONAL LINGUISTI cs 
Timothy C. Diller, Editor 
Sperry-Univac 
St Paul, Minnesota 55101 
Microfi che 34 
Coovri$.ht @ 1975 by the Association for Computational Linguistics 
PREFACE 
The talks presented in Session 3 of the 13th ACL rneet- 
ing dealt with three facets of the language understafiding 
process: the integration of knowledge sources   axt ton and 
A. Robinson), syntactic processing (Bates and J Robinson), 
md semantic processing (Hendrix, Sondheimer and Perry, and 
Cercone). The first four talks were oriented toward speech 
nput while the talk by Sondhe~er and Perry assumed phrase 
structure trees as input and the talk by Cercone assumed 
textual input. The paper from which Bates's talk was drawn 
was much too extensive for inclusion bere and will be pub- 
lished separately My special thanks to Stan Petrick who 
agreed to chair this session on very short notice. 
--Timothy C. Diller 
Program Committee Chairman 
TABLE OF CONTENTS 
Syntactic Processing in the BBN Speech Understanding 
Systen! Madeline Bates ................ 
4 
System Integration and Control in a Speech Understanding 
SvbteEi William H. Paxton and Ann El Robinson ........ 5 
A Tuneable Perf~rm&~ce Grmraar Jane J. Robinson ..... 19 
Semantic Processing for Speech Understanding Gary G. 
....................... 
Hendrix 34 
SPS: A Formalism for Semantic interpretatton and its 
Use in Processing Prepositions that Refeience Space 
Norman K. Sondheimer and Doyt Perry ............. 49 
The Nature and Computational. Use of a Meaning Represen- 
tation for Word C~ncept s Nick Cercone .......... 64 
!375 ACL Meeting 
SYNTACTIC PROCESSING IN THE BBN SPEECH UNDERSTANDING SYSTEM 
MADELINE BATES 
Bolt Belranek and Newman Inc. 
The syntactic analysis system presented here is composed of 
~vo parts, a modified augmented rransition network grammar and a 
7arser which is designed for a speech understanding environment. 
The parser operates on partial utterances called th&ries. 
A t3eory may be thought of as a set of words which are hypothe? 
sized to be in the uttezance. The parser processes the words in 
a theory by building partial syntactic paths using the words of 
:he theory. These paths do not depend on left context, which will 
Le missing if there are gaps in the theory. Syntact-ic constitu- 
exs are built where possible and, whenever a constituent is 
it the parser can interface witrh the semantic zomponent of 
 he total. s~eech understanditzg system for guidance and verifica- 
tion. 
The parser tries to predict words and/or syntactic categories 
50 fill or reduce gaps in the theory, particulatly small function 
riorls which are difficult to detect reliably on acoustic grounds 
aloone. Thk parser does not follow all possible parse paths, but 
nctc~;:~pts to select the.nost likely ones fo; extension. It uses a 
'-:dicious mixture of top down, bottom up, depth first, and breadth 
J 
" I 
rlrst parsing strategies to take advantage of local, reliable in- 
forrLstion. It saves all the information gained while following 
al~crnative parse paths, so that several parse paths which share 
a comnon part, even if the paths are in different theories, can 
stare chat portion without ~eparsing. This is true even if the 
-.rse r/ '.A paths split before and/or after the common part and even if 
the common section analyzes only part of a syntactic constituent. 
American journd of Computational Linguistics 
Microfiche 34 : 5 
WILLIAM H, PAXTON AND ANN E ROBINSON 
Artificial Intelligence Center 
Stanfoyd Research Institute 
Menlo Park, California 94025 
ABSTRACT 
Two impartant Problaq8 in Speech understandiqg arc haw to 
afteotlvtly integrate multiple &ourcts of K~ow1edg.e within the 
8YSttm and haw ta control tht! ~ctlvltles sE tDe SYete~ to arrive 
at approprirta LntetpretatLbnr tor uttcrsnces, Thir paper first 
d@&crlbc8 the roles plrytd by acoustics, spnter, ~ccantlqs, and 
dlIcOUT8e1 and Shows haw a lanQuaOa definition is Used to 
lntegrrtc thcr lnta a rysten in a way that allows the 
interactienr 'to be easily virialtt The second part of the paper 
derctlber an exteutivt that uca~ intormaf ion from there knowledge 
8aurctr in its control strategy, 
Acknowledgment 
Thll rrrcrrch was rupportrd by the Dctcnsa Advanced Rasearch .- 
Pr0je~tl Aucney of the Dcprrtmcnt of Defense and monitored by the 
U,6@ Army Hcrerrch Otficc under Contract NO, DAhC04-75-C- 006, 
A Speech underotandlng system muse use many kind5 Of 
kn~wlcdg@r each Playing a particular role -elurlmsg the 
interprctatlen sf an utterance, #hi le thede rolcrj are 
intcrrclatad, it Is lmportant to oc agle to separate the 
kn0wIedge sources so that InterPelatlans arc vlslbbe and so that 
the cantrlbueisns from the variaus sQurecs can be studiedo The 
knowledge sources uged in tht system belnq devclo~ed jointly by 
SPJ and SBC can bc characterized broadly under tns headings of 
860~6tic~~ syntaxI semantlcsp and QiscQusse (wslkcr ?t alQp 19951 
Pablnban, 18753 Hendrlx, 19751 Deutgch, 19758  loc cum, 19751 
Pitea, 19751, 
Thc acoustic component relates 1f-nqu1~tbc @nFithe~ ('#~rd$ 
and phrases) to the s~gcch wavefst~, An acsus~%c-~hon@t$c 
processor enalyzcs tnc digitized waveform t~ extract paramctcrj 
based on Spcceh ~r~dU@tl~n characteristicsa The Darametefs 
ihcludc fundamentab Lgequency~ volein3 Labelo fQrfl&nt frequencyp 
energy data, and athers, FollowinQ parameterlzatlan, varis~us 
rules @a@ ~ppbtard to gepa~ratc BPI ac0~3tf.c feature d~5e~:l~ti~n of 
the uttcranct, The parameter6 and ksatures are ~ubs@quen?ly used 
by the Laxlcel m~plpiglca proccdiaza,  he mapp@r 1s called fILarhci 
the parsing of an utterance to q$ve a. deed~icsn score as to 
whether a progored word or phrase could actl~ally he prqaent in t~ 
sPaci f Icd time region ka l %PAC Input, Phonoloqlcal ~nd 
&~a~stle=phbn~tlt rules are used by the mapper to r~lqt~ nhonsttc 
spelling$ ta acauetic AaEa, 
Syntu provides rclirblc~ reasonably incxpenliv~ lndlcatlanr 
of ikleh words or groups of words may ca~bine and of ho* xcil 
thty fit, Syntactic ruler plve general psttrrns for conrtructlng 
noun phraser, claustrr and scntencgs a?d ~ro~ide censlstency 
eRccks tor such iterr as nurrber agreement, In testlr:q word or 
Phrarc c~mbinati6nSt syntactic Infarmatior? alone can often rule 
Out a candidate without the I7ecd tor nore costly serantic ana 
dttcourac analySlfir 
Tht semrntlc component includes a general p~del of- the 
domain of dlgeaurser and e set of a2gorithvg for ca~biniqg (ox 
tcjcctlngl concepts in the domain, For txrmplct given a verb apd 
two noun phraser, semantic routines can build the corresponding 
@ctMintie ralation between the Items indicated by the noun 
phrases, 
The df~course component deals with the pelattanship of the 
current utterance (or a portion of Lt) ta the dialog context and 
ta entfties In the task dsmain, DLscourse functions U6e 
infarmatlon fron prtvious uttcrancaa to fill out elliptical 
sxpresslons and to find referent8 far Pronouns and derrnite noun 
P~TL~~CI I 
The Irhguage detlnltion is the foe41 paint for Integrating 
the#@ kn~wl~d~a sources, A lanquage definition Includes (1) set8 
of Units out of which Uttafanesr in the languaue arc csnotrbctcd 
and (2) rult8 for combining the units into lar~cr structure@, 
The bcrit unit8 will be cdll@d 'wordsc (although th16 technicel 
use deer not exactly correspond to the common use), The 
camPorttion rub66 indicate how Phraseg Can be combined into otill 
l~rgar phta~c~~ Mere ~~@c~scZY~ a, phrarcp Ir 4lther a word in 
the input ox thc result of 8pplylnP a composition rule to 
constituent phPe@cBc Tne rules give the Lldtar Pattern ot 
constituents and .rpeclfieations feu calculating values for bath 
the attrtbutel of the tcrulflnp Phrase and i~r factors used in 
judging the rasult, 
ft its at the phraa~ lave1 that the knowledge sources are 
tntrqrattd into the 8ystrm, Thcrc are two aspects to the 
~OntflbutlDns trom eren sourcel the values df properties of the 
phrah ar computed by the knowledge aourqe, and tne saurcec8 
arr@rrmsnt of the cart6ct:naser or th1,lo phrase a8 an interpretation 
of the inpot, There two aopectr are reflected in the attribute 
and factor statement8 that ,are assotlattd with Bach of the words 
and Phrase, in the lrhguagr d@tini&lonr Thr atrfibutc statements 
provide lnatryetlonr for eomputlnq var lour prapcrticc of the 
Phra*a, There inrtructiahr may call Won any or all of the 
oaurccs of knouledua, ~sr axampla for a phfdse spanning a 
particular ragmarit, an aenustic &ttrlbuta may specify the ward8 
in that sagrnantl an attribute Supplied by the syntax ean sPacleY 
r farture rueh rr thr valce (*retlvrP or Qgaaol~cC)~ an attributa 
Lbpplled by remrnticr errn 'Ipecify r semantic net kdterpretatton 
built trom the rernantfcr of the con@tituantap and an attribute 
rupplied by the die~ourae componant can indicate a tofercnt or an 
impLLed meaning, 
P8cfor strtltlntntr tall how to u.r@ these attrlbutrr in 
dWm!tlnihg the likelihood that the Phrarc it corrret 
$ntcrpretatlon ot thr input* The result of eosblnlnq the factors 
tor u prrtlcul&r phrilt 1 eJlcd a reorc, Thr urc of such 
gcortr by the executive in determlntng averall itrateqy fl 
described bt10~1 Fact098 are nonbinaryf tinca they can have a 
range of valuer, rLgid *yere or 'no* dt~ist~nb 64 not have tc be 
lad@ in esrersgng the qUalltY of a Pht.#8ee For axa@Ple, the 
cloransre of the acourtlc match asy vary @nu this CB~ be 
rrtlecttd 1 the corrctPondlng factor, weak evidence from one 
rource of knowZtdgc could lower the rcora, r hi la strong evidence 
from amther 80urch cwld c6mpen8ata for that and actually raise 
the SCO~-4, 
In summatyr a phtare is 4 eompo~lte intarpratation of a 
~articu1.r portion of tne utterrnei, inttgr~tinp contributLans 
from a11 ialavant knawledg@ oaur~t~ r This mean8 that each 
pertian ot the lnput is lntt'rpreted and evaluated by the system 
ar fully a@ P60Sfb&~r as 800n 41 po~~lble , Tne 8y8tw is navcr 
faced with the problem of rtXatlnp ar eomblninp freqmvntary 
thearier conlttucted Inde~tndantl~ bY dif fr~rent knowlrdus 
8eurcl)s~ and cvrluarlanr mrdr by differant sources are 
Lnirdlrtely mrrqcd to e~ntrol and CoOrdlnrte ovrr@ll ryatcm 
activity, For QXI~PICI a8 goon &f r dafinlre noun Phr&I@ i8 
found# the acourfic component chcckr thr corrttculatlan of the 
eanrtltuentr, the ryntretic component check@ for agr~ement in 
trrturer 6~eh ar number, the rcmmtic component build8 II 
r~prasrntrtlon of the meaning, and the direourre component looks 
far r referent, 
The fallmin~ tx@apke lllutttatQp haw Several knowledg@ 
~b~tces ate used together to htet~tet and evaluate phraree, The 
Submarine.* or 'their submarinar@ and t1l;urtrates t integration 
QS acauotret ryntactle~ remantic@ and dlrcourre Lnformarfon, 
RULEa~CI NP9 NP d DET NOM) 
8TRfKG 8 APPENQISTRINCIBET]r~TPING(NQY])I 
NBR r CfNTCR8ECT(HBR(DET]tM0R[NUM)), 
CMU @ GINTERBECT [CMU (DET) ,CHU (NOM) 1, 
IEMARTZC6 $EMCA&L(fl$EMMRNP7,SEHANTtCS(NOM]o 
MOOO(D~T),CCAS~(O~T)~~NTERPRETATION~DET~), 
DIlCOURSE 8 It HOOD[DtT) EQ VOEC THEN 
DIICALL(~~DIBRWP~,$&MAN~:ICS) ELBY ~UUND~EPXNEV, 
IMTtRPRETATXON t IF DI8C0URSE bU HUNDEFINtD THEN 
DIlCOURIE ELSE SEMIHTXCBt 
COIR'T a MAPPER(L~~TWOFD(DET),VIR~TWORB~~QM)), 
NBP 8 IP NULLXWBR] THEN OUT EtbK b~, 
CMU 8 IF NuLL(cnu1 THEN OUT etaE OK, 
6tMAHTIC8 + IF NULL(GEMANTIC$).THEN OUT 6L8L OK, 
DXecouPsE xr MOQR(DET) NQ RDEC THEN OK ELSE 
Ir NULL(D~1ICOURLE) THEN POOR ELSE 
IQ AHBICUOU&(DIICOUR~C) THEN OK ELSE GOOD? 
rr8ultrnt phrralb, which In an acoustic attri.bute indicating the 
ward8 campori~g thir phrwe, NBR (nllmber) and CMU 
(ca~ntmmr88-Unit) axe UynZ&ctlc rEttibUtgS tar the Phrraa. each 
being d~rlvsd from the Intar8rctlon af the eorra$pondinp 
rttrlbutea of the con&tltuanrrr  he ravrnticr attribute is a 
piece of semantic net that is constructed from tnt remantles of 
the canrtttucntr by the stmantle routine (SkMRNP7) araociated 
wlth this rula, If the MOOD att-ributa of the DET can~tituent is 
'DEEC, 1 declarative determiner, then the dlscautse rautincs 
wilL leak for a referent for the Phrase in the dialog context and 
8srlQh its ramdntlc structure a$ the value of the attribute 
PXSCOURSE, The INTERPRETATION of the phrase 16 efther the 
referent found by dl~c~urrt or the semantic net structure In case 
no direct reference is found@ 
The frctoi statements use these attributes in carnp~tfns 
cantrtbutianr tawardr the seara for the Phrase, As ha$ been 
aentloned~ there fc 4 rmpe of aecaptaBlc valuer for factorr, 
Par rimpllcity, Symbolie valuer are used (VEFYCOOP, GOOD, OL, 
POOR# BADt and OUT], Zn the cxamPlt ruler there are factors 
deterrnlned by saeh af the major knowledge sources, The CaART 
factor reflects rp acoustic :crt of the coartlculation of the 
kart w~zd of the clet~~mintf and the first word of the nominail 
NeR and CMU arc ryntictlc factor8 that will eliminate the phrase 
it either attribut4 1% ineomprtlblc bttvccn the conrtitucnts. 
The ranantic factor *ill rllminate the PhFrgc if no remantie 
Intrrpt+tstlan can be tarmulatad, While the current remantic 
camgonsnt bat6 not hrva a metric for d@thrmining the likelihood 
of @n intcrpretdbtlon other th.n whether or not a Semantic 
rsprertntation can be butit, it ir pos~ible to intmducc auch a 
nrwic and hrvt the irmrnt%c factor# be nonbinrry, The discourse 
trctor i& nonbinary, ~f the determiner 1s declarative, the 
dlseourse h&8 tried to find a referent, If no referent War 
faund, the factor is given r low value, *POORpr but the phrase is 
not diacsrded, If scvcrrl porrlbls referents were found, the 
phF88e 16 kept and the score is not lowered because the ambiguity 
can perhags be resolved later, It just one referent war found, 
It is taken ao evidence that the phraoa 1s a correct 
interpretati-an for t,hat Porttan of the Utterance and the factor" 
is given a higher value 'GOOR@, 
The example discusead i#bOv@ shows how the langua~,~ 
definition sygtom can be used to integrate a variety of knowledge 
rourccr in a Wry that keeps the cantributlonr and interactions of 
the ditf went sautcrr earily vis&bltr Tns teprercntatlon 
com~lncr procedural Infatmetion (in the exprtsliana for 
c~lculating attxlbutt and factor Values1 end declarative 
inforrnrtlan (in the constituent pattern) in a form designed to 
rimplify thc task of writing a large! definition containing many 
rules, Howrvrr, brfore fhc ruicc can actually ae Usrd, thr~ muat 
bcs convert cd to a ditfcrcnt rapraoenration d~eigned with 
afficdency In mind, This tWhn8lati0n is dona by s language 
definition compile'l:' that cemtructr an internal rcprcrentation 
of the language definition that depend% in an intricate way on 
the rtructura of the "a~scut~~c~~ the portion of the $YStem 
g!@#pOnlibLI for rchedultnp and eontrolling the various tarks to 
be Perfarmad in conrtructlnp an Inter~ratation of an utterance, 
The operation of the exrcutiva is the subject of the rest of thir 
The axhcutivt maker a dirtlnc'tlon between tho phrases bclnp 
built and the tarlcr rkquircd to build these phrases, A data 
StrUetUrer called the 'parre net', tdpreeants the growing 
collcctlon of phrase#, snd another structure, called the 'task 
qUeut@, Cnc0de8 the rltcrnrtlve operrti*nt available for taklng 
another Step toward Understanding the input, Each enttv in the 
tarK queue specltier r proccdurt to be performed at a particular 
locatton (node) In the parse net, The ~e%torrnancc ot such a 
proctd~~C typically entails both modifying the parse net and 
Seh@dUlLng naw task8 to makc further rnodltlcattons, Each task 
has rssocl~ted with it r priority for pctf"ormlng It. The method 
for determining priaritier is described below, 
Taskr can include lpaklng for a new word or phrase to finish 
an 1ncomBlett Phraoa [one mtssfng some of itr constitutnts] and 
frying to ure r word or completed phrase in a larger phrase, 
This men8 that the ryrtem can work both 'tap downr and 'bottom 
upel btcaUlt it can look in a goal-driven manner for mlrring 
eonstltutnts of hlgher level phraiest and it alto can accept 
word8 from the aceurtlcr to bulld into lrrgar phrbrcr in r 
data-driven mannerr AS an exampkc, COnsidtt the simple grammar 
S 8 NP VP 
VP F VP NP I VERB 
VERB s awn I last 
NP a thty I the house I It 
Assume that,the word @theyo has been7 found initLally either 
by the acoustics directly or as a result of confirming a 
Dredietlan made by The language ?ertnitLon, .They4 constitutes a 
compltte NP, This NP cah be gut into the S rule, cau$ing the 
partially filled phsase 'they VP* to be added to the parse net, 
Alread~~ Some of the attributes and factors tor the S rule can be 
drtermlned, and r score computed for this phrase, Bullding this 
partial phrase loads to the creation of a ntw task4 to look for a 
VP falloWLng the hP, That task in turn leads to two alternative 
8ubtagK~I Laak for a VP NP or look far a VERB, PriarLtLrs far 
both these tasks are computed and they are put on the ta$k queue 
to be ProC@$Sadr Tho executive then removes the next task frov 
the queue and continues, 
In Penepalc decldin$ which task to perform is of great 
Importance, bccaure only a subset ot the scncduLed taws will 
actUallY Prove ro be neeassarY to Understand the ingut; the 
ethers will be *false steps* leading to dead anqs, ~deailyr in 
dcclding wh-ich task to do1 the executive would always Choose one 
of the ntecsoary fa8ks and never take a false step, The 
utterance would ba undarrtoad with the unnecessary tasks still 
left $h the Queue, Ta aP~rgech this idealr thc actual system 
must spend rome nf its effort In chodaing tasks, Such offorg is 
well spent if it producca a net decrease in processing time, In 
other Wordrr the etticitncY of the system will b~ 1rnPcoVed by 
deci$ionn regarding the order in which tasks are perfotmed, if 
the cost a!! the decirlons la loss than the coat of the false 
rttpr that wauld otherwise hsvc been taken, Since ac+ouatic 
uncertainty in $patch undersfanding makes tnt potontlai for 
rartlng effort en unnecessary operattons particularly large, the 
system can afford to carry ,out rather complex computations In 
deciding what to do next and Still obtain a large imrovcment in 
overa$l efffcltncy. In the current system, the decisions are 
based on the rclativt priorities assigned to the varlous tasks 
Waiting In the queue, Taska ere aSS*oclated with phtaser, and 
task pzlorltier lrrqtly depend an how important the system feels 
it is to Process the Phrase, 
In addition to the scores of Phrases, which combine a 
variety of factors but awe LndePendant of the larger gentential 
COntUtr the system form8 another arstrsmcnt of the quality of 
the phrase called the phrase 'valuee, which depends on the 
context of proPoScd complete lntcr~retrtionr for the entire 
utteratwe, The phra~t value is an estimate of the highcat score 
€or ell posslbla intetprrtationr spanning the utterance that 
Include the PhtaSt, The tltlnatt 18 tzornputcd by means at a 
heuristic atarch of the space at por.8 lble rentcntial contexts 
tltablilhed during thc PreViQUS tarks Pertarmed by the cxeeutlvc, 
The ~~16fif~ of a tagk Is inttially rut to the value of its 
aSS@cia0cd Phrrstt but thr Priorlty 18 lowcrcd it the tsar 
conflicts with the cxccutfvc@s cuortnt 'focur of aetivity", The 
phrarr Vlfu@ that dttcrmln.8 the initial prLorlty reflect8 an 
evaluation of both the internal rtrdcturc of the Phrase and its 
te)rtlon to its context, but it does not reflect its competition, 
If a Phrase he8 e high OalUe, other Similar Phrases are also 
likely to have high valbes, If values alone determined 
prlorltles, then even after successfully completing a Phrase, the 
system would tend to cantlnut looking for minor variations In the 
lame area rather thae moving on to Look far ways to construct a 
cgmplate interpretatton, The afacup df activityc mechanism 
provldts a way for phrases to inhibit the executive from looking 
for cbmpt@lnp phralts that would nece~sarily fcplacc them. This 
focusing 11 btoughg about by lawexfnq tha priority of tasks that 
look far r~~laecmcnts for any of e Set at focus ~hrases~/until 
the potential replacement pramlses to lcad to e aigniflcant 
irnpravcmcnt in Value for the final interpretationd The effect is 
to bias the sxecut-ivc toward bullding UP a complete 
inter~xetatlon using phrases in focus rather than exploring 
competing interprttatians that wauld not use focus phrases, X f 
the focus 8 wrong, the attempts to extend it to a complete 
Interpretation will be unrucceS4fUl. Eventually a taok that 
eanflicts with tne rocus wiir Decame the highost pr'idiity 
c>Perrtlan far the e,xtcutl,vrc to Pstfarq in aplte of the bias 
49rin8t Itr At a result the focus Will be modlfl~d so that it 
Is consistent with t,he new task, and tha, executive wllL then, 
cancentrate an urlna the raVl$ed ssr az Phrases, 
In addition to ealeulrtlnq priotltLes of tasks on the basta 
of Phrrrt value6 and focus of activity4 the executive must cnrurc 
that tha informattan gained through tho performanca of the task8 
s used etf~ctlvely, This ir done by strueturinq the parse net 
and the tasks that operate on it in a wry that brings together 
related actfvltier and caordinstes thrm to tllmlpctc duplication 
of effort, By avoiding dupLLcrtion, the rytttm rcduleeb the fU 
effects of the false step8 It will inevitabty taKe, Work dane on 
a false path fr not necerserily wartedr slncr it may produce a 
Phrase that can be ustd in soma other Way, for exampler a RPiraSe 
eonrftucttd ir part of an unsuccerrful search fog ant! type of 
Sentence may later appear in fhc tlnal interpretation sS Part of 
a dlfiarent kind of stntsncce AlSOr false steps arc not 
repeated, since the system only makes one attempt to build a 
partteular type of Phfrse Ln a P%rtl~Ul~t Zoca,tion in the 
utterancar rtgardl@ss at how many larger phrases might include 
it, Mlatakta are intvttabls, but at least the oytttm will not 
make tbf &&me rn1~rrre twgce in ant par$- 
Ta SUmwarl~t!, the language definition Is designed to 
facilftrre the IntWrrii~n of many knowledge sa~tc@$r Pules in 
the language dtflnftion cdntaln atttibutes and factors from all 
of there ~o~tces, The attributes are Urtd ta indicate partitutar 
proptttitl of phrasest and factors then Urt thc~c attributes to 
detrrmtnr th. #core of the phrase, The external reprelentation 
of the language, derlgncd tor easy U8t by people, in converted by 
a languapc definition compiler into an internal reprarentrtlon, 
derl~ntd tor efficient use by th@ rxecutlvr, In a step by step 
manner, thr emeutLvc user this information to create, rvrlurte, 
and coablne phrrrer, The of the next operation to carry 
out taker the form of asrlgning prfaritios to alternative tasks, 
Prlaritie~ ratltct bath the expected values of campltte 
intarptetation8 toward Which the task would lard and the relation 
of the task to the current focus of aetlvlty, Flnallyr the 
entire process is organized so that informatian gainad In 
Performing e t~6K 16 Shared and recordtd in ruch a way that It 
does not have to be tedfscovertd, 
American .fournal of Cornpntational Linguistic8 
Microfiche 34 : 19 
Artificial Intelligence Center 
Stanford Research Insti tute 
Menlo Park, California 94025 
ABSTRACT 
Thlr paper de1crfb~l a tunaable performance grrtamar) 
currently being developed tor SpWGh un4~tstrndlng, It rhows how 
attrtbuter of words ate drfinad and propaqdtcd to ruccrrtivrly 
larger phraser, how other rttrlbutrr are required, how 'factorr' 
reference them to 1 the Parser ehoorc araanp cowctin? 
d.et1nttionr in o.rdet te interpret the utterance correctly, and 
haw there facterg can easily be changed to* adapt the grammar to 
othar blseaurrrs and cOnt*xt#r Factors that @iiQht br clu8iftld 
rs a8Yntcctlc* are amPhrriZedr but the rttributrl they rrteresace 
nrad not be, end 8eldam are, pu~rly ryntrctic, 
Thlr ramrtch was rupportrd by the Dtfense Advanced R8@8&rCh 
Project8 Agrncy of th~ Dlprrtmanr at D~tenae and mongtoxad by tha 
U,s, Army R*r@rrch Offlca under Contract Ha, DkWC04-75*C-0006, 
8 perfotairnce grramrr (PC) d*tihrr tha fatm and meaning at 
the kind8 of uttQrancer that occur in rpaatanooua dialog, When 
tnr detinltianr of the grammar RtaVid~ tnformrtfon that htl~s a 
prtrcr ehaora these rule8 ao8z likely to laad td correct 
int@tpretatlonr of uttrrtanCs&, the grrar@ar 18 raid to be 9tunrd*, 
Uhan the tuning 1.8 crarily chrncraa wnrn thr dornrin ot dircaurrr, 
ehanges, thc grasmrr 16 said to br *tunrrblr#. The ability to 
tune r grammar &a particularly inportant in spreeh undrrrtanding 
where the inhatsnt Uncmrtrinty of the input caurer falre path8 
through the grammar tb be multipll@d, 
Thil pa)rr dercribrr a funrablr PG baing developed Jointly 
by 8RI and SDC for r camputrrrnbrred 8P@@ch underrtanding eystam, 
7[t# vocrbulrty and phrr,ro typesr ralscltod frsm pt'otoc44rr rra 
appropriate fat rrelng and anruering qus8tionr about propettirr 
ot 8ubarriner. Thr PG now dafiner War 70 word and phraae 
crtrgorter, Its reape rxtandr far bayand syntax, A dl#~ours@ 
coaponrnt enrbl~o if to nand&@ rnaphora and ~llipiit~ 4s int 
*What ir the surtrce dirpL~crsrnt of the Lafeyattr?.l.. What 11 
It8 arrf t?", and "What in the irnuth ot tnr baEayrtte?...,, The 
Ethan i~len?~ X remantic8 component definer a common meaning far 
paraphrrsar, b8 gn Vhr Sperd at thr Lafayettr ir $0 knotrH and 
*the Laf&y)tpe h.8 a rperd of thirty knatrW. (Sar Walker rt al., 
f97St Paxtan rnd R~bin~ont 1975; nrndriw, lW5t Dautlchr 197dr1 
Each dr@lnLtion comporlng thr PC ha8 threr prxtr, The flrst 
name$ r word crtegoty or a phtr8e Catrgory and pravld@r r 
cont'rxt-ftrr production far ftr caaparltianr Tha racond part# 
ca1Ied @rttributcs@, tells how to drtrralne fha propertlea of an 
Lnrtrncr of thr cateqory. Any detinltibn ern r4tferrn~8 nUltlp2r 
rourcer of kn~wled~e~~aeourtic~ syntrcttc, rrmanticl dlrcour#e, 
et prrgmatle-mior infotartton nreded to dutrrminr rttribute 
~11~8s I The third part, @fcctorrO, define8 #corer for 
combtnrtLonr of rtttlbutrr, indicrtlng how wall they Qtlt', 1 t 
ir through factor rcorer th&f the granscr is tuned, Thr 
indtvidurl scorer re rmbtned into r eolporltr rcorr which I8 
used by the parsar te eh008c rnong competing plr#fnglr A 
putpattrl lnrtrncr of the deflnitlan rlth r reate of OUT for any 
factor Ls Lamedlrtely rli@lnrtrdr a Lor score may riinlnrtr r 
p&tslng path! r high reorr rnhrnerr the priority of I prtrlna 
path that appller the drLlnLtien'@ 
Our ~nemanie term8 for factor rcorer 1re VERYCOODt GOOD, OK, 
POOb BAD, md OUT, There rre trtlartet af likeiihood, They arc 
neclssarily v8uu1t beCaU8e wt arr drrlinq with grrdurl phrnamrnr 
and Probabillstfc ttndenct@ar They nrcn tomething like npultt 
Ilk*LyN r ui~pd~t@dRt rardinary@r "odd but ~c?~~ibl~~~ 
*unt/kely-lJ8teh rgrinRf rrd "8o SPrelrl that wr do not drfinr 
Itw, Rlpid, prercriptive judqnontr rrr avaided, Combining 
wfaatm rkth R-8F a8 r plurrl noun 18 Ind~ed w~dng and therefore 
OUT, On the other handr @fueln doer Cb@bln@ with plural @-an 
rlth thr r~relrlitrd ntaning Rklnd@ at furlu. At prerrnt 
Vlu@Jc\ like Vfsatra 1s judged ta br OUT tot out lrngur~e, but 
$hi8 judfiteent can ra.r$ly be C%t@tQdr &f we find thrt our lrhgtrrgr 
uarrr refer to kind8 of fuel rr kt~rlru, 
22 
liner frgtor scorrr aan br changed without tttecting thr 
rent of the d@tlniZibnr thr grammar tr tun@rbl@ to diffarrnt 
direoutre domrlnr and rtylrr of SPrrking. hirot 1f one factor 
drtlnas r low rcorr for an lnrtrntirtlbn, other$ ary stlL1 tats. 
the cenporlte score, A rtatlrticrlly improbablr phrarr thrt 
~rkrr renrr and !I uttrtrd intrlliglbly should not be unduly 
dJtilcult to racogn2zr and lnterprrt, 
The rrrt of thir paper @xrrLnrr tepurncer ot drtinitiont 
rrquired for prrring and Undrrrtandlnp a typical utterance. We 
begin wtth word aefinitionr~ and rhow how the attributri sf wardr 
rfr ~ra~~grted to ruec@rfiivrly larger phrraer, haw athrr 
rttribut@S prcufiar to high@r*l*V*l phrrrrr at@ addrat an6 how 
frctorr rsfrrenc* thaw in tuning the grrmmrrr 
Preceding d$ficourfi@ and unberl~lng asmrntic dlrtinctiona 
con# train the rurfrce ryntrx of an utterance, Bscaure 
ruprrfici4l ryntrctic prapcttirr signal thorc constraints, it 11 
often econoaicrl to usa ryntretic trcforr in order to alrcontirm 
r wrona parsing path or confirm r corrdct an*, avoiding @all# on 
rrtmrnticr P ~LIcou~~@~ and r~ouotlcr bat axpanrlva 1x1-dapth 
avaluatianr, For r#im~1e, if 8omranr ray8 Vucl rupplirrn, wa do 
not want th~ prrrrr to rnplorc, In depth the application of rul,rr 
that build r plural noun-phrlsr tram *furl $,,, II wkthaut: 
conlldrring an altqrnrtlvr deLlnltLon in which wturl~8 a 
aoditi*i 0-f r auntablo nomin.1 beginning with So thir end, 
ua incLudr a frctot that chackr the countablrnarr ot qturlw by 
rrfrrraclng a countlmarr/unlt (CHU) attribute, which Lu syntax 
ortented but rrrrntirlly rra&ntlcr barad, 
txrmple$ of soma uraful ayntrx~orirntrd rttrlbutrr drflnrd 
for the word category W (noun Btel) rpPIar in (1) belor, Evrry H 
ha6 r value tor the CNU rtttfbut. drawn from th@ &at (COUNT HAS8 
UNIT), NI With the CNU value UNIT (auch I8 "t~~t% 't~fi', 
Mknotw) eoabinc rrsily WLth pIurrl auiflxer and number 
cx~tt86~ort~ c~~Q~~ @two kknotaUr nffl~~ fertRlr but not go well 
with ast initr aetsrrfnrcr  these twe knotam), or ~rnfti~~ 
ruftlrer tathr twenty knot#@ 8Perdal. (Ct, "he Ethan Allrner 
(1 1 WORDS ,0EF 
FUEL cnu CMAIII~ 
FOOT 
CMU 8 cunxl], ~~swrr 8 WQ~ 
LAFIYETTE CMU a (COUNT)) 
SURFACE,DISPLACEMENT CHU r (COUNT), RELN 8 TI 
TON cna 8 (UNIT) 1 
Like th* CMU attribute' the REtN attribute it errrntlalLy 
Sirn(lnt IC. It mrtks such words as @surface di8placenentwr 
"rp-@adr- fflen~thr, and sdraftr ra rpecL&l *reLstlanare noun 
Word8 SyntrctlcrLlY, relrt3onal Nu do nalb combine rlrdllY' nlth 
plural tuffixer and nuabcr exprarrl~n8, and when thry do, the 
rrantng 18 rpcretrllsrd, To @om@ degrr~., they era like maw Na; 
@tht+e splrdsVthraa Fetes at 8$@@d) I8 analogour to Vhrre 
IS itoePtrblr~ while R& fudl of two tgnrfl 18 iL1 farmad, 
The Cttributr PLIUPF dirtinguirhrr irrrguiar plural8 like 
*t00fq, Unllkr thr CMv aria RELN rttrlbutrr~ it is purely 
Atttibutal 4tfectlng the rbildtg to combine with the plural 
~~tfi~ br3u @re retatrnced fn the two cornpa8itian rules of (21, 
defining the crtagoty NOUN, The attribute Statamenfr Fropagate 
the rttributer at the item, adding a numbrr atfributa (WIR). Tho 
flrrt factor or N-1 reirrencaa the CMU 4ttribUte and rtrt~la that 
if the value ir &A$$, then the acora 18 GOOD, This juagmsnt 
lncorporrter our knoulrdpr that the ether ruler N2, cannot 4pply 
to mast noon-Stam%, It tha token i8 r mar8 noun-atam, N1 is the 
rlght campotltion rule Za apply, 
(2, RULE,PEF N1 NOUN a Nt 
ATTRXBUTELI 
CIU,RELN,PLLUFF F~OH N, neR a mcsc)l 
FACTOR8 
CHU IF CMU EQ "(MASS) TH€N GOOD EtSE OK, 
REZN x IF RELN EQ *T THEN GOOD EG8E OK, 
PLSUCF IF PLSU~F EQ #no THEN GOOD ELSE OK) 
EXAMPfrE8 
SURFACE DflPtACEMENTc FOOT, FUEL [GOOD) 
GUBMARINE (OK)# 
kl.F3hErSEF N2 NOUN * N lPfrj 
ATTRXBUTES 
CMU,RELN~PL$UFC PROM N, NBR 8 "(PL); 
rACTGIR8 
PLSUFF 8 IF' Pb8UFF EO RNO THEN OUT ELSE OK, 
cnu IF cnu ta fl(nAa8) Tnrn OUT ~b8t OK, 
UMIT 1 XF @UNIT T!4 CHU THEN QDOD CLUE OK, 
REtH 8 LF RE&# EP V THEN POOR ELIB OK! 
EXAMPrtEU 
FOOT *$, FUEL -8 (OPT), TONS (GOOD) 
8URFACE DISPLACEMENTS (POOR), SUBMARXNES (OK)! 
bike the CMU frctarr, thr Ft8UFF factarr rnhrnca the #ear@ 
tor applying N1 to stems that do not take a piur.2 rutfix and 
Cohltrain NZ not to apply. A RELN factor Onhrncrr thr score whrn 

s~ba&rlnefl, @thole fualrvnnd accept What furln 8.r OK, while 
awhich tonsa &end "that draft at f'ive featr ara POOR, Factor8 for 
NPl1 e;iainrta fU*lHr "1 draft ot tha LafiyetteU~ and la 
$ubmarlne.r@9 4Cccbpt "4 subrnatinr@# tanw? *the aubmrfneM, and 
"he 8Ubb.!WS@d spmdU, and Brcote @the tonm and *the dtrtt at tiva 
feetm rs POOR, 
3 RULE,PEF NP4 NP 8 NUMBERP NOMj 
A1[LTRIIbUTEB 
FOCU$ w "LNDEFl HOODINUM FROM NUMBERP, RELH FROM NOM, 
NBR 8 GZNTER8ECT(NBR(NUMB1ERP)INBR[NOM))p 
CMU 8 GZNTER$ECT(CMU (NUMBER?) ,CHu[NOM) ) ) 
FACTORS 
CMU 8 IF NULL CMU THEN OUT ELSE OK, 
HUN 8 IF FITWD(NUMBERP) IN @(HUNDRED THOUSAND MILLXON) 
THEN OUT em OK, 
NBR r XI- NULL NBR THEN OUT' ELSE OK, 
UNIT 8 IF *OMIT IN CMU THEH VERYCbOO ELSE OK, 
R,ELN r IF RGLN EQ T THEN OUT ELSE OK! 
RU&E,DEF NF7 NP a DET NOMt 
ATTRIBUTE8 
rocus 8 WEF, RE~N FROM NOM, MOOD FROM DET~ 
CMU i CfNTERSECT(CMD(DET]rCMU(#OM])p 
NBR s CIRTERbECT(HBR(DET),HBR(lOn)); 
r ACTORS 
CMU 8 ZF NULL CMU THEN OUT EL8E OK, 
UNIT 8 IF VNIT LN  MU THEM PUQR ELsE OK, 
NBR * IF NULL WBR THEN OUT EbSE OK: 
RUL&,DEF NP11 NP 8 ART NOMj 
ATTRIBUTES 
'RELN FRQM NOMp QOCUl FROM ARTl MOOD @ 'DEC, 
CHU s GINTER8CCT(CHU[XRT)tCHU(NOM)), 
NBR u G~MTER~ECT(MBR(ARTI,NBR~NONII~ 
FACTOR&' 
CMU 8 fF IUbL CHU THE# OUT ELSE OK, 
NIR 8 IF NULL IBR THEN OUT etsc OK, 
UNIT r IF IUNIT IN CMU AND FOCUS EQ WDgF 
THEN POOR Et6E GOOD, 
RE&W r fF RELN EQ T AMD FOCU8 EQ WINDEF AND 
CHU EQ q(COU#T) THEN OUT Et8G OK1 
Zn arch drttnltion, r UNIT trctot tetrranees the CMU 
rttrlbuta ot the HP, If the vrlue L8 NIL, the drflnltion 18 not 
appZIcable. Zf UNIT is I valuI~ than the UIZT factor Yar NP4 
score8 the applicrft-on as VERYGOOD. Them arr two rraronr for 
this judgment. Numb@? axprrrrlons are tYpicrlly found with unit 
rwpresslonr to fora me&rure rxpterslons, and unit8 are more 
likely to occur with indetlnitr than wfth dtftnlte focus, rr the 
ptecedlng example8 Catventy Knots* and so on) have Lndicated. 
Since the focur for NP7 i8 dlwrys drflnttet the UNIT frctor 
8ecrerses the wore fat rpplylng It When the UNIT vrlue rppe~rr 
in the CMU rttrLbute, got NPIIt the UNIT frctor scorer the 
application GOOD if the artlcle 1s *rw and UNIT rppears in the 
CMU valuerr but POOR if thr artac~e rs w~heRr 
NP4 rpplisr esprcially Well to instancrs ih which units ara 
prermntt but does not apply 4t all If the hard of the nominal 
canatitusnt $8 a RE&# Stear In dlsceursd about warhlng mscninea 
and bleyc"les, 'three rpeeds@nti~ht occur in m drdinrry way but 
for our curtpnt discourr~~ we do not antlclprte Such 
comblnrtlon, Certainly, we lo not expect Wthtee surface 
dl@Pl8C@IR@~tl', 
luch conrtralnfr rrlirvr the nr.6 for detailed rnalyrir. 
For Lnrtancer rrluar that thr @CoU#tLc mapper ham tentatively 
~tf~teb both *SUbm~ttnaw 8nd ~UUblmrgad $Peedn IS a~0Ulfl~11lY 
pL1u8Lblr altrrnrtlv*r tar tilllng the gap in the parttally 
analyrra Phtr8e wtnr@@ -8 OL the U.8, Navy\ Thhl8 Is not 
ilpr~brblr SLnce arubnrrLnrr" and nrUbnrrg~d ip@ad#* terembla 
arch ether in @any ways. They both rtatt with Vr*r their flcrt 
ryllrhle8 hrvr central vowelrl their last ayllablrs have high 
front vowel#) and so forth* If NP4 Is to be applird, however, 
the RELN frCtbr will resolve the do~b~t in favor ol w8ubmarlnen, 
and theta +ill be no nerd to teat In depth how we11 ulubmatged 
8pqabw maps onto tha rcauttlc data at fit8 the aemantlc a04 
dlrcour8a conrtrrintr, 
The UNIT factor of NP1& guider t'Ae choice between V" and 
atti@\ where acoustic evidrnce tor 4 choice is typicrlly lacking* 
semmticell~y, qrl lrlamblrr wonsw in it8 ability to comblne with 
nUgber8 and unltap erg,# tan5 ma tonm, none hunbr@b\ 
hundredH, It the inrtancs ot the NOM Is mtonMr mfoot" wkn~tw, 
ar Some other singular rxpresaion wlth the value UNIT tot it$ CMU 
 attribute^ then mrVs judgsd to br more llKely than wthe", On 
the Pthw hrnbt it tho NOM la atuelM or VubmrrinrBst the artlcla 
cannot br T~U CHU attribute tot "aw is (COUNT UNIT), which 
doe8 not intargrct wtth the value (MASS3 of the CMU attxlbute tot 
mfualwt the NBR attribute is (SG), which does not intersect with 
the value CPt) tor wrubmarin~UM, The factor8 r@ferenclng theae 
attrlbutsr rule out rppllcation wh4n the intersection Ir NIL, 
There ate typical ryntrctlc agreement ta$tte 
41 longer phrare8 r built up# tha vartour attribute8 
interact in other wry#, For instanc8t tht4 ryntactic pxaptrcier 
ot ralrtl~nal rwprrrrionr depend on which arpect at the relation 
t 8 prllacrrnt Ln an accomPanYLng preposltianal phrase, 
Prrpositional phrrs'er hrva the aftributrr of thclr NP objacti. 
Whrn a pr6positlonel phrare modlfirr a noun with the RELN 
attribute, thl CMU rttrrbutr tot thr rorurtrnt phrrrr ig drflnrd 
by trK$ng the union of the vrloar for tha two nomin&l 
constlf~ent~. As 8 re#u&tr phrrtr8 like urusL~rr Ulmplrerment of 
the brfaYettru have tkr VrZU8 (COUNT) ~nd those ilk@ Vutfroe 
dirplrc~er\f OL seven th808and tonsa hrvr the vrluc CCOUW? UWlTl, 
The dttferdnce in vrrurr mark@ the tact thrt the two exaapirr do 
not fit rltn equrr rare in a11 syntretle 4nvlronacnts. Xt is 
taterwzewi in thr UNIT and R~LN factors in (31 above, to 
LnfLUln~e the choice betwetn the two rrticlrr, which rrr srldom 
dirtinguishrd eler~iy by sound, The rule i8 tunad fa ptrfar 
%aa@ in the rbrence~ a# the UNIT vatuet rr In -the surrrce 
dlspl~~#mant of the Latryettcu~ end mra when it ir prrmrntr am in 
"a 8urfaea df8placemlbnt O-f seven th6U11nd ton#*@ RA SUI~ICI 
dleplacutent ot thr &~L&y@tte% wrhteh impliar the porribility of 
havgng sort than one ruttaee diSPlaccaantt is ruled out 
carP&rtetsr, 
NPs rlro haye & HOOD atttibUt*r deri~@b from their initlrl 
canst~tuents. It 3 either declrrrtivq (DEC) 4s in Vthtt 
rob.rrinr% er WH-int*rr~uetivc (WHI rr in wunich rubmar &new . 
The WH valur 28 PtoPagrfed to thr lrrg8r PhtrSr8 in which NPI are 
con8tituent8, 8entrncea (Sl md uttrrrncrr [Uj t&lr the vrlur 
f4t their MOOD ettrlbutb ftem an lnitt~l WP, Our currant 
Voc.butaty doer not include Vetbr like Wknow@ and wtrllfl, which 
Cub embed %H QuaStionl likr @Do YOU knbw whit tha Iuttica 
dlrp%acamrnt LS?VQ~ ttha t imr bring, rr rcrusr that nantnitirl 
abon PnrdMlr rrr qot likely to have the vrlur WH tar HOOD. Echo 
qutstionrr *,ger ayb~ laid what?B ate not ruled ~Utp but have 
The convergence of many attribute6 at the hlgher level# ot 
Some of them are thown in (41, 
ATTRIBUTed 
MQObrFOCU8iCMUlREtN FROM #PI, 
AFFNEC FROM AUXBt 
FACTORS 
NBRAQR1 s IF CMU EQUAL m(UNIT) TWEN 
(IF NBR(AUX0)EQUAL R(~O)THEN OK ELSE OUTlELIE 
IF GINTER8EeT(NBR(NPi)INBR(AUXB))THEN OK ELSE OUT, 
NBRAGR2 8 IF CMU(NP2) EQUAL I(UNIT) THEN OK ELSE. 
IF' GINTERs~CT(NBR(HP~),NBR(AUXB))TWEN OK ELSE OUTl 
FOCUS 8 IF FOCUS(NP1) EQ UINDEC AND FOCUB(NP2) EP WDEF 
THEN POOR ELSE OK, 
GCAIE1 8 IF GCASL(NP1) EQUAL "(ACC) THEN OW?' Ct8E OK, 
GCAIE2 8 fr GCAIELNP?] EQUAL *[kCC) THEN OUT ELSE OK, 
MOOD1 8 IF MOOD EQUAL *(UH) THEN GOOD-ELIE OKt 
HOOD2 r IF MOOD EQUAL *(UH) AND MOOD(NP2) EQUAL .(WH) 
TWEN POOR ELSE OK, 
AFFREG 8 IF MOOD EQUAL a(WN) AND AFPNtG EQ gNEC THEN 
BAD ELS~ U!$, 
RELN a IF- RE^ EQ WT AND CMU(NP2) EQUAL m[UNIT) 
THEN VERYGOOD .EfrSP: OK, 
PERShCR 8 IF GINTERSECT(PERS(NPt),PERS(AUXB)) 
THEN OK ELSE OUT; 
EXAMPLfES 
THE LAFAYETTE IS A SUB MAR IN^ (OK) 
THE LAFAYETTE 18 $UBMARINESr WHAT IS THEM (OUT) 
k UFAYETTE IS THE dUBHARINE (POQR) 
THEM ARE SUBMARINZBr rT AM A SHIP (OUT) 
WHAT XS IT' WHAT X& THE tENCTH (GOOD3 
HOW MANY ARE WHAT (?,ObR) 
WHAT fBNeT THE .IURFAC& DISPLACEMENT (BAD) 
THE 8URFACE DXaPLACGMENT Id 7006 TONS (QERYGOOD); 
The PERSAGR (perron~agre~mtnt] factor trrtr Ear rgrsamcnt 
betwean the ro~crll*d pronoun8 and the auxiliary conrtltuanr 
The two grrnaiti~al carr tretorr, CCAdEl and CCAdLZ, rrqulrr thrt 
the $;~tclmnz&tf~&1 c8sar of the two NF8 rra not accu8-rtlve, Thcrt 
tridltianal ryntrctle agreement tests black applieatian of the 
roapealtten rule to putative axprerstonr like Yt rrr" and @they 
S 'Them I@" $S d~~bly bl~~ktdc 
Some of, th~ rr~rlning factor (tatcmrnta in 4) rtr lrsr 
trrdittonrlr On8 of th~re Is the AFFNEG frctor, which raferenear 
bath the MOOD &nd AFFRtG rttrlbutas @nd reducer the resrc Qrratly 
Lf the Anrtrnca ir purgortrdly r negative WH qurrtion Ifkr "what 
Itnet the surfrec dl.tplac(rnent?a Canulna rcqutltr tor nagatlvr 
infsrsrtlon accur in highly tircumseribrd rltuatianr, Tha 
rh@taricrl questtan I8 not r qQnUfno raquert fbt Infor~btfbn 
[a,Qet wWh@ wauldn't ltkr to b@ rich cnd fanourtW), "Who ltnCt 
herela is mhS~nabl@ only it thrra is rn a8tabltrhcd and limited 
List of paopla who aTe rxpeeted ta bc $rerant, ar in a elaroroon, 
@What isn't ye~t nrsel-rn =Where don't you live?-are patently 
Ub8U?dl 
The constraint on negative WH quartianr 18 arrrntially dur 
to Prcgritkc totcer aa we11 rr armbntlc onrr, 8lrniLlr forera arc 
at work in obrervrd tendencies for the first NP in the 
corvorLtlan detinrd by 53 to br indefinite in Faeur enly when the 
rreond one is also. Qtdbted ovrrrimply, in eohrrrnt dlrcour#r, 
the thing8 rltcrdy trrkrd rbwt-the %olw inforartlon-trnds to 
C~BI tir8tr What !. s pr@dicrt@d about if-tha wnaw'5 
Intorrrtlon--tend# to fallow, Old infarmation ir Infarmrttan 
thrt ha& rlrrrdy been trlkrd abwt and eatablirhrd in the 
dircours~~ 14 that bt 11 lfkrlY to be ancabad in degLnit@ nsun 
phrases, Thcra ire likely to br In rubject porltlonr so that the 
oentrnca they intraduce Is conairtenf! With preceding Hntence~, 
Naw Sntermrtibn tandr to b introduced in Indefinite naQn 
 phrase^, Tha next mantian of the Vmnc thingA will than ba 014 
infarmrt$,onr 4Llplblr tot definite Paau8, Canraqurntl~~ uA 
Lafayetta is that bubmarlnsw racmt peeu1latr ralatlve to @That 
rubm8rlnc ir r tatay~ttu~, HA brtryettc is itn ia 6t111 mote 
pcbeul tar, Thrrs dlt~~~~~lrmba#@d ptpOk,abiLi#tic tsndencl8r lrr@ 
exprerre$ in the POCU8 factor at 83, 
The CMU att~lb~te~ al Pr~Vio~sly nBt@d, i8 not puw4'y 
ryntarctic, Qn the ather hwv!lr rnlrrttcrrlrr Ilks number agreement have 
rlway8 bean c~ntrrl to ayntrx, It it grxticol%rly intap@$tlngl 
th@r@f0rer that the numbar rgxaemant conrttalntr far 53 cannot be 
properly rtrtad wlthout rppaallng to CMU, To atat@ number 
rgres~tnt contfraintrt NS denoting unit% muat ba warked 
raparstely, dsntencer like *Theas ara a tubmatinem, nTThr8s Ir r 
tarpcdo tubam, wTht86 Lt nf$!Bll@ Irunchrrs*, end wThir are aubrR 
are clC4rly unqrammrticrl, and the ungaammatLcallty is ururlly 
attributed to tha fact thrt one of the can$t&tqtntr differ8 tn 
grammatical number from tha other two, Haw~vdr, "The rurtace 
dl%pl&cement 18 rsvrn thousand tonsn la wnaAlY grrmmatic~1 @van 
though two ot the sanrt$tuantr are ringular and the third ir 
plutaA. 8ueh UIC of rrmantic rttributrr in ryhtactlc Lsctorr, 
paint8 to tihe conclu6ion thet the intcrgrrtlan ot Ilnfarmatlon faam 
dlffdr~nt rourea8 OF knawladge is well motlvrtad an bath 
l$ngUtttlC and haurLatle grounds, 
BIcruse ot the high frequtncy et HH qutrtionr in the 
protocols from which the V~Clb~la~y and phralc type& ware 
seltCt@dr the PG 11 nor tuncd to rxpcct them. A sentence dcflnrd 
bY 53 r@etlvt& i higher ScorQ troa the MOODi factor it its MOOD 
ts UH, This tuning can earily be changed without altering the 
ryntrx or senantics at the language, If the user bath @~tr&etl 
data tram the data bare tnel enters data into it, wfth n6 
prodfefrble pltt+rn of altetnctlont Facfbrl lik) MOOD1 can riaply 
be removtd. A mete intrrerting alternative ;a to ra8et them 
dynrsicully In r diseaurrc context whara tha computer raaatlmsl 
askr quertlanr far tbr ur~r to rnawrr, Atter each Ular Qulrtlant 
the grraarr could be tumb to rxpcet r drelrrrtlve utterrne* 
who#@ gyntcx an6 ramrnticr wrrr rpproprirte and r@lavant, 
American Journal of Computational Linguistics 
Microfiche 34 : 34 
Artificial Intelligence C~pte 
Stanford P~search Institute 
Menlo Park, California 34025 
ABSTRACT 
The semantic component Of thc tpecch understanding systgrn 
being Q~velopcd jblntly by and SBC rules out phrase 
combination8 that are not mtanlngfu) ~nd ~radueag semantic 
intcrgretatlong for eombin&tlons that area The system conslstr 
of a semantic network madsf and rou%lfl@$ that intaract with it, 
The net 1s partitioned into a set of hlorarehicalby ordered 
subnete, fdsilltsting the encoding of hlqher=~rde~ predlcstet and 
the maintenance of multiglc parring hypotheses, Composition 
€aUtinttr mrnblning Utterance earnpanant% into Phrases# consult 
nctwqrk dQcctlptlong of prototype ritortiana an3 surface-to-deep- 
cafe mspr, Outputr from th@r@ routines are network frnqmcnts 
canrixtlng of raver.1 rubnet8 thet In aggragatc capture the 
intcrrolrtia~rhipf betwean a phtr8ces syntax end rcmantlcs, 
Thi~ r@ocarch war supported by the Defsnsc Advance Research 
Projects Agency of tho Dcgartncnt of Defcnre and mon-f5torcd by the 
U.S. Army F18errcn Office undar cantrrct No, DAHCOI-75-C-0006, 
OVERVIEW 
This prpcr describe8 arFtctr of the stsintic co~pQncnt af 
the rpccch undargtandlng systcr currently bclng developrd jointly 
~y SRI and SCC, [Tor a ea~prchehrlvc discussion of nonacourtic 
partlonr of this sY&tomr tee WeLktr et al,, 1975,) The senantic 
eoQpQntnt conrlstr of two majar partrr a semantic network ccdlng 
a model of the trek domain and & batkqry of crmantic ca~~positian 
routines (SCRs) that art caardinat~d with the languag~ definition 
(roughly1 the ngrammarw for the eptcch undorrtandlng syttrmt gee 
Faxton and Rabintanr 1975, and ReblnsonI 19751, This paper 
cenctnttrtcs e%cluSiVely an the inttrplry b'atwecn these two major 
pact6 during parrlng. However, the rcarntic component also plryr 
importent roLcS in knorladgr managementr di@cour~c snalYsigr 
prrdlctLon, &nd question rn$narlng, 
An SCR jr called with n~twosk rcprerentatlsns of comPaments 
that the assacS~ted language dctlnitian rurc hag found ts be 
ryntactically capable at cornblnln8 to farm e larger PhFar@c 
UIIng knarltdge from the Semantic net, the SCPI adlminatc 
cambinations thatr although ryntactically acctptabla, do not meet 
scmantlc crltcria for msrningful unification, For eombinrtiant 
that are receptablrr the JCRI build network ~truct~rM ts 
reprerent the meaning of thr cosporitt phrarc, using the network 
rtructurer ot the camponant& ar building bloclcs, Thcae net 
atxuctwres QFO eonstructad so that (1) multlplc hYpothCSaS 
csncernlnq the proper lneorperatton af a given Utterance 
conrtiturnt fn larger phrrrer my be oncodrd rimultrnaourly in 
one netr (21 compttlng users of r canatltuent may share a single 
netrotk structure representing the constituentr and [3) tho 
assaciation between aaeh syntactic unlt at an input; and its 
translation image Ln the network Is cxpllcttly encoded for use In 
discourse analysis, 
THE 8EMANTfC NETWORK 
The ramantic network ie tha ptinclpal InfsIYhatisn source  PO^ 
SCRIP cncodlng euch diverse rntltlrr as objectsr situatlansr 
ertrgorler, taxanomlra, dcfinltlonrl and quantlflad 8tdtrmcnt~. 
Network structure8 indtcrting portible telatlonbhl~s between 
objects rrc uacd to drttfminc the moaningtulnc~ss af ghkasc 
cambinskionor while tne network itself serve8 at the medium tar 
recording lnhrprdtrtlona of uttrrancc fragment8 during Parting, 
The ~ttucturt of this network differs fr~m that of canvrntionrl 
netr in that nodcr and srct arc ~~rtltioncd mta 8tg~~~~A, These 
8p@c@gr Dlryln~ in networks tola roughly analogous td that 
played In strings by parenthcsarr group infornation into bundler 
that help to candanre and orgsnlzr the network*$ knowltdg~~ An 
introductlan to net partltionlng 1s provided allawhetr (Hrndrix, 
19351, 
An 1lluatrativr portlan of the permanent knowledge secttan 
at thr rrmintic network 1s depicted in Pipire i. In the uppar 
left cbrnsr ir.ndda 'Uor rrprr#enting the univbrral set U, Ta 
the right 1% nod* QPHY80BJ$*, rrptalantinp the rat PHYdOIJB of 
FIGURE 1 A aFAIMPLiNO FROM THE GENERAL KNOWLEDGE NET 
physl~al ~bj~eta. That PHYSOBJd I8 1 sublet of U 11 lndicrtlld by 
the smart from 'PHY80BJ8* to 'U'. A subset of PHYBOBJS is SUBS, 
the set of all rubmarinss, A particular elemant 02 SUB8r as 
indicated by the r~are'lrom *DOLPHIN@ to 'SUBB', is the DOLPHIN. 
The DOLPHIN is r DartltfPant In r Qattlcular situation, HB, 
the rlturtion in which thr DOtPHfH has a beam of i9 feat. HB is 
rn clrmtnt of <HAVE,BEAMBP the set of all rlturtionr in which a 
physical object is chrractarizrd by a rnrarurc at itr breadth, 
Cartrin autgoing arc# fro% r nods rsprrrrntlnp r rituatlon are 
ured to specify rbtuatlon rttributrr through derp rrmrntlc crtcs, 
lor txrmPLe, the outgoing obprrc from 'HB' rveettirs the value 
of the wobjw [(ObjftCt) lttrlbutt ot HB to be DOLPHLN, Hera@tttr 
the notatton a~~ojw will bt uled to indtchtc hthr vrjoe ot 
the attribute (Q) Obj** 
The network of Fluure 1 hat bran divL4e6 into five spacer, 
KSP S4r 85, 86, and 87, Pictarirllyc each of thrle 8paeCa 11 
tdOreatnted by' boxc The mo8t global informrtion In the network 
ir encoded in rprca KS rthr outerrnofit bog, ramrtimrs erllQ4 the 
wKnowledqa Spacea) which Includqs Such ~n~tltirr rr node8 'IJ' and 
'PHYBCJBJ6' and thr r-arc connrctlng them. Thr baxar rrprerentinp 
#vrcrr 84 thraugh 37 may be thought of as hole8 ln the box of KB, 
Patrllaling the rrlrti6nahlp botweqn an inn~r and an outer black 
of an ALGOL program, qach of thrm sprees 8~ecitle8 a nore lac41 
rrra of the net than 18 rpeciticd by KS. From the pcrrprettvr ot 
85, far aXamDlr~ it &S Pasribls to accrtr bath lacrl node *Po and 
(rtlatl.vely) Pl~brl node .PHYIOBJIF, Haravrt, from KI the nodal 
and rrct Inride $5 ate not &cc&##ibl@, The hierarchy of SPlca 
1oealiertlon nay be ttprerrntcd by r partial ordering suoh rr 
that of Flqura 2, From any FPICC 251 the node8 and arcs arc 
aCC*LISblc that lie in 6 or fh any mace S' above 8 in tht 
hierrrchy, For rxawpler from 83 onty nodes and arcs in 53, 62, 
$1, and KS &?Q &ceersfblc, 
Pictorfrllyl kt may ba nscegaary to 6raw an arc crossing box 
baundariar , In such erars, the arc belbngs to the space (or 
spacer) In whose box tht ,arc Zabrl is written, Sp&~et may 
overtap, For exampler in figure 1r node 'ED,HBe ller In both 
rpacc 84 and rpaeo 85, Further, r spate may tcrve 4r a node in s 
more global Spice, Both 64 and S5 bhhsve &I nodes in KS and arc 
connected by s conssrsrc (~anseq~rnce]~ 
FIGURE 2 SPACE LOCALIZATION HIERARCHY 
Typleally, locrllzed rpaeQ8 ruch ar $4 and 55 art used to 
@ncoda higher-order "~rtdicrte~,~ such as quantiflarrr logical 
eonnrctlvrr, and hyp~thetic11 data, Here, 84 and 55 rm used to 
ancoda an fmpllcatlon, Thr rpaet 84, doubling a8 a node in apace 
KS, Lo connected by an emarc to '*fHPtY>' and by a eanrwrrc to 
'SSPr The inttt~rttati~n of any tiamant of set cLM~ty.> $8 that 
if rntltlos can be found matchlng the structure of the element 
Egticer then the exirtence of rnttties matching the rtructurc of 
tho rrsoclated con14 Wac@ may be interred, The only StnI~tUm 
encoded In elomrnt Speec. $4 la a nod@ PED,#B0I with an @-arc to 
QcHAVE,BEAM>@, This structure mateher any concrete instance of 
eHAVE,BEAM> (such as HB), Thus, fat any instance of (WAVE.BEAM3, 
entitlss matching thr tttuctore of 55 JnUIBt rxlrt, The Stru~tuxe 
of 65 indicates that the element ot cHAYE,BEAM> will have a 
IQobj, rhlch is an tl~mcnt of PHYSOBJSI and a #@mcarurt, whlch is 
rn element of LINEAR,MEA8URES, 
The ImPlication mcQd1)4 b3 $4 and $5 arervclr to delineate the 
8@t GHAVE~BEAM~, Thrt Ia, the im~lieatlon lndlcatar all the 
attribotrr tdrrp ear~sl of r gHAVE,BEAM> riturtion and their 
r@nWr of 4e~1ptrbla vrlu~~, Thlr d~llnrrtlon may ba urrd during 
parring to teat the plrutibiAity of r given group of rntitirr 
baing united in r (HhVE,BEhM, situation 6rr in a prcdlctlvr mode, 
ta rUug@rt &o#glblc rdntance partlclprntr, Such delineation8 arc 
mcadrd for every rltuation and want rat known to the system, a 
recond cXsmPla in Figure 1 baing the drlinration at rrt *BUILD>. 
THE SYSTEM IN ACTION 
The urs of She SCRs and remantic network in translation may 
be seen by conriberltlg the parr.Sng of 
*The paw- plant at the sub was built by Westlnghouserm 
The Ultimate result of the translation Process for thts utterance 
is the network structuge recorded in the SCRATCH $pact of F.1uure 
3, StrucEurar repre8enttng new i~puts arc construeted in a 
scratch space (or &pacer) to prevent them from becoming confured 
with the Sy8twnq% Petmr~nt knowledge (recordrd in K8Ir Slnco 
the syrtem unclerrtand8 new l~putt by appaalfng to PY@V~OU~ 
knowledge, thare art many link81 In the form of cmarcs, from the 
SCRATCH Space Into KS, (Note! Only a fragment of KS it Sh6Wn in 
the varlous flouras of this paper,) 
FIGURE 3 PARGE TARGET STRUCTURE FOR "THE-POWER-PLANT OF THE-SUB WAS-BUILT 
BY WESTINGHOUSE" 
The interpretation of the network in the SCRATCH space I8 a8 
follcwr: Node '0' reprerant8 an element of the ret <BUILD> of 
building event8 an wnAcn a #@dgt W built a #%bj PI The agent W 
of thr building event is an element of the rat ot WESTXNGHOUSES. 
The #@obj built by W is Pt an elemant of the art POWER.PLANTS, 
According to node 'H*, thia power plant is the $@subpart in a 
<H-AVE,PART, relrtionahlp in which SI the PartlCular mqmber of 
SUBS currantiy In context, is the #@su~p.art (Wtrp8rt)a 
Discourse analY8Ls mrchani#mS dlrcusred Ln Deutrch [1'975f and, 
rare fully, in Walker rt al, (1975) will br used ta rrraclate W 
with the unique Wertlnphb~rr Cerporatlon known to the semantle 
net in space KS, The other definite NPS ("the sub" and Vthc 
power plant ot the rub9 vwill Llkewira ba r@~~lv~d, 
ToesUpp~$s 8sCOndary details whtle considering the building 
of this structuret rrrum~ the highly airn~llfiad language 
deflaitiont 
Crammrr Irsrricon 
Rit 6 a> NP VP NPr, the-~owar*plrnt, 
R2 a NP a> NP PRCPP thta~~br Westinghou#e 
R3t VP r> VB PREPP VP; was-built 
R4t PREPP 8, WEP NP PREP1 ofr by 
(Note$ wthe-p~~~~~PllntR 18 not treated 4s an NP in the actual 
8Ysttm0 Rsthcr, NOM npowrr plant98 first comb-tncd ~tth PREPP 
rnof the rubw and only afterward is "thew appended to produce the 
NP "the Power plant of the subR,) 
In the trrnalatlon procerrt rpacer are created to reprarent 
the ternantleg of each grammatically defined conrtltuent of tne 
total utterance, These spaces Ire shown In Figure 4 with he8VY 
arrows indicating the space hlararchy, 
SA -3804- 19~ 
FIGURE 4 MULTIPLE SCRATCH SPACES FOR 
"THE-POWER-PLANT OF THE-SUB WAS-BUILT 
BY WES'I'INGHOUSE" 
At the rtrrt of prot~ssing~ oprcr K8 contains knowledga 
about Powst=PlantSt cHAVE,PART> rlrlationships~ SUbmarlne~~ 
<BUILD> eventst mb Westinghousa, On rpottlng the noun phraae 
athe-powrr-plantc, an 8CR ir called to set up a rpaee, NPI, below 
KS tn the partial dtderlng, Withln thfr rpace, a structure is 
crested reprerenting the meaning of Hthe~pawer-~lant~, 
8lmilarlY~ new space8 are bet UP to encad@ the 0thN Sentence 
cbnrtitucntr that correspond to cxpllcit lexLcrl antrlas, 
A8 the ~hrrar gtbu~s rubphrase8 into lakper units, SCRr arc 
~~llcd to aid in the praters, Usinq rule R4, PREPPl (MbyR) and 
NP3 (@WestingMu~c~l are combined to form PREPPl C ''by 
Wcstinuhousew). PRePPl is allu%ccrted it8 gwn $pace, out no new 
rtructures are created rtthln it. 
When ryntaetic considerat ions suggartr combining VPk 
(wwa.r-bul~~tn) with PREPPI, the appropt. &ate SCR it cac.icd, 
Con8ulting a rurface~to~desp*cast map aosoeiattd with the lexlcal 
entry fat he verb wb~lldR, the 8CR dttetmknt8' that I @byw PREPP 
following the verb often rignalr the drrep rrqt ear* Iri a plrsrlvc 
conrtruction, Operating under thir hypothesis. the 8CR checks 
the voice Of VPI, PISSing thir taltr the SCR naxt cheek8 the 
stmantic fearlbiLlty of the NP of PREPP1 serving 4% tne bQagt in 
a eBUIbD> event, To $0 this, the SCR Consult8 the #@d@lincstlon 
of tBU13LD> in 8pQcc KS C14~ Flgute 1). The dclinration 18 
tncoded ra an <IMPLY> 81turtlon in tatma of spacer 86 and Sf, A8 
dlteurrrd crtllerr this drlinertlon indieate8 that any #@rut ot a 
<BUILD% situation wurt be an altwent of LEG&L,PERsONS, The 
candldatr for the #@apt position is W of space NP3. Sinec W I# 
an alcaont of WEbTINGHOUsES and UESTINCWOUSES ir a subsst of 
LEGALcPERSONSt W in accepted, A canstruction such ar wbuilt by 
the rybmrt~ncm would have been rejected. 
Once VP1 and PREPPl have paS8ed the acceptability ta~t~, a 
new #pace, VP2, fs constructed to encode the rarultsnt VP, This 
new rpace link8 node '0' of VPI with node 'W* of WP3 via an 
agt-~~rc , This new arc L8 accerslblo only from space VP2 (and 
lower spaces in the hierarchy) and La not rect~aiblt from either 
VP1 ox NP3, Thtg leaves ths com~onents encoded in VPI and NP3 
free to comblne in alternativtr to VP2 if need be, 
continuing the Parse, NP2 ('tha~SUb"1 ia comblnrd With VFZ 
(wwar=built by We$tinghoure~.) to form, sl, 6fter parrtng trrts 
~imlhr t& thebe above, The obj~alc linking the constituent 
phrart6 of 81 ir aontalnad Ln space Sl and hence Is Inacc~srlble 
from the spaces of the conrtituentr, Notice that t construct 
"the-~ub war-built by WaStlnghoUsaw whlch is encoded bY 81 i# r 
rpurious intrrprttatioa ot uttcrahcr companent8. 
UIinq rule R4r PREP may be eambined with NP2 t~ form 
PREPPZ, Th. n.t*brk itruCt~fe8 ~cc4ssibla from PREpP2 do not 
include the ispurlou8S obj-arc f.rom '13" to '8' that llsr in space 
I 
When the syntax at rule R2 rupgrrtr cunbAning NPI and PREPP2 
to form r new NP (wthe-p~w~r~pl&nt of thr*rubw), an SCR is 
called, This SCR checks NPI to $40 if It is relational In nature 
us Is 'bermH in *beam Of the DolphinV and hence rwprcting an 
argument to b4 8UPPl$,&br Since NPi fall8 this ta%t, the 8CR 
check8 the proprrtirr of the PREP *ofw and d28covet# that it may 
be used to encode <HRVE.PART> rlturtl~ns. Calling upan the 
dellneation of <HAVE.PART> and a~proptiata #urtaea.to-dccp~care 
maprt the ICR detetmlnas thic to be 4 legltlmate interprctrt'ion 
and hence build# space NP4 with a node 'W' and three arc8 at 
shownr Whlls these new Cenrtructt 4t4 4C~ddslbLe from space NPl, 
they ere inrccrrrible from conatltuantr NP1 and PREFP2 (and NP2f. 
Furthermore# they Cannot be accerred fram ~puxlsur #Pace Sit 
hence the construction sf NP4 hag not altered the view of the not 
from 51. 
U$ing rule R1, 52 is canStruct@d from NP4 and VP2, In 
addition to the QbJ-atC contrlned in space 82 it$clft the view of 
the net from 82 includtr all the infarmatlan acCtodibla from 
either tpuce NP4 or rprca VP2 and hence is idantical to the view 
from apace SCRATCH at Figurc 3. Slnca the ~arae mrrsrpon@lng to 
space S 1 doe8 not succ4rrfully account for the fragment 
uthe-power-plant ~f'r it 11 rbje~tadt and S2 is accepted rr 
rxpEb!rring tha meaning of the inputr 
The partial ~rd~ring of spacer fram S2 to KS indlcatrd In 
Figure 4 1s idcntlcrl to thrt rcprraentcd more elearly in ~iqure 
5, which, brcauae of the choler of $pace labclq, may bc 
rrcoani~cd rr the parse tree of the input rentcnce. The syntax 
of the input and the assoeiatlon between each ryntactlc unit and 
its corresponding rcnantlcr has thtrefo?~ been captured in the 
structures built by the SCRs, 
'JP I ?PEP; '\ P3 
FIGURE 5 SPACE HIERARCHY ABOVE 52 
DISCUSSION 
Paxtitlaning is a recent lnnev&tion In rrmrntic nrtworkr, 
Aa shorn lbOV@t this new fe&kuI's enabler networks to maintain 
altt&nat~ve hypothe&e# (a,glt 81 and &21 concerning tha uor of 
Utterance c~n#tltU@ntr and enrblc8 such competing hypotharta ta 
rharc network 8ubprttr (a&?,, VPZ), Wlthout prrtitlonlnu, the 
pack-linked nature af netwotkr caurar a constituent to be rltetsd 
when it 18 incorporatad into & larqer unit and hanca renders it 
unurrblt in e1taa;nctLvc conrtructianr, The highly rmbigu~ur 
naturr of WoU8tiC input mrktr thest ablLit1W to maintain 
Partitleninp also allows rrlcctrd portion6 of a network to 
be arroclattd with ryntretic units, shoring 'the correspondtncc 
bttwtrn network cntitttg and the syntactic strocturrr that *@re 
used to communlcat@ them, A& dlrcussrd In the section on 
rs6ociatlon i6 crUcf81 in analyzlnp the tlllptie uttcrrnctr that 
arc 60 characterlrtic of rpecch, 
American Journal of Computational Linguistics 
Klcroflcno 34 : 451 
S-P S : A FORMALISM FOR SEMANTIC INTERPRETATION AND ITS USE 
NORMAN K. SONDHEIMER AND DoYT PERRY 
Department of Computer and Information Science 
The Ohio State University 
Columbus, Ohio 43210 
Introduction. 
This paper presents a formal ism cal led Semantic Processing 
Scheme, SPS, for use in describing semantic interpreters. SPS is a rule-based 
system with a rule-ordering scheme that can produce deep case structures 
from phrase-structure trees. It was originally developed to demonstrate 
how English prepositions, such as "up", "down", and "through", which refer- 
ence location, motion, and orientation in space could be semantically 
interpreted. This paper presents SPS in its current form and shows how it 
can Landle these prepositions, call ed the locative prepositions. SPS is 
continuing to be used in studies of semantic processing. 
Computational linguistics has seen a considera'ble amount of work on the 
development of general model s for 1 anguage-unders tandi ng sys terns. Among tile 
4 5 7 
most we1 1-known examples of this is the work of Schank , Simmons , Wlnograd , 
839 
and Woods . On the whole, these rnodgls have been tested on broad but 
shallow subsets of ~nglish, in that they have been applied to many different 
phenomena but few extensively. The authors of this paper are taking a 
different approach. We are studying a few phenomena and attempting to allow 
for them in considerable detail. At the least, this approach should lead 
to better treatment of the particular phenomenon. 
It can also lead to the 
development of new general models or the revislon of old ones. 
The paper is written in five sections. The f4rst describer the overall 
Interpretative framework. A second indicates some of the difficulties 
inherent in the processing of locatlve prepositions. An overview of SPS 
Is glven in the third section. The last two sections expand on the SPS 
description and discuss how the locatives are all owed for. 
Syntax, Semantics, and Pragmatics. SPS is developed for a traditional 
three-level system, with syntactic, serna~tic, and pragmatic stages. Based 
on the level of abstractness, these stages compare most closely to 
Yinogrd't and Woods'. 
The syntactic processing stage is assumed to take strings of text and 
produce underlying syntactic structures in the form of cansti tuent 
structure trees. We are attempting to keep these as close to surface 
eonstftuent structures as possible. However, some divergence from the 
surface form is currently assumed. For exampl e, imperatives, Interrogatives , 
and relative clauses are assumed to be shown in a declarative-li ke form, 
and preposi ticns are assumed to have their complement immediately following 
then. 
An SCS based interpreter takes these syntactlc structures and produces 
output whlch ref1 ects underlying semantic structures. The form of the 
semantic structures is also a topic of our research. We are uslng Case 
structures 
7 
*2'4v5 and PI anner-1 l ke assertional forms . It 1s Interesting 
to note that our results to date tend to indicate the need for a level of 
ahgtraction somewhere bet~een Simn 'IS and Schank's semantic nets. 
In developing the semantic lewl-, we are trying to make it the one 
where "general knowledge of language and its relation to the world" is 
applied. This is in contrast to the pragmatic level, where situation- 
specific information is used to interpret the semantic structures. 
In sumnary, a system employing SPS would construct syntactic trees, 
use SPS for the production of Case structures, and employ a pragmatic 
processing scheme to interpret these structures. 
Problems in Processing Locative Prepositions. Part of the problem with the 
semantic interpretation of locatives is the complexity of the structures 
necessary to represent them on the underlying syntactic and semantic levels. 
'This section discusses these problems and introduces our semantic 
structure notation. 
The representation of locative prepositional meaning in Case structures 
has been problematic. The number of cases that Fillmore has postulated for 
them has risen to four--Location, Source, Goal, Path. He a1 so features 
2 
locatives in a paper on problems within Case grammar . The worst of the 
problems involves not being able to interpret the semantic weight o.~ meaning 
of the representation. An example of such a probl em comes in the represen- 
I I 
tation of the following: "Bill held his daughter on his lap in the tunnel. , 
Both of the locati* phrases w~uld be assigned the same case - Location. 
Howeverj they actual ly locate different objects. 
Bi 17 's daughter was said to 
be on his lap while both of them were said to be in the tunnel. 
Similarly, 
the use of an unordered set of cases fails to a1 low for the difference in 
meaning of the following two- sentences, where the first two prepositional 
phrases in each would be in the Path case: 
'!He went down the hi1 1 across 
the bridge to the chapel.", and "He went across the bridge down the hill to 
I I 
the chapel. . 
The Case representation we are using deals with these problems. This 
representation uses only one case for all spattal references. This case, the 
Place - case, identifies spaces which derive from the location of participants 
i.n its action, event, qr state of affairs (or event/state). Which participants 
and how each space relates to them depends on the type of event/state, 
The basic structure of the assertional notation can be seen by showing 
how a Place case wul d be represented: ( :PLACE #E/S $PO). The ": " 
II II 
identifies a relation, the # an event/state, and the "8" objects (note that 
many of these will be replaced by variables in the actual assertions 
produced). The first element of any assertion is always a relation, which 
forces interpretations on the other elenients. With the relation :PLACE, the 
last two elements must be references to an eventlstate and a spatial object 
(space), in that order. The specific spatial objects that are referred in 
Place assertions are call ed Pl ace objects. 
The prepositional elements on the semantic level can relate Place 
objects directly. An example of this is the representation of "She died 
away from where she 1 ived.", i .e., (:PLACE #E/Sl 
$PO1 ) (:AWAYFROM $PO1 $PO?) 
(:PLF\CE iiE/SZ bP(12). here a prepositional element relates the Place 
object of the two event/states corresponding to "she died" and "she 1 ived". 
Prepositional elements can also relate spaces derived from Place objects. 
This is seen with the representation of motional meanings, such as in the 
mu1 tiple Path sentences above. The Place object of "go" and other notional 
event/states are taken as indicating the space traversed by the moving object 
or objects. For the example sentence, the Place object would show the space 
through which the person travelled. This is acceptable since the static 
positioning of these spaces (or paths) as "across" the bridge is logically 
equivalent to his going across it. The predication of derived spaces 
arises in the handling of the ordering problem. The motional Place object 
can be taken as composed of parts that are ordered like the parts of other 
objects (from front to back or top to bottom). The ordering here is based 
on the time the component spaces were occupied. Using relations to select 
segments of the path and the end points of these segments, simple mathe- 
matical relations compare the orderi ng of the component spaces, coirpari ng 
parts of the journey in time. A semantic structure might look 1 i ke the 
foll.wing: 
(:PLACE #XI08 $XI 09) ( :SEG $XI 09 6x1 10) (:SEG $XI09 $11 11 ) 
(:FINAL $XI10 $X112) (:INITIAL $XI11 SX113) (:LE $XI12 $X113). 
The Place case proposal avoids problems 1 i ke that with 
the Location case exarnpl e, through the representation of certain syntacti - 
cally simp1 e clauses with more than one event/state. The representation of 
"He held her on his lap in the tunnel ." shows an event/state corresponding 
to "he held her" and one corresponding to "she was on his lap". These are 
constituents in a causative event/state, with the first causing the second 
*Fillmore roves in this direction in 121 Similarly the representation 
resembles those of Rurnel hart and Norman 3 and schank4. We Wave attempted 
to systematically work out the event/state analysi , as far as it concerns 
locatives, for all verbs taking locative objects. 8 
This complex structure solves the case problems by a1 lowing each preposition 
to predicate a different Place object. "On his lap" predicates afi existen- 
tial event/state showing where the female was located. "In the tunnel " can 
predicate the Place object of the causative event/state. The interpreta- 
tion that space is that it is composed from the Place objects of its two 
constituent eventlstates. Hence, both peopl e wi 11 be predicated by i t. 
While these last two devices enable us to avoid representational 
problems, it should, of course, be remembered that semantic interpretation 
must support these forms.* 
Tied in with semantic complexity is a1 so complexity on the syntactic 
1 eve1 . Assuming sentences are normal ized in underlying syntactic' structures as 
specified, locatives appear in four positions: 
as the qualifier of a head noun 
in o noun phrase; as the compl ement of a copula; as the adjunct to a clause; 
and inside a clause as a locative o,bject. The adjunct usage can be differ- 
entiated from the locative object by its tendency to give overall predication 
to the event or state referenced by the clause. In "He held her on his lap 
in the tunnel .", the first phrase is a locative object and the second is an 
adjunct. 
To summarize this section has presented a variety of points about the 
semantic interpretation of locative pr'eposi tions- that they can require 
complex case representations, and that they appear in a variety of syntactic 
environments. SPS has been designed to relate the syntactic to the semantic 
*There are other phenohena for which the Place case proposal a1 lows. The 
co'mpl ete representation is descri bed el sewhere.6 
What has been given 
here is enough to show the difficulty of interpretation. 
environment of locative prepositions. How it deals with these problems will 
be described after a brief over vie^ of the formal ism, 
SPS. The SPS formalism is mst closely related to a fam'ily of semantic 
interpretatian schemes deriving from Woods' 1968 The close similarity 
to that work 1 ies in the basic form of rules. These rules have the form 
"pattern + action", where the pattern side specifies tests to be made on 
the syntactic structures, and the action side specifies forms to be added 
to the semantic structures. The tests are mainly based on the matching of 
tree fragments against syntactic structures and the testing of semantic 
features associated with those elements matched. In SPS, sets of features 
can be directly examined or compared to other sets of features. Each lexical 
entry may have mu1 tiple sets of features associated with it. SPS a1 so a1 lows 
these tests to be made against features associated with registers by other 
rules. 
If the tests are successful, the action element is executed. This 
principally adds assertional forms to the semantic structure, but can a1 so 
set values of registers. In the assertional forms, means are provided to 
a1 1 ow references to the syntac4ic constituents and 1 exical entries matched, 
as well as to other forms through the registers. 
SPS uses a finite state transition net for ordering the appl ication of 
rules. 
Each noun phrase and sentence is analyzed under the control of a net 
associated with it. The process of forcing interpretation through constituents 
is guided by marking completely interpreted nodes. The overall tree is 
processed from the bottom up. 
SPS Rules and Locative Prepositions. 
To see how SPS works in detail, and ,," 
explain how it allows for locative prepositions we look at a typical rule: 
Rul e 2-STAT-LO: 
((*I-S5 (1 2.3 4) 1(4) *1-S7 
(( EQ #2STAT 1-1) (COMPATIBLE 1-1 2-1) 
(COMPATIBLE 1-2 OBJ(I-i 1) (COMPATIBLE R(SS) SUBJ(I-1 )))I 
======+ 
(((:PLACE R(CAUSED) !X(1 )) (1-1 !X(1) ! X(2)) 
(:PRED !X(3) $BE) (:OBJ !X(3) !l-2) (:PLACE !X(3) !X(2))))) 
This is a rule that might be applied to interpret the prepositional phrase in 
the sentence "He held her on his lap.". 
The rule is identified as 
2-STAT-LO. This particular name indicates that it deals with a preposition 
with a certain static type of meaning (2-STAT) used as a locative object (LO). 
The pattern portion of the rule consists of two parts. Tbe first 
describes the syntactic environment in which it applies, while the second 
gives the semantic feature tests. 
The specification of the syntactic environment is done through reference 
to tree fragments that must be matched in the syntactic structure in order 
for the rule to apply. The reference is made through the asterisk-number- 
dash-1 i teral 
fons in the rule, e.g., 11*1 -S5I1, where the 1 itera1 s identify 
fragments such as the following: 
PROP 
I 
VP 
I 
PP 
/\ 
PREP NP 
I I 
I 
PROP 
VP 
I 
PROP = proposition 
These-fragments would match a locative object use of a prepos~.~lur~ arlu LH~. 
verb of that sentence. Other fragments are needed for other usages. The 
two forms in the rule after the reference to the first tree fragment will be 
described in the ~ext section. 
The second part of the pattern side is a set af triples used to test 
semantic features. These tests are of two types, EQ and COMPATIBLE. The 
EQ or "equal" tests ascertain the presence of a single feature in a set. 
Its 
first parameter is the feature and its second the set. The primary use of 
this test with locatives is to identify the cases where the prepositional 
tree fragment has actually matched a locative use of a preposition, since 
the syntactic parser can only be assumed to identify prepositions and not 
differentiate their senses. SPS allom for this discrimination by providing 
reference to the 1 exical entries associated with a preposition .* These 
references are made thro+ugh the number-dash-number forms where the first 
number refers to the number associated with an occurrence of a tree fragment 
in a rule, while the second refers to the leaf number in the fragment. 
The COMPATIBLE test is meant to allow for the semantic co-occurrence 
restrictions. It takes two sets of features as arguments and evaluates to 
true if the sets share at least one element. The above rule il 'lustrates how 
this test can be used to allow for three types of restrictions affecting 
locatives. These are between a verb and its prepositional object and 
between a preposition and the two elements it relates (Winograd's semantic 
subject and semantic object). 
The fact that SPS allows three sets of features to be associated with 
lexical entries is used for the three restrictions on 1o.catives. One set, 
accessed through number-dash-number , is for restrtc tions placed on the 
*With ambiguous entries, SRS tests each sense individually , therefore, any 
of the lexical references can be considered to have a unique meaning 
at 
any one time. 
preposition by the verb. The other two sets, identified by the OBJ and SUB 
prefixes are for restrictions on the elements related.* 
The final triple in the pattern differs from the othprs in that the test 
is against a register. 
SPS allows for registers that can have sets of 
features associated with them. The registers provide communication between 
rules to a1 low for some contextual effects. Tests may be made against 
registers both before and after they are set, with the test held in abqance 
in the former case. 
The use of the register here is to identify the semantic subjest of the 
preposition. This is necessary since it can not be imediatelv said where 
the subject is situated in the sentence. In the following sentences it is 
initial, median, and final: "He held onto the rope.", "Hk held her on his 
lap. ", and "He held in his hands the letter I sent Mary." 
Given that everything is successful on the pattern side, the action 
side is executed. An example of rule application is given below: 
(:PLACE !X21 !X100) 
PROP (:ON ! xi oa LXI 01 1 
/ \ ( : PRED ! Xi02 $BE 
( : OBJ ! Xi 02 !'x4 
I: PLACE ! XI 02 4 Xl 01 ) 
I 
he 
held 1 
her 
I 
on hi:, lap 
Note that ":PREDt' ide,ntifies the predicator of an event/state, " :OBJH identifies 
Ir 1 II 
the element in the object case, and that the 1 i terals beginning with . are 
*Note that the test using OBJ is on a noun phrase. At the moment SPS takes 
references to noun phrases and sentences to be to the lexical entries of 
their head noun and verb, respectively. 
variables representing some event/states or objects. 
The purpose of the rul e 
is to relate the location of the object being held to the location of the 
complement. 
These locations are available through event/states which 
identify where each of the two objects were. We use the predicator $BE for 
these event/states, such as in the one for "his lap" which is produced by the 
rule. Hbw the correct assertions are produced from the assertional forms is 
illustrated in the above rule. 
All the direct references to relations and objects that start with ":It 
(I tl 
'I#", or $ are inserted directly. The number-dash-number forms provide a 
reference to a 1 iteral stoAd in a lexical entry . For prepositions this 
1 iteral gives the physical relation that the term refers to. 
The two Place objects are formed by the use of a variable generation 
feature using the " !XIt'-number-")" form. References to the $BE event/state 
are also formed in this way. The other event/state is referenced through a 
register. SPS allows registers to hold variable names as well as feature 
sets. The register used here must be set with the variable name used when 
the event/s tate was const~ucted. 
As the above example shows, the registers are used here in situations 
where mare than one event/state results from a clause. When only one event/ 
state exists, a simple reference to the major covlsti tuents of a sentence is 
necessaty. SPS allows for this by automaticall v associating variables with 
the S and NP nodes in trees. These are referenced through forms like "!I-2" 
which here gets the variable associated with "his lap" (presumably !X4). 
This variable will also appear in the assertions describing the object; 
hence co-reference is achieved* 
A facility of SPS missing from the exampie is register setting. Two 
operations can be accompl ished. Either a variable is loaded, or both a 
variable and lexical entry are loaded. 
These registers are essential to the development of the complex 
structures that must be produced at the semantic level. Besides he1 ping 
produce mu1 tip1 e eventlstate structures, they also provide the means for 
ordering the partial predication of a path. 
In any list, the variable identi- 
fying the location of the last mentioned space can be loaded in a register. 
Then with the next phrase on the 1 ist, the variable can be referenced to 
form the comparison. The new f'inal 
value can then belodded in the 
register. 
The Ordering of Rules and Locative Prepositions. The SPS system appl ies its 
rules in a strictly orderea fashion. Major constituents have rules appl'ied 
to them on the basis of an ordering shown by a finite state transition net- 
work. The following is a hypothetical network for ordering the application 
of some rules: 
R1 
+O Initial State 
R4 >@ OFlnalState 
The 1 iterals on the arcs name rules that must be successfully applied be'fore 
a state change can occur. These nets are set up for noun phrase and senten- 
tial elements, and are used with a marking scheme such that interpretation, 
of a constituent is complete only wMn its net is in a final state and all 
its constltwnts are marked as interpreted 
These nets are set up for e,ach head noun or verb 
interpret noun 
phrases and sentences. 
Their utility is in allowing for the orderings among 
case elements. The constituents fill ing semantic rol es in sentences can 
only appear in certain positions with respect to each other. This is 
particularly true with respect to verbs since the roles and orders differ 
from verb to verb. Hence, the net used depends on the head noun or verb. 
There would be no need for a net if the number dP constituents were 
strictly 1 imited. However, with locatives there can be no 1 imi t on the number 
of intermedlate points or on the syccessively finer specification of 
location, e.g., "He lives in New York near the Battery by a park.. .". Nets, 
wlth their ability to loop, are useful for these structures. 
Interpretation proceeds from state to state until success or inabil ity 
to progress further. In the latter case, SPS can back up to the last state 
that still had rules to apply, a fact useful in allowing for pemtnic 
ambiguity. 
Reglster tests have been mentioned as bejng postponed until the register 
is set. It could happen that the register never gets set, e.g., "He hits 
into the stands." does not specify what went ihto the stands. This is a 
case of semantic ellipsis. SPS allows default condit4ons to be associated 
wfth registers that are left te~ted but unset. 
The maans of progresslng through a constl tuent and assurlng its complete 
Interpretation Is provlded by forced anchor! ng and marking schemes embedded 
in the rules. An example of each is seen in the rule shown in the prevlous 
sectlon, i .e., "*I-S5 (1 2 3 4) I(4)". Both schemes refer to nodes in the 
tree fragments uslng a preorder - root first, then subtrees left to rlght. The 
numbers in the parentheses in the example rule refer to nodes of S5. ?he 
anchoring scheme restricts these nodes to being matched to the 7 eftmost unin- 
terpreted nodes in the structure being processed. When a node is prefixed, 
by "I", it and the nodes 
it dominates are marked if the rule succeeds. tjence; 
the example rule marks the prepositional phrase as interpreted. 
Because of 
this marking scheme the noun ~hrases and sentences of a tree are interpreted 
from the bottom up.* 
Conclusion. A formal ism for writing semantic interpreters, SPS, has been 
described, It alfows for a semantic feature scheme that can describe the 
restrictions on locative prepositions. SPS also has registers that can be 
used for these restrictions and for building up the case structures that 
represent the meadings of locatives. A rule-ordering scheme is also 
heloful here. 
It car] be said that SPS is a good vehicle for 
dterpreting locative prepositions, and that any system for semantic 
Interpretation with these features will be able to analyze locatives. We do 
not claim that SPS is a cornpl etely successful semantic interpreter. However, 
the formalism seems to be clear and expressive and it does work for locative pre- 
positions which, to the authors' knowledge, have not been as effectively dealt 
with elsewhere. It could well provide the basis for a uniform, coherent 
structure for semantic interpretation, especial ly for Case analysi-s. The 
authors intend to continue to experiment and develop it as a tool for language 
understanding . 
SPS is implemented in LISP 1.6 on the DECSystem 10. 
%re detail on a somewhat earl ier version of YS can be found in Chapter VII 
of [6]. 
Acknowledgements. Thanks are due to David Brown and Roberto Pardo for their 
reading of the paper and early programming work, respectively. The guidance 
of Richard L. Venezky is a150 grateful lg acknowledged. 
AAI -nKn 
American Journal of Computational Linguistics 
Microfiche 34 : 64 
NICK CERCONE 
Department of Mathematical and Computing Sciences 
Old Dominion University, Norfolk, Virginia 23508 
ABSTRACT 
Varicus representations have been used to partray the ribmiqs of mrd 
(mtably action) mncepts. The mst pruninent of these include decanp3sition 
trees, 1- repreentations such as the Predicate Calculus, and senantic net- 
wrh. The propsition-based semntic ne-rk notation developed by Schubert 
(1974) is especially well suited for including praptic and sepantic Mom- 
tion as part of the meaning representation of irdividual wrd mncepts. The 
attempt is made in this paper to explore the nature of mrd concepts dse mxm- 
ings are represented as senantic netmrks and ,W .investigate their cmpu~tional 
use within the fr-rk of a natural language prpcessing systan. 
The meaning of a cxmce1ps. is explained in lx%rns of other concepts and thmug 
its relatimship to other concepts. Various representations have been used to 
prtray the reanbgs of corrcepts. The mst pranhen~t of these include deccmpo- 
sition trees (Idcuff, 1972; Wilks, 1973), linear represenbtipns such as the Pre 
dicate CalcuL21~ (Sandewall , 1971) , and mtic nemrks . 
Natural langage processing systans can comrenimtly utilize factual. know- 
ledge represented in the form of sa~ntic mmrks. The visual suggestiveness 
of semantic nebmrks aids both" in the forrmrlatian and eypsition 05 the ccmputer 
data structwes~ they resemble. The use of r;~xnantic mtwrks can b fourd in the 
wrks of my authors writing on natural language processing (ir~~loding *hank 
Anderson ard: Bcrwer 1973; ar@ Palm! 1971) as well as other forms of understandhj 
(Wluding Winstmi 1970; and Guzman 1971). 
In utilizing semantic network representations, these authDrs have made use 
of the foll~ characteristics of smtic nets. First (and most important) , 
nsdes that denote the same cmncept are mt duplimtad (in mst cases). It is 
then mible that distinct propsitions my impinge on a 
via arcs. Semrd, 
propsitions are £om& by linking predicate ms to their -t nodes us- 
i.rq arcs. Third, since mncepts are rot necessarily word concepts, particular 
and general concepts are represented as labeled or UnLWed des of a graph. 
Propsitions my also have des associated with then. 
Finally, propsitions in 
a semantic net are rot asslsned to be asserted (@ven though sane researchers treat 
all nodes as implicitly asserted). 
The propsition-based -tic nebork mtation of ScM (19745 is es- 
pecially wdL suited for including pragrmtic and -tic informtion as part of 
the meaning representation of individual mrd concepts. These mardq representa- 
tions are netuorks based on ppsitims that consist of an wary predicate with 
a finite number of agumnts. Terms used in the netmrk to represent a given 
mrd mpt can also be represented by semantic netmrks. 
Thus there is rro in- 
sisw that a given set of "primitives" form the basis for the nemiq of a 
W0rd.l 
The nat section illustrates the use of senantic =rks to represent the 
ws of word mncqb. Subsequent secthris sketch netbds that involve the 
-thd. use of these meaning representations in parsmg ard interpreting 
natural, language text. 
1 
M'mNmG -w FOR wm - 
Cercare (1975) divides kis lexicon Fnto open class itaos and closed class 
itazls. Typically, closed classes have a strictly limited *ship which can- 
not be increased by adding new fomaticms or loanmrds (which are mrds that have 
been bmrpratd by one language fran another language). The significance of 
closed class itens is beat aqxessed by their gram- function. 
contrast, 
open classes have a large, readily increasing rrrmberrihip. New fomtions and 
loahmrds are easily integra-. 
Associated with open class category wrds are meanirq representations: one 
for: each sense of the mrd, The structure of xredng representation is based 
on the smmtic netmrk notation developed by Sch* (1974). Pragmatic and 
semantic information are included in the manhg representation. 
Figures 1 through 6 dmw netmrks that illustrate same of the min senses 
of the word drink, concentrating on action aspects. For illustrative plrFoses 
Figures 1, 3 and 6 are divided into a pragmtic section and a semantic section. 
The pragmatic section indludes the taplate(s) that guides the parse of the utter- 
ance and tm lists: the first list contains proposith that represent the im- 
plications that are fikely to be needed for the ccmprehensicm of subsequent text; 
ard the second list contains propositions representing critical implications that 
we expet xpect mtch in the surface structure. In Figure 1 this first list is (P3) 
and the second list is (Pl,P2), The mtic section oontains the netwrk that 
represents the mankg of the mrd sense. Figures 2, 4, and 5 show various d- 
nal senses of the word. "drink". 
Notice that Figures 1, 3, and 6 aU have the notion of change in contain- 
ment location in ccmmn. 
This rnrre~nds to a general ooncept that subms not 
only differing senses of "drink" but also other more specific concepts as well, 
like "eating" or "receiving an enana". This obsenmtion has led b the follow- 
- consideration. 
When creating the ming representations (netwrks) for concepts it is de- 
sirable to amid the duplication of propositions in storage. If we extractrrore 
general concepts £ran the specific concepts that they sub- (totally or in 
part), we can ayoid duplicatian by associating the crmron propositions with the 
more general concept. 
In a sense the mrk of both Scharak 11972) and Wilks (1973) wrts the am- 
tention that the mmhg of a mpt is best representd by precatjons at the 
highest led. of ga-ierality that adequately explain the term's -. Thus we 
extract frcm "drinking" (and eatbg, etc.) the stmetye shown in Figure 7. 
We might reasonably label the corcept expressed by this structure - "ingestT1. 
It is impDrtant to note, Inever, that while Schank and Wilks might mnclude that 
"ingesting" is a primitive action, that I cansider it a general concept. This 
applies to all primitive actions prt  orw wad of Schank and Wilks. l&amimtion 
of Figure 7 shms clearly that ingesting is -- mt a primitive action but one wbse 
meaning is expressed in tens of causes, mtion, time, and other concepts. 
At this pint the original representations for the various actb senses of 
"drink", i.e., Figures 1, -3, and 6, can be replaced with shplified diagrams 
based on the general concept "ingest". Figure 8 'shrrrtss the representation of 
"drink" expressed in Figure 1 redrawn in terms of the general concept "ingest". 
In similar fashion Fiv 9 diagrams one meaning of "eating", again based on the 
general (mncept "iIlgest". 
The key to making effective use of the meaning representation fok ~~=- 
sicm centers on the propssitbns that contain aqummts that we expect to mtch 
in the sur£ace utterance. The lexical itan for "drinkrt wuld contain, amng 
other things, pointers to a list of pmpsitbns; these propositions contain the 
argments that we arpect to mtCh with mds in the text ard are mst frcqxx&ly 
needed for ccmpmhensicn. At times, bwever, other propositions may be required 
for ampmbmh. For example, the mrd sense illustrated Fn Figure 1 S~-QWS 
that we expcxk to fM, in an utterance atout drinking, an- anim(x) and a liquid (y) 
propoisitims P1 ard P2. But the question can be psd, "What is the effect of 
John's drinking". To answer this question would entail a £urther imtestigatim 
of the other pmpsitions in the ktmrk, especially the first list of implica- 
tions. Altbugh it is %licit in the senantic structure, ~e make explicit in 
the pragmatic structure the inference that "x - drink - y" necessarily implies 
that it causes y's 1.ocation to be - in x at -sane tk after x initiates the drink- 
ing action. Of course, she this implication is cxmmn to all senses of "drink" 
(and eats; inhales, etc. ) it is abstracted into the same general concept "ingest" 
as well, as sbm in Figure 7. 
The -tic structure for each wrd sense for "drinks" is represented as 
properties attach& to the mrd sense. me properties include AW;S, the 
aqmmt list containing aquments used in the mrd sense; IMPLICS, a list of 
implications bt acc(npany *the mrd sense; the propsitions P1, P2, etc. that 
relate the arguments and predicates that make up the netwrk explicating the giv- 
eh wrd sense; and tarrplates of the £om 
argl arg2 ... argi WRD argi+l ... argn 
The implications make the mst camonly used inferences part of the meaning re- 
presentation of word concept. The propositions, for eample P1 and P4 are sb 
Figure 10. See Cercone (1975) for sarnple laical entries, in particular the en- 
try for "drink". 
&my &vantages accrue by represent% meanhg fodas in this way. First, 
unlike Wilks' (1973) meaning fodas, the representation is suggestive of the 
meaning of a word. I see m justification for \(binary) lexical decanposition 
trees as meaning representations for mrds as such trees are neither susqestive 
of the type of processing required nor of the propositions they encoae. 
A seoord and mjor advantage is this. The meaning representaf:~ for a word 
is not required to & explicitly i6 terms of "primitives". Rather, each of the 
predicates in the pmpsith that £om the network repres&ting the meaning of 
the word can, in tum, be represented in an analogous muma. 
In particular the 
notiin of a "cause" seems to me to be m more "primitive1' than "drink". 
This met- 
kd of representing wxd meanings enhances the representational schm for the 
of c~n~rehensb since any munt of detail can be included in the m~an- 
ing representations by addhg propositions to the ne-rks. 
Third, inference ~~, heuristic processing algorithns, and superim 
psed knmledge-oxpnizing schams can be inaqorated US@ this representation 
for mrcl meanings as easily as in, any other representation. IMxmplete infoma- 
tion in surface text can be inferre~3, when necessary, directly frcm the meaning 
representaeon, in scme cases as a missing argument. 
The use of this type ~f mdq repressitation for lexical items is further 
exp1aim3 in the next tw sections. 
111. PARSING AND IXlEKP~ION USING NEIFXlRKS 
~raditionally, the object of pars- sentences has been to outwt syntac- 
tic trees. 
These trees serwd as inpt to scamtic mutines charged with the gen- 
eration of mmiq stn&ures. Whgrad (1972) and kbds (1970) tried, wtth sarre 
degree of suaxss, to integrate the tm processes an3 use each process to guide 
the other process. shank (1972) and Wilks (1973) have stressed that syntactic 
processing was secordary to nwning analysis and should be necessary only when 
the resolution of ambiguity by meaning analysis alone had failed. Utilizing net- 
mrk meaning representatbns the parsing phase is rkmst ccmpletely smtically 
orient&. QE important q-product m the methcd to be described is the detection 
of the correct sense of naninals, modifiers and actions. 
The parsing weeds as follows. Words, in a clause that has been classi- 
f3ed2 are i3camed fmn left to right in search of a suitable cardidat& for an 
action. Once found, the senme is separated into ( (FIRST PAKI?) (ACITON CANDP- 
m) 
(s= PMF) ) . The action candida,p contains, mng other things, a list 
of possible action senses that this particular root fom may have. These senses 
are ordered by a scheme, albeit a very superficial SC~~-E, described in Cermne 
(1975) . Associated with mrd senses are templates as described above. For exam- 
ple, the sense *GIVE1 of the root form "give" has a taplate "X GIVE Y 2" and an 
alternative (ALTERN) tanplate "X GIVE Z TO Y" associated with it. 
The template, e.g.. "X GIVE Y Z" , is used to guide the parsing. In this ex- 
ample XI Y and 2 are variables representing the argUru3nts of the predicate "give" 
that we expect to find in the surface utterance in the given order. hbre detailed 
infomatian concerning the argunwtnts is obtained by exmhing the netmrk proposi- 
tions, for the sense of "givett in question, that involve the arguments. Thus X, 
in this case, wuld represent ap AKtM?SE rvJnindl capable of "giving". 
This is very similar to what Shcank does when parsing in conceptual dependen- 
cy theory. 
If the wrds in the surface utterance do rmt satisfy the constraints 
for arguments of the predicate being examined, it is due to one of four reasons. 
First, alternab syntactic constructions could exist. Secohd, a different sense 
of the qctiorr is "correct". Third, the particular action cadidate in question 
is not the action of the clause. 
Finally, sane other reason, like slang expres- 
sions might be the cause. 
Whenever arguments fail to satisfy predicates, a search for alternative 
inplication templaw begins. 
The result of this search is shclwn quite clearly J 
in Figure 11 of Section IV for tbe ternary predicate "give". 
In that example 
"give" is u&d syntactically in two different £om to distinguish the indirect 
object, one with the preposition TO and one witbut. 
If this approach fails 
then the list of senses for the root £om is fvther examined. 
If other senses 
of the action candidate exist, they are examined further to -e if anymmts of 
the action candidate in the surface me mtch'variables in the tanplate, 
TWs pIocedure is repeated until the mrrqct sense of the action candidate is 
faund or the list of senses is exhausted, 
If th@ sense list is exhausted, scan- 
ning continues in the surface clause for another suitable action Wte and 
the process is Eepeated. 
Part of the process of mtchirq argllpeptS of predicates in &ace text to 
variables in implication tiernplates fnvolves fhxlhg the correct sense of rwminals 
and Wiers as well. sentace "A drin)rer drinks my drinks" has as the 
seam? aqmmt of the predicate "drinksn the wrd "drinks". Possible rwmindl 
senses for that "drinks" linclde an a1mbli.c beverage, a body of water (thrww 
John into the drink) , or a thirst quenchere Thus, if the first sense of 'a d- 
ndl fails as afgment, other senses mst be examhed before dec- not to 
accept it as -t. ms rqp&hg applies with respect to nodifiers in a 
similar kt not identical fashion. For instance, a "yellow cake" is a type or 
cake like a chocolate cake whereas a "yellow car" is sanething that is yd- 
low and something that is n car. Using these ~~, sentences such as "A  drink^ 
-- 
er drinks nary drinks" ar8. "The pilot banked his plane near the river bank wer 
- - - 
the bank that he an for good bpdcing service" present little difficulty. 
- 
Mxpblogicdl analysis is ~~t sin& only tbse £om that tan autplen- 
ticaUy be considered as actims need be exmibed. In the example, "A drinker 
drinks my drinks" the word "drinker" is elimhaW hmdiately as an .action 
cadidate due to mrphological analysis. Thus, we are very qtlickly able to get 
a right &ice for an action axtiidate, 
The next section shows an example of parsing anl the reSulting smtic net- 
work con~trtlclt-Rd -- meaning representations of the type described. 
me - 
72 
The following example is taken fmn Cerc~ne (1975). 
Many other examples can 
be found we. 
Tkae mle lis- preceding Figure 11 gives the results of the 
parsing phase, clause by clause, der the headhg -Ht ASSOCIATED m1m-m 
# R NEW:l!4EmsP 
# 12:23.49 
= (&$mKE 'CMBT) 
- 
m 
= (-1 
- 
RmDY 
- 
=JCMNGaVEJUDYTHF, 
= RED EmK. THEN, JUM! GPVE 
=THEBRtYWNBOOKrn~. 
- 
- 
-1-c-t ASSOCIATED 'ACTION--WRIX&E TRIFLES +ti- 
FT 
- 
 GIVE^ *BOOK1 Z) (*GIVE1 *JUDYl Y) (*GNEl *JOHN1 X)) 
- 
- 
+HxDIFm 
= 
((N CADJ CLASF ((0 0) (*rnl)) )) Z) 
- 
- 
+t+ THE: sEzmMrIcm tt3- 
- 
w* 
= *JOHN1 
*PROPERLYk 
- 
X 
- 
moo01 *mn Y 
- 
PROP0002 WK1 PRED 
- PFOPOO~Z INST0003 ARG 
- 
- 
~0004 INSTOO03 AXIL; 
- 
~0004 -0005 Pm 
PW3POOO1 INST0003 Z 
- 
- 
PFOP0006 -1 PRED 
- 
PIiL1POO06 3&$7!0003 MG 
moo01  GIVE^ PRED 
- 
= 
- 
+.+tASSOCIATH AcqON-v- TEUgLES +ti- 
- 
( (*~m - Y j  G GIVE^ wn Z) (*em1 *my1 X) ) 
- 
- 
-+HMmIFm 
- 
((M - ((0 0) ("B-) 1) Z) 
- 
- 
- 
tts. THE SEM14NTIC NE2 -ti-t 
- 
- 
*A!KM* w* mm* 
~POOO~ 
- 
"JUDY1 
- 
X 
- 
FJFDPOOO~ m PRED 
- 
- 
PE3P0008 ~JSTOOO~ ATI(3 
- miaPOOl0 INSTOO09 IIEil: 
- 
- 
rnP0010 'wNso011 PRED 
- ~~0~0007 INSTOO09 
- 
z 
- 
- 
~~0~0007 '!mRn. Y- 
- PR13POOL2 *BFmNl PRED 
- 
- 
PRDPOO,~~ INSTOO09 AFG 
- PNlP0007 *GNEl PRED 
- 
(W) 
v. ~SIONS 
The above sections outline what I believe to be the wrrect approach to re- 
presenting the meaning cantent of wrd mncepts. Hopfully the use of mmhg 
representations such as these will simplify the problens inherat in representing 
the conceptual amtent of natural language- utterances in terms of meaning struc- 
tures. In prtiCULarr I see the foll~ desirable features inherent in this 
approach. 
(i) Interpretive directness 
The meaning !structures wrrespw to natural. language utterances are 
TO& according to s-le structural rules. mful heuristic criteria, based 
on the central mle of verbs and on preferred -tic catqorieS for the sub- 
jects ad objects of verbs, guide Bach choice ih the creation of meaning struc- 
hnes. 
Interpretation of utterances then takes on a "slot and filler" character, 
rather than requiring extensive trial and error search. 
(ii) Wsis of syntax 
In ordinary discourse it muld, IS absurd not to accept "ungrm&l" 
-11s like -ling participles or fanciful locations such as mtapbr. 
In the abwe awch a syneetic straightjacket is mt bps& on assible 
utterances. Therefore the ahxmnal is mt excluded as it is in my linguistic 
w'tJ=- 
(iii) -sis an events 
A mjor part of our interpretative effort in understanding natural lan- 
guage is focused on events, i. e. r tkne-depen3ent relationships. By contrast, 
"static" ~~ps in the mrld are relatively easy to understand. Therefore 
the 
for lkxxlm&tal senantic structures Wuld mncentrate on the repre- 
sentation of events. The use of meaning representatiO11~ as described above facil- 
lihtes this emphasis on events. 
Phe handlira of vagueness, events, the lexical manb-gs of mnp1-m concepts, 
ard the problem of cwmill knowledge organization may raise additional problans 
when processing natural language with & representations rmch as the ones I 
have used. I-Immm, ths meaning repre~entati~ns UW in this paper can be viewed 
as an extension of several successful but mgerf icially disparate schmata, such 
as Schank's (1972) conceptualizations Qr Winston's (1970) descriptions. This 
indicates that their use should prove of real value in the design of understanding 
Many thanks are due to Dr. Len Schubert; his ideas and amrents are inter- 
- in this research. 
I am also j,nd&ted to Dr. J. R. Smpson @ Dr. K. V. 
Wilson for their dful r&iq ar$l suggestions. 
Notable systans currently in Mgue that utilize "primitives" in this way in- 
clude tbse of Wilks (1973) ap3 Schank :t al. (1973). 
2 Nxds in clauses are rrorpblogically analyzed and, based on that analysis, they 
are classified to determine all of their pssible syntactic funqtions in an 
utterance. 
In WWirogradfs (1972) wrk, "gives" is recognized as a transitive action that 
requires tw, objects : his classification is TRANS2 . 
r & ir 
I.'"$ "b' .-5 
Fig. 3. 
1 
drinks 
I 
7. ' " 
Fig. 6 "(MY at= dvin --. ..- ks (q~5!j I ! .E rr 
7y.0: then 
I JDIR I 
r- '-, , "I 
'I,;, , 
pig. 5. ,$3!I) \yd$?;*, 
(PI, PZJ 
I 
?r~gn;a! i~5 
L 
, emu? ics 
THEM, JUDY DFlVE THE 
BROWN- WOK TO CIARY 
MOTE8 I 
1. ?RU? LIW8 ARE MOT 8HOHW- 
2. DOUbLY-CIRCLED NUDES RUE 
rlEOlCRTlYL COWCEPT MODE8 
IMO P~OPOBITIOW NUDE8 c 
3. TllPLV-CIRCLED WOOLS 8PECIFY 
~RERDY EXISTINO nmEa BUT 
n1O PERSPICUITY- 
4. RRWllEWf8 Of tiUMROIC PREDtCRTE8 
IRE LRBELtEO QROc 
L* AROUKNTS OF NOW-HUNROlC rREOICRTE8 
#UE ORDERED #NO LABELLED X. 'I. ETC 
rigare 1 N-ary Predicate Network 
-tic retmrks present special problet~~ with respect to the use of logi- 
cal amectives, quantifiers, descriptions, rmdalities, and certain other con- 
strvctiaas. $Wq&ert (1974) has propased systematic solutions to these p& 
law by exterding the expressive power of mre or 'less ~omtional samntic net- 
wrk mtation. In this appendix only the elemntary part of the folmalisn, 
namely only as mch as. is necessary to clarify any yisconceptions than may arise 
fmn the figures used in this paper, is explainel. 
In saantic netmrk notation, the distinction between labels designating 
storage locations and labels designating ~inWs to storage locations requires 
clarificatbn. This distinction is used & Quillian (1968) to designate "type 
nodes" (unique storage locatims) verv "token mdes" . The notation can be 
made uniformly explicit as in Figure A.1. Here "part-of", which in sane rota- 
tions corresponds to a token node, designates a type node (as suggested b$ 
Winston, 1970). All mcircled nodes correspond to storage locations andmall 
arrows to addresses of storage locations. What fozm~ly were token nodes are 
mw called--sition nodes; they serve as graphical nuclei for propositions 
as a wble. 
At times the explicit natation- of Figure A.1 will clutter tho diagram lead- 
ing to a loss in readability. Wefore, when the meaning is cle, binary 
predicates will be represented as in Figure A.2 for visual effect with the under- 
standing that the use of explicit propositidns underlie the structure. 
In Figure A.1, A, B, and REL are mere distinguishipg mks. They are ana- 
logous to parenthesis or ccmas in the Predicate Calculus & that they serve to 
relate demthg terms syntactically; they are mn-demtative thenselves. 
When- 
ever possible they will be cbsen to be =-, i.e. to enhance readability 
and be suggestive, but they amid be chosen as numeric labels as well. 
One adwhtage of the explicit.notation of Figure A.1 is that it works for 
n-aq (1172) predicates. 
The sentence "John gives the hk to Mzty" involves 
"gives" as a three place prediate. * 
It is diagram& as in Figure A. 
3 
Figure A.3 is appealing because of the significance we can attach to labels - 
agent, object, and recipient. 3y no means is Figure A.3 a graphical analogue 
of "case-skructured" grmmars. 
Cases are not view& as conceptually primitive 
binary relations as Filhre (1968) and res~ar~hers influenced by him, notably 
Schank (1972), view than. In a case structured system the central ncde would 
denote a specific action or process with the property that it is a "giving" 
and involves John, the book, and bkzy as agent, object, and Pecipient respective- 
ly. Case relations can be understood as ccmplex mnprimitive terms derived 
fm such causally and telmlcgically related sequences of states. The wble 
notion of a case derives frcm the syntactic and sa~ntic similarities bebeen 
the role played by the argurrrents of many predicates. Nevertheless the mtion of 
an "agent" to depend in part on causal priority of a state of the supposed 
agent in the sequence of states mer consideration, and in part on the extent 
to which purpsive behaviour can be ascribed to the supposed agent in general, 
and in part to the extent to which the particular sequence of states which he 
initiated can be assum3 to be intentional on his part. See Cercone and S&ndxrt 
(1974) for a further discussion of cases. 
One f inaL notational pint by way of introduction needs to be made. The 
"camn labels in Fim A.3 are to be med as me mnem~cs, altbugh indi- 
cative of mre canplex relations. To avoid confusion, predicate names will be 
designated "in mall letters and markers by capitals. 
Other conventions that are 
used-include: solid loop for propositional nodes and existentially quantified 
amcept nodes; bmken lwp for universally quantified concept nodes; solid lines 
to the parts of a propsition to a propsition node; dotted lines for 
dependency links joining each existentially quantifiedmde to all universally 
81 
-+.ifid des on which it depends; and broken Lines for logical I-. 
1 PRED 
Fig. A.2 'qlberka is part & Canada 
Fig. A-3. "John gives the book tc Mary. 
8t 

References 
Deutsch, Barbara G, Establlshlng Context in Taakmorlented 
Dimlogg, Presented at the Thirteenth Annual Meeting of the 
Association for Computatlonsl Linuuirtics, Boston, Massachusetts, 
30 October 1 November 1975.
Hcndrlx, Gary G , aemantie Processing for Speech 
Undarrtending , Presented at the Thirteenth Annual Meeting of 
the Association for Computetl~nal LinQuisticsr Bostonr 
MassQchurettfir 36 October - 1 fioV@mbt? 1975.
Ritca, H, Barry, Automatic Speech understanding Systems, 
proceedings, Eleventh Annual IEEE Computer Society 60nf erenee, 
Washington, D,C,r 9-11 September 1975.
~obinronr Jane J, A Tuneable ~erformance Grammar, 
Prtsrnted at the Thirteenth Annual Meeting of the Ars~ciatian for 
Com~Utrtl,ona~ hlngulstlcr, Baotonr MaSSachUSett6r 30 October 1 
NoVsmbtr 1975.
Slacur, Jonrthrq, Speech Ccncratlon from Semantic Nets, 
Prerentad at the Thirteenth Annual Mt4Btlmg of the AsabCiation for 
~omputation.1 Ling\lirticr, ear,tonr Mar~achuaat'tr, 30 ~ctobcir a 1 
November 1975. 
walker, Donald E,, ct al, Speech Underatandlng R~search, 
Annual Report, pralect 3804, Artificial Intelligence Center, 
Strnford R@s@a&ch Institute, Manla Park, CaLlfatniat June 1975.
DeUt~chr B, G, Eatcbltlhing C0ntcxt in Task-Orientrd 
Dirkage a Pc~tenttd at the Thirtecnlh Annurk Beating of thr 
Association tor Canputrtisnai Lfnpuitttci, Borton~ Ma#laehureFtrr 
30 Octebrr * 1 Nsvembar 1975. 
~~ndli~, Grry GI Jearrnt tc Procoa8Lng for Speech 
Under8trnding. Prclantad rt th8 Thtrtaenth Annual Naatlng sf thr 
Asso(S.t$on tot Coaputrttonrl Ltnqultttc8~ Bogton, ~arrachurettat 
36 Oetaber - i November 1975. 
P~xton, Wiltlan Me# and Rsbln86nt Ann E, &Y#~@R Inteutatlan 
and Contra& in c $peach Undsrrtandin~ dyqtra, Prelentrd rt ths 
?hittrentht Annumi M~~tgnq ot thr ~880cl~ti~n for C~m~ut&tLan.l 
tln9uirtfe!#, Bastan, HcrrachUs~rttr~ 30 Qctober 1 Novambsr 1979.
Wrlksrt Dah~ld Eer et rl. 8perch Understrnding Ralrrr~h~ 
Ahnus1 R*p~ttr Projact 3804, ArtLtieP~l tntelliqrnce Canter, 
&tanford Raarrrch Xnstitutr, Hrnla Pvk, Callfornb, June 1915. 
Dcutreh, Barbara C, Estrblishlng Context in Tark~Orlentcd 
Dislogr, Prtrented at thr Thirteenth Annual Heating of the 
Atroclrtlen for Computrtionrl LLngulrtfc8, Barton, Hrrrafhurcttr, 
30 October a 1 Nsvcmb@r 1975.
Xendrlxt Gary G, Expanding the Utillty of Srmrntic Hetworkr 
Thtough Pactltloning, Advance Paperg ct the Fourth Internat tonal 
Joint Conference on Artf tieial ~ntrlllgencr, Tbllltl, ~eergia, 
USSR' 3.8 Bepteabtr 1975, 115-121.
Prxtont Willlam H.1 and Roblnront Ann E, Syrten Integtrtlon 
and Control for Speech Unberrtrnding, Pterented at the 
Thjrteenth Annual Meeting at the AS8oCiation for Coaputrtional 
Lin~uirticr~ Barton, Ha~tachurcttr~ 30 October - 1 Novarnber 1975.
Roblnront Jrnr JI A Tunaible P~rfarmsncc Grammar, 
Prercnted at the Thlrtarnth Annurl Wetting of the Alloelrtton tor 
Computational ~in~bictlcr, Boston, Irrrachurcttc, JO October - I 
November 1975.
Hllk@&r Donald Err at a, Spdrch Undqt8tandlnq Reaeafch, 
Annual Report, Project 3804, Attltlelal Intrlllqcncr Centrr, 
ltrnford Rertrrch ~nititutr, Manla Park, California, June 1975.
'I. Fillmore, Charles J., "The Case for Case" in Universals in Lin uistic 
- *rt Theory, eds., Emnon Bach and Robert T. Har'rns, New York: Ho t 
and Winston, 1968, 1-88. 
2. Fillmore, Charles J., '!gome Problems for Case Grammar", in Working Papers 
in Lin uistics No. 10, 245-265, Columbus: ~e~artient of Linguistics, 
me T?-r- lo tate Un~versity, I- 1971. 
3. Rumelhart, David E. and Donald A. Norman, "Active Semantic Networks as a 
Model of Human Memarv". in Third International Joint Conference on 
Artificial Intel 1 igence, .Advance Papers -- of the Conference, 450-4v, 
Stanford, Cal ifornia, 1973. 
4. Schank, Roger C. , "Identification ofLConceptual 'izations Underlying Natural Language", in Corn uter Models of Thou ht and Language, eds., Roger C. 
+ -&n7~~nC3sco: W. H Freeman and Schank and Kennet M. CWIW.
5. Simnons, Robert F., "Semantic NetNorks: Their Computation and Use for 
Understanding English Sentences", in Corn uter Models of Thou ht and 
Language, eds. , Roger C. Schank and Kk.-F,63&aflran- 
cisco: W. H. Freeman and Company, 1973. 
6. Sondheimer, Norman K., "The Computational Semantics of Locative Prepo- 
sitions", Unpublished Ph.D. Thesis, The University of Wisconsin, 
Madison, 1975. 
7. Wi-nograd, Terry, Understanding Natural Language, New York: Academic 
Press, 1972. 
8. Woods, Will iam A., "Procedural Semantics for a Question-Answering 
Machi ne" , AFIPS Conference 
, 33, (1 968), 457-471 . 
9. Woods, William A., "Progress in Natural language Understanding - An 
Appl i cation to Lunar Geology", AFIPS ConLrence Proceedings , 42, (1973).
Anderson, J. and Bmer, G. (1973). Human Associative mq, V. H. Winstan and 
Sons, MShingtDn, D.C. 
Cesccme, N. (1975). PhB thesis. University of Alberta. Fdmnton, AlbeYta, Canada. 
cercxme, N. and SchuberP), I. K. (1974). "%ward a state-ksed Conceptual Repre- 
sentation", mt of ~tiq science, TR-74-19, University of Mbrta, 
EEbntcm, Alberta. 
I)a*, D. and IEarman, G. (eds) (1972) . Semantics of Natural Language, D. Reidel 
mliskirrg Ccmpany, Boston, ~assa~tts. 
Fillnore, C. J. (1968). "The Case for Case", in Universals in Linguistic Theory, 
E. Bach and R. Harms (eds) , mlt, Rinehart, and Winston, New York, New York, 
GupMn, A. (1971). "Analysis of Curved Line Drhwhgs Using Context and Global 
In£omtiontt, in Machine Intelligence, B. Wltzer d D. Michie (eds) , Ameri- 
can Elsevier Publas my, New York, v 6, pp. 325-375. 
Palme, J. 
(1971) . "Making &x~~~t.ers Understand Natural Language", in Artificial 
J3telli.S- and Heuristic 
Emming I N..Findler and B. Wltzer (eds), 
~merican El~ier Publisww-y, New York, New York, pp. 199-248. 
Wllian, M. (1968). "Semantic Memrry", in Smmtic Information Processing, M. 
Minsky (d) , MIT Press, Cambridge, Massachusetts, 227-279. 
Wlian, M. (1969). 
"The Teachable language CanprMer", CACM, v 12, n 8, 
pp. 459-475. 
# 
Quine, W., (1970). bbrd and Object, MIT Press, Cambridge, Massachusetts. 
Runelhart, D. , Lindsay, P. , an3 Norman, I3, (1972) . "A Process Model for bng Term 
bkmry", in ~rganiiati& of mry, - E. Tulving ard W. Donaldson (eds) , Aca- 
dmic Press, New York, New York, pp. 198-221. 
Schank, R. (1972). "meptual Demw: A mry of Natural wge Under- 
standingt', Cbgnitive P~lagy, v 3, pp. 552-631. 
Schank, R. (1973) . "Identification of GmeptWizatiofls Urderlying Natural 
Language", in Ckqutesr Wels of 'IIYXEght and Lmguag 
w 
e, R. Schiurk anl K. Col- 
by (eds) , W. H. FYecmm ard Capmy", San Francism, Qlifomia. 
Schank, R., (1975). "W mle of Mrrory in Imguage Fmmssing", Yale University 
Tkhdcqd. &port, New Haven, Conn. 
Schank, R., Goldmn, N., Rieger, C., an3 Riesbeck, C. 
(1973). "Margie: ?&r~@y, 
Analysis, kspnse Gmerahn, and Inferme on mlishtl, ~hird- 1ntepmti6nal 
Joint Oonference on Aftificiai ~ntelliger~le, SRI, WO Park, California, 
Sdnhrt, L. (1974) . "Cb the Ekpressive Ad of SaMntic Netsmrks", TR74-18, 
kgartmnt of mting Science, AZberta, E3rmntun, Alberta. 
See also Proc* LJCAI-4. 
Simmns, R. F. , an3 ~ruce, B. C. (1971) . "Scme Relations Between Predicate- Calcu- 
lus and %antic Net Representa~m of Discourse", Pdvarrce Papers of Seoond 
International Joint Conference on Artificial Inwligence, The. British mter 
Society, b&n, Ehgland, pp. 524-530. 
Wilks, Y. (1973) . "Understanding Witbut hwf s" , Third 1nterrbtiona.l Joint Con- 
f- on Artificial 1ntellkg6, SRI, -0 Park, California, pp. 270- 
277. 
W-, T. (19721, Understanding Natural. -ge, manic Press, New York, 
New York, 
Winston, P. (1970) . "- Seal Descriptions fm Examples", PN) Thesis, 
m, m-'E?-76, Cambridge', Massachusetts. 
Ws, W. (1970). "Transition NeMrks for btural Lanwge Analysis:', W, 
v 13, n 10, pp. 591-606. 
