American Journal of Computational Linguistics Mi crof i C~Q 2 7 
A FORMAL PSYCHOLINGUISTIC MODEL OF 
SENTENCE COMPREHENSION 
PETER REIMOLD 
Copyri ht @ 1975 
Association for Computational Linguistics 
This paper outlines a psychologically constrained theory of 
sentence comprehension The most prominent features of the theory 
are that: (1) syntactic structure is discarded clause by clause 
(where the traditional notion of clause is modified in certain 
respects so as to conform to short term memory requirements); 
(2) the syntactic and semantic processor work in parallel. 
The semantic analysis proceeds from the preliminary semantic 
representation (PSR) via the intermediate SR (ISR) to the final 
SR (FSR), making crucial use OX an encyclopedia which codes se- 
mantic knowledge. 
The three stages of the semantic analysis are discussed. 
Concatenation Rules establish the PSR, Meaning Rules and Ency- 
clopedic Rules the ISR, and Semantic Linking Strategies the FSR. 
At every stage, the semantic representations are in terms of a 
modified predicate ca~culus notation. 
Syntax-free as well as syntax-sensitive Linking Strategies 
are presented for clause-internal linking. Finally, syntax-free 
linking of constituent clauses of complex sentences is described. 
TABLE OF CONTENTS 
I l SOW3 PSYCHOLOGICAL CONSTRAINTS ON SENTENCE 
COMPRXHENSTQN MODZLS a...amommamo*.amoe.oemaoaoamom 
4 
11. TRADITIONAL LINGUISTIC APPROACHES rmmm.aamoammmamem 
11 
LII l A THREE-STAGE THEOXY OF SENANTIC ANALYSIS l l l a a . a l 
13 
IV A BIODIFIEB PREDICATE CALCULUS NOTATION FOR 
SSdANTIC REPRESENTATIONS mm.m.maeeo.ameoaoom...mamm 
15 
Vm THE PSRB CONCATENATION RULES .o~m.**eaoaaeme*.mmomm 17 
VI. THE ISR: SEMANTIC KNOWLEDGE RULES aoemmaaoaoeme*aoa 19 
VTI THE -FSR t SEMANTIC LINKING STRATEGIES l l l 6 l a l l l a 3 22 
VIIIe DIFFERENT MODES OF PROCESSING a a l a l a moo l a em or l a l rn ma 
30 
IXo "ALTERNATIVE LINKING STRATEGIES'' ..aaoaaoaaao~mamma 33 
Xm CLAUSE-TOWCLAUSE LINKING mm*aoao*o+mom.m--*---*--.. 35 
XI THE "TIDIPORAL SEQUENCE STRAT=YW l I, . l l l a a l l a l a a . 
38 
4 
I . SOME PSYCHOLOGICAL CONSTRAINTS 
ON SENTENCE COMPREHENSION MODELS 
In this paper I consider the question of how an automatic 
sentence recognizer would have to look in order to be compatible 
wLth present psycholinguistic knowledge about speech comprehmh 
The basic premise is that psycholinguistic considerations are of 
potential interest to computational theories (see, e.g., 
~chank(1972) ) 
Let me begin by summarizing some characteristics of speeck 
processing which we know either from experiments, or which are 
intuitively clear. 
First, there is some evidence that the clause is a unit of 
processing. For instance, ~aplan(1972) showed that after 
a clause boundary is passed, the cons-bitaent words bf the 
completed clause are relatively inaccesai'ble, as measured by 
wmd recognition latency. The effect was independent of the 
serial position of the word for which recognition time was 
tested. This suggests that sentences are processed clause by 
clause, with only the semantic content regularly retained after 
the clause boundary is passed. The surface words (and g fortiori 
the syntactic structure) of the clause would tend to be erased 
after each clause boundary. 1 
*Thie paper is based on chapter VII of my doctoral dissertation 
(~eimold (forthcoming) ) . I wish to thank Thomas G . Bever , 
Jame-a Higginbotham, and D *Terence Langendoen for helpful 
suggestions. 
l~he fortioriJ refers to the fact that the syntactic structure 
Another study supporting the clause as unit of proceasing 
is Abrams 6 ~ever(1969). These authors found that reaction time 
to short bursts of noise "clicks") superimposed on sentences 
was longer for clause-final clicks than for clause-initial 
ones. This would point to the clause as unit of perception, 
under the assumption that processing is more intensive towards 
the end of a perce~tual unit, and that reaction time to external 
stimuli is a valid ineicator of the intensity of internal 
procesging.  or a review of other studies in support of the 
clausal processing theory, the reader is referred to 
Podor, Bever & ~arre%t(1974), where arguments are also given 
for the clause as a decision point across which ambiguitiis 
are ,normally at least, not carried-. ) 
Secondly, it seems that as we listen to speech, we 
simultaneously have access to both the syntactic and semantic 
propertiee of what we hear. That is, there appears to be ~arallel 
orocessing of the syntax and the semantics of a clause. One 
finding explained by this assumption is that so-called 
"irreversible" passive sentences like (1) are perceptually no 
more complex thnn their active counterparts (the air1 ~icked 
the rlower,in this case). By contrast, 'reversiblew passives 
like (2) take longer to verify vis-a-vis pictures than the 
~orre~pondin$ active sentences (~lobin(1.966) ) . 
presumably contain8 surface wards as terminal nodes. Hence if 
the eyntax were regularly preserved the surface words should 
remain easily accessible, too. 
(1) The flowe~ was picked by the girl. (irreversible) 
(2) The boy was kicked by the girl. (reversible) 
It appears that the syntactic complexity introduced by 
the passive construction is somehow circumvented by a 
predominantly semantic method of analysis in the case of 
i&reversible passives m2 We thus get a picture of speech 
processing as in ~i~~l.3 
yntactic pr-l $ syntactic structure of .S 1 
input I' 
[string SI 
semantic ~rza-----d semantic structure of s 1 
Fig.1: Parallel Processing Model 
2~nother experimental study su parting the Parallel. Processing 
theory is larslen-lilson(l973 !' . 
3~y view of the role of syntax is related to that expressed in 
~chank(1972), who believes that the function of syntax is 
""as a pointer to semantic information rather than as a first 
step to aemantic analysis?' (~~555) Similarly, winograd (1971) 
allows parallel operation of syntactic and semantic analysis. 
However, me syntactic and semantic processor in Winograd's 
systehr have full power, in minciple, to question each other 
about their respective success before proceeding with their 
part pf the analysis* This powerful device has been aeverely 
restricted in the theory described here (for details, see 
Reimold (f orthc oming) ) The main reasons for thie are the 
greater reliance, in my theory, on "syntax-free" semantic 
interpretation, and the generally shorter life-apan of 
syntactic structure (see the discussion of "peripheral 
clauses" below). Woods (1973) also discussee a system with 
certain facilities for parallel processing, for instance, 
the "Selnctlve Modifier Placement" facility (sMP)~ The 
function of SMP is to ~elsct from the list of syn%actically 
adrnissibLe alternatives the one which is semantically most 
appropriate, and return only that alternative to the parser 
before golng an *o analyze the rest of the sentence. The most 
important difference between Woods' proposal and the one 
presented here is that his semantic processor only chooaes 
among tact tructQrsd alternatives (and, in that 
sense,% a fu%$%$tax-sensltve method), whereas my theory 
partdates rn gyntactfclm men wh modifiers and their heads. 
Let me return now to the principle of clause-by-clause 
processing, If we assume that "imqediate processing" takes 
place in short term memory, then we must automatically 
requkre that the unit of processing must not exceed the known 
Limits of short term memory. Now since that limit is generally 
taken to be about 5 wordsp the clause-by-clause principle 
cannot be literally true. For instance. (3) lists some "clauses" 
longer than 5 words. It seems to me, therefore, that we have 
to revise the traditional concept of olause* 
(3)a) John and Bill and Otto stroked and huaaed the goat am 
the goose. 
b) The man with the dog with the c~llar,with the bell 
Laughed, 
c) John met his friends yelafzerdau morning around ten o'clock 
ih .a liCtle cafe near DL-U~UW~Y. 
Z propose to take the underlined phrases in (3) out of the 
sentence proper and process them as if they were separate 
clauses. That is, f draw a distinction between the "nucLearH 
clause and "peripheralw clauses. The non-underlined 'portions 
in (3) are nuclear clauses. Peripheral clauses include: 
Prep-clauses ("with the collar"), Comparison-clauses ("than 
the old colonelw). Post-clauses ("yesterday, 'I "around ten otclo~ 
'in a little cafe"), and Coordinate-clauses ("and Bill," "and 
hugged") 
This treatment of certain phrases as peripheral clauses 
seems plausible too, if we consider that "adnominal" Prep- 
phrases, for instance, are semantically like relative clauses, 
as shown in (4), and that adverbs are parallel to certain 
"adverbialU clauses, as indicated in (5). 
(4)~qirl { 
with a green hat 
who wore a green hat 
greeted John. 
5) John ate the cake 
C 
afterwards . 
aster the guests left. 
Zvldently, witha green hat in (4) is related to who wore 
a meen hat, and the adverb afterwards in (5) can be replaced 
by fill1 adverbial clauses like after the m1es.t~ left. 
We are presently testing the validity of this notion of 
peripheral clause. e use snntence pairs like (6a-b) I 
(6a) The officer threatened to give the woman * a +icket.b&e- 
internal position of click "*") 
(6b) The officer threatened to fine the woman * without B 
a license. (clavse-final position of click "*") 
Y 
Our goal is to determine, mine a 'click detection paradien, 
whether or not there is a "clause boundary effect" before the 
final peripheral clause witho7l-t a License in (6b) * Notice that 
according to my hypothesis, there is a clause boundary after 
woman in (6b), but not in (6a) . It has been shown in n humber 
of studies that clause boundaries (but not phra~e boundaries, 
in general) have certain measurable behavioral effectn Crf. the 
review in Focbr, Bever L ~arrett(l974))t so this should apply 
here too, ff peripheral clauses are indeed psychologically 
reab clauses 
'Now, the last principle-I want to discass is that in 
understanding an utterance, people mage creative use of their 
knowledge -about the warId. 
4 
For instance, if I only hear 
you say* 
(7,) 
The cat just caught a -- 
I can immediately guess that the last word was something like 
bird or DDUW SimlLar-ly, if you say: 
(8) Put the freezer in the turkey. 
I know that you really meant "put the turkey in the freezer," 
4 
This general point has been made, 
in one fortn pr another, by 
many authors. For instance, Winograd~1971) notes that correct 
understanding of they in *The city councilmen refused to give 
the women a permit for a demonstration because 3;he-y. feared 
violezkmm and "The ckky coundlmen refused to give the women 
a permit for a denaonst6ation because they advocated revalution" 
needs the "information and reasoning power to realize that 
city councf lmen are usually staunch advocates of law and order, 
btlt are hardly likely to be revolutionaries."(p~ll) Similarly, 
~chank(19-72) envisages a theory of natural language under- 
etanding whlch."has a conceptual base that coneiate of a 
formal etructurew and "can make predictfons on the basis 
of this conceptual structurew( ,556) The principal differences 
between these approaches and m f ne have to do with (1) the 
of the stored semantic information (PLANNER and "conceptual 
case networkn representations vsc predicate calculus 
representations) and (2) the proposed ~ccem ~chan5s~ to 
thia infobmation. Schank's theory relies on lexLcal decornpoeLtion, 
while I uae the "meaning postulates" methodr Winograd opts for 
a broad procedural approach, representing "knowledge in the 
form of praceuuree rather than tables of rules or lists of 
pattern~~"(p.21) By contrast, my proposal remaine closer to 
the traditional WdeclarativeH approach, as will become clear 
because J. Know someznlng aDout turkeys and freezers. No model 
excluding the possibility af matching speech against stored 
knowledge of the world can explain such facts. In this 
connection, consider also the sentences in (911 
(9a) I'm leaving the dbor open so I won't forget to wina &=p 
(it= the clock-- there was no previous mention of a clock 
in the dialogue, but the speaker was looking at a 
Grandfather clock with open door) 
(9b) They published Wodehouse immediately he came over. 
(= published books written by Wodehouse) 
(9c) Italy was sitting in the first row, and France in the 
second. (= peo~le from Italy and France) 
(9d) We'd better put in 20 minutes. (= money for 20 minutes -- 
speaking about a parking meter) 
(9e) He's sitting bv his olate that isn't there.(= by where 
he wishes* his plate were, by his   late in his wish-world - 
epeaking of a cat) 
These sentences can a11 be understood without difficultyr 
and the way we understand them is by using our general semantic 
knowledge. 
What this means, then, is that the comprehension model 
needs to incorporate an encyclopedia which somehow codes 
semantic knowledge. 
In slm, to be compatible with the psychological model, the 
al~tomatic sentence recognizer should have the following 
properties 8 
(1) it should be a clause-by-clause processo2, where my notion 
of "-lause" includes some things traditionally regarded as 
phrases; as soon as the interpretation of a clause is 
completed, its syntactic structure is erased; 
(2) there should be parallel syntactic and semantic processing 
of each clause1 
(3) the recognizer must make systematic use of an encyclopedia 
which codes knowledge about the world* 
IIt TZADITIONAL LINGUISTIC APPROACHES 
Putting together the above observations, one can already 
see that current linguistic theories are not very helpful for 
the eolution of our problem. Fnr inetance, linguistic theory 
would claim that sentence (10) has the syntactic structwe in 
(11)~ which then undergoes various syntactic transformations 
until it is finally mapped onto its appropriate semantic 
s true ture . 
The rrentence recognizers moat directly meetfng this descriptior 
are robabl those developed by Stanley Petrick (see Petrick 
(~969, 197)y) . Wlth some modlf icatione, however, this 
deecription also fits the theories preeerited in Winograd (1971) 
and Woode (1973). Whkle these syetema are feature-manipulating 
rather than transformational, they nonetheless assume that the 
life-span of syntax extends over an entire sentence, and they 
make crucial use of integrated syntactic structures for 
corn lax sentences  or instan~e, winograd (1971) presents 
an f ntegrated syntactic structure for the sentence "Pick up 
anything green, at least three of the blocker and either 
a box or a sphere which is bigger than any brick on the table. 
Recagnlzere ueing an inverse "Geherative Se'lllantics" grammar 
would aldo fall under this deecriptioh. 
(10) The man with the beard claimed fiercely that he was 
innocent. 
fiercely that h~ ent 
But in the vLew I have just sketched, sentence (10) never 
has any integrated syntactic structure like (11). Instead, as 
shown in (121, the string with the beard, for instance, is 
processed as a separate clause, and as soon as. its meaning 
has been extracted and added as qualifier to the preceding 
noun phrase the man, its syntactic structure is erased. 
(12) 4 successive "perceptual crrruses" for sentence (10)r 
[claimed MVB] 
Lq ma< 
Similarly, the "Post+ Lause" fiercely and the entire 
complement clause that he was innocent must be linked to 
the main clause without referring to the syntactic struc~ure 
of the latter, which is assumed to be erased as soon as the 
word claimed has been semantically integrate& This would 
seem to be a more economical procedure, because it minimizes 
the size of the syntactic ballast that has to be carried 
along. C~mparsr for instance, the size of the chunk in (11) 
to the size 0.4 the little chunks in (12). 
Secondly, transformational grammar is hardly compatible 
with the principle of parallel Processins of the eyntax and 
semantics of a @lausem The reason is that accdrding to 
transformational grammar, the syntactic analysis precedes and 
determines the semaqtic analysis. By contrast, Parallel 
Processing means that at least some of the semantic inter- 
pretation rules must be syntax-free. 
111 a A THREE-STAGE THEORY OF SEMANTIC ANALYSIS 
Lez US reT;urn Lor a moment to rlg.l. That rlgure contained 
a box labelled gyn*actic pracew, and another box labelled 
semantic proceseoP. As I have stated, these components cannot 
be identified with the eyntactic and semantic components of 
current transformational grammar . The syntactic processor will 
not be discussed in detail here (see ~eimold~forthcoming) for 
a fuller diacuesiori). It is a predictive parser using d-ency 
* 
notation. There are-no syptactic transformations at all, but 
the output is a simple surface tree for each clbause, with 
certain nodes marked by functional features like SuBJect, 
OBJ1, OBJ2, or MainVerB. The trees m (-12) above are exarpples. 
For the remainder, let me concentrate on the semantic box. 
I suggest that there are three stapes in the semantic analysis. 
as shown in FigS, namely a preliminary, intermediate, and 
f.inal semantic representatton (PSR, ISR, and FSR). 
Preliminary SR f-$~~terrnediatew Final SR I 
c onc at enati on Meaning Rules & 
(Rules1 '-Encyclopedic Rules ' 
Fig.24 +Stage Model of Semantic Analysis 
The PSIi corresponds to a simple comblnatip~~ of the lexical 
meaninsrs of the words. Clearly, as we hear the words in a 
sentence, we immediately grasp their indi4idual meaning, even 
though we may not be sure yet how they fit together. Thi-s' then 
f s the preliminary SR. 
But we als~ immediately have access to some of the 
immJ,ications of the words and phrases. For instance, if I hear 
"catw I immediately also know wanlmal.w Adding such implications 
derives the intermediate SR from -the PSRe 
The final SR is lake the preliminary one, except that the 
appropriate semantic roleshave been assigned to all the 
conetLtuents. An example for the three stages is given in (13). 
(13) The boy laughed. 
PSR r (THEX 1 BOYX) (By) (E t.: PAST t ) 
I: LAUGHY~ 3 
ISRI 
(IIIEX~BOYX &HUMAN% & ~ADULTX m0*}(3~)(3 t~ 
PAST t & -FUTUaZ t 
* . ) C LAUGHyt & HUMANY & 
ANIKATEY~ & BLNFyt .a 2 
(simolif ied to a (~Ibr BOYX) (E t t PAST t) 
C LAUGHxt 3 ) 
Before translatAmg the structures in (13) into English 
let me remark on the form of semantic representations. 
33: A MODIFIED PREDICATE CALCULUS NOTATION 
$OR SmANTIC REPRESENTATIONS 
Bgch semantic representation consists of a number of 
mef iaeq- and a matrix, where the prewf ixes oorrespona roughly 
to the noun phrased of the sentence, and the matrix to the 
main predicate. For easier reference, I have marked this 
distinction in the text by always enclosing the matrix in 
square bracke%s * c f " . 
6 
For instance, in (13) there are three prefixes, and the 
matrix i& LAUGHVYI. 
Each prefix consists of a quantifier (e.g., THE, g --which 
reade "there is at least onew-- or ALL), followed by a 
variable g &,y,=,&,g --represented by lower case letters 
6 
The lfnear notatLon used throughout here is an abbreviation 
defined over dependency stru~tures . For details, see 
~e~mol4 (forthcoming), where definitions are also given for 
tranelating them structures into standard predicate calculus. 
in the examples), and optionally followed by a backmounded 
proposition. Backgrounded propositions are the expressions to 
the rl&t of the colon within the prefixes. For instance, the 
first prefix in (13) contains the quantifier THE, the variable 
&@ and a backgrounded pl?oposition BO'lhc, and the entire prefix 
is readt "The entity x such mat x is &.boy. #I 
We can now translate the structures 111 (13) into English. 
The first, Lee, the preliminary SIi, sayst 
"%e x such that x. is a boy is invowed in somE 
vent suoh that there is some y and some time 
which is PAST, and .y is laughing at time t." 
Notice that this only asserts that the boy is somehow 
,invo&%ed in this, but it does not specify just h.0~. But 
in order to describe what the ligtener actually understand$ 
when hearing %he bov lauded, we must of course specify 
which role the boy plays in this event. 
Now, looking at the final SH in (13). it can be seen that 
it is like the PS3, except that it also contains a role 
pssiment (or $ink, as I will caL1 it), namely x=y:. That is, 
5, the boy, playe the role of y, who was the one who did the 
laughing. By executing this equation x-Y, we can of course 
simplify the representation, which gives us the last line 
in (13) 
The intermediate SR in (13), furthermore, is like the 
preliminary SR, but in addition contains certain implications 
of the words. Thus we have: "the x such that x is a boy and 
(by implication) human and not adult. etc . * And in the matrix 
of the fSR we get "y laughs at time t and, by implication, 
y is hman and anima-te and alive-at-time-t.* In other words, 
one cann.ot laugh unless one is human and alive 
Vs THE PSR J CONCUENATION RULES 
Let us return wain to Fig.2. It shows three different 
blocks of rules which are responsible for deriving the three 
stages of the semantic analysis, namely, Concatenation Rules, 
Meaning Rules and Encyclopedic Rules (collectively ~ef erred 
to as Semantic Knowledge ~ules), and finally Semantic Linking 
Strategies. They will occupy ue fn this order 
The Concatenation Rules take the semantic definition of 
the most recent input word and add it to the current 
preliminary semantic structure. For instance, (14) lists the 
semantic definitions (namely-for the, b6v.and laughed) which 
are relevant for the example in (13) above. 
7~ have made the simplifying assumption that there are lower- 
level components providing the syntactic an& semantic components 
with a lexically analyzed input string* This, of course, is 
almost certainly incorrect, and should be refined by making 
the matching procaea partly top-downi(1n the case of the 
syntax 
this has been done to a certain extent, since it is.based on 
s predictive analyzer. It has not yet been done for the 
eemantioel but it seems that it can be bui t into the present 
s stem relative1 easj.1 .) See Nash-Webber 1974) for further 
P P 
t 
drscuseion.eepec ally h s description of the SPEECHLIS system. 
(14) (a) rthe DDI a (THSV~ --) re-1 
(b) ~~OST ~fl I (EX) EBOYXI 
(c) Elaurrhed MVB PAST], (~'y) (E tr PAST t) L LAUG~~~ 1 
Notice that each of the deftnitions consists again of a 
prefix and a matrix. There are two Concatenation Rules, 
namely Joining and Ba~kmoundinq~ They are sta-t;ed in 
abbreviated form in (15) and (~6). and are illustrated in (17). 
(15) Joir.;_?gc 
Let (x) tM1l be the current preliminary SR, and 
(Y) t M2 3 the semantic definition of the last input 
woPd (which may pot be   art of an NP), where (x) and (Y) 
are the prefixes , and ml'J and tM23 the matrixes Then form 
(K)(Y) 
[ MI & M23 
(16 ) Backeroundin~: 
Let 
{QTFV~ (x) (hll)) be the partial 9R for the current 
NP. Then join the semantic definition (Y) [P121 
of the 
last inp~~t word (if it is part of this NP) as in 
{PTR~:(x)(Y)( MI Rc M2 )I 
(17) Joinin& and packarwndinq applied to (13-14): 
Start r (THEVI --) t--3 
Backgrounding: (THEV~~E x) LBOYx'] ) c--l 
Joining: (THEV-:(E x) ~BOYX])(E y)(~ t:PAST t)[~~ud~yt] 
I will not discuss this aspect of the theory in great detail 
here. No+e only that the only syntactic information needed for 
Concatenhtion is whether or not the input is part of an YP. 
Othexwiee the semantic definitions of the words ar8 added 
from left to rF&t, prefixes behind prefixes, and matrix 
be ind matrf x. 
8 
VIt THE ISR I SWANTIC KNOWLEDGE RULES 
Going baqk to Pig.2, the next step was the intermediate SR, 
which is ilerived from the preliminaPy SR by applying Semantic 
Knowledge Rules. namely Meaning Rules and Encyclopedic Rules. 
Meaning Rules deal with strict implicatioli, while 3ncyclopedic 
Rules are typically probabilistic. For instance, Keanfng 3ules 
tell us that if somebody is a baker he must also be human and 
hence animate and hence concrete, etc ; while Encyclopedic 
Rules tell us that he - tends to wear white clothes, tends to 
sell bread to people, and similar facts. 9 
f had stated earlier that speech involves using one's 
knowledge about the world. There are two separate problems 
with this kind ~f semantic knowledge: (1) how to code it! 
(2) how to retrieve it. 
Coacerning the finst problem, I assume encyclopedic 
8~n the present version of thd theory, the semantic processor 
accept6 the aemantic definition of a word only if that word 
has matched the syntactic predictions. It is probably desirahls 
to liberalise this procedure so as to handle ungrammatical 
86QU6n216BB r 
9 
The distinction between M-rule8 and E-rules is akin to Katz& 
~odor'~-(1963) distinction between semantic markers and 
giktinauis~, the main difference being that Kate & Fodor 
information to be coded essentially in the same form as the 
semantic representations themselves. This makes it easy to 
transfer to the encyclopedia informatior received in the 
cutrent dialogue. Such "active" information would be added 
continuously to the "situational chapter" of the $ncyclopedia, 
where this chapter is thought of a3 containing The linguistic 
and non-linguistic context of the current utterance. 
Concerning the retrieval problem, it is clear that the 
information in the encyclopedia must be extracted selectively. 
The solution chosen here is the one characteristik ~f networks 
Each Knowledge Rule is given an address, and each lexical 
entry, as well as each Knowledga Rule, includes pointers to 
the address of some other relevant Knowledge Rules; Only tnose 
gules are called up which are associated with the sentence 
constituents through some pointer. 
In the case of Meaning 2ules this restriction seems 
sufficient, because there are only few for each lexi-cal entry. 
Not so in the case of Encyclopedic Rules. For instance, thebe 
are all kinds of things I know about bakers --say, that they 
tend to wear white clothes for work--, but most of which are 
irrelevant for understanding and verifying the sentence1 
(18) The bay was sold a nice cake by-the baker. 
used features whereas I use meaning postulates@ For practical 
purposes, the most important aspect 'af the distinction is 
that M-rlxles are applied obligatorily, while E-rules are 
applied &electively, according to the intersection technique* 
(see the discussion below.) Fot: a discussion and critique of 
the ''featurelln approach, see Weinreich(l966). 
For instance, the sentence is perfectly true even if the 
baker happened to be wearing a f lremanB s uniform while 
selling the cake to the boy. 
A good way of restricting the number of Knowledge Rules 
called up for a given sentence seems the Intersection 
Strategy (cf. ~uillian(1969)) in (19) . 
(19) Intersection Stratemc 
If a clause contains two different constituents A and 3 
both pointing to the same encJIclopedic rule E, then call Be 
This rule says to call up only those Encyclopedic iiules 
which are associated wkth a% least two constituents of the 
sentence. For instance, in (20) 1 
(20) This bread was eo'ld to John by the Italian baker. 
baker. sell, and bread all point to the same encyclopedic 
"patternw (21), which states that bakers typically sell 
baked goods to people4 
Hence, the Intersection Strategy Would call up this pattern, 
which could then be uaed to help interpret the sentence. But 
the rule specifying that bakers typically wear white clothes 
for work would not be called up for sentence (20)~ because 
only the constituent baker in the sentence would point to 
this particular pattern. 
VIIt TAE FSAr SmANTIC LINKING ST2ATEGIES 
Let me return once more to Fig.2. The last~CIle-block in 
this figure was labelled Semantic Linking Strstegies. They 
are responsible forrderiving the final Sii, by assigning the 
appropri&tte semantic roles to various parts of the clauim. 
There are two aspects to this# clause-internal linking, and 
clause-to-clause linking. I would like to discuss clause- 
internal linking first. 
It seems to me that when we listen to speech, we have 
a choice of how much attention we pay to syntactic details. 
This is what Parallel Processing is all about, and it means 
that thsre are syntax-frse linking strztegies besides syntax- 
sensitive ones. (22) &ves a synopsis of tba linking strategies. 
(22) Semantfc Linking Strate~iees 
1. Linking by Variable Type 4. Canonical Order Strategy 
2. Pattern Matching 5. Syntax-sensitive 2ule 
3 . Contradiction Zlimination 6. "Alternative Linking" 
Types 1, 2, and 3 are syntax-fre8. Type 4, the "Cxnanical 
Order" stratecy (cf. ~ever(1770)) relies on the shallowest 
aspect of syntax, namely simple linear order of the major 
clalrse consitllents. Type 5 is nensi-f;ive to "functional" featllres 
occurring in the syntactic surface tree. The 6th type, 
"Alternative Linking Strategies," will be explained later on; 
they, too, are pi~rely semantic, Lh-gugh they apply only after 
the syntax-sensitive rule8 have applied (and failed). 
Nbw, within this system of linking strategies, there see3 
to be different levels of detail. At the shallowest level we 
have Linking by Variable Type. Pattern icatching requires 
more semantic detail 1 and Contradfc~tion Elimination is sti 11 
more thorough. Furthermore, thp syntax-free strategies rnay 
be assumed to be simpler than the syntax-sensitive ones, 
because the latter must keep track of two separate structures 
namely the semantic and the syntactic structure. I assume 
that the strategies are ordered according to their relative 
simplicity, which would give us the order as listed in (22) 
Furthermore, I assume that once an acceptable reading 
has been derived for a clause, application of further 
strategies in the hierarchy becomes o~tiont&. This of course 
is subject to empirical tests. For instance, Pattern Katching 
would interpret the sentence: 
(23) The baker was sold some stale bread by the butcher. 
incorrectly as "the baker sold the butcher some stale bread. '' 
If application of further strategies is indeed optional, such 
sentences should sometimee be misinterpreted. 
There is in fact some intuitive support for suck a position 
I 'think most of u% have experienced 
situations where a Slip of 
the Yongue like put the freezer in the turkev passed unnoticed 
atfirst. This ie moat naturally explained by assuming that the 
syntax never got a chance to apply to the sentence, due to the 
fact that Pattern Matching resulted in an acceptable reading. 
Let me now discuss these Linking Strategies in more detail. 
The first type was Linking by Variable Type, which is stated 
in (24). 
(24) Linking by Variable Tne: 
(1) Link the head-variable of each SR-prefix to the 
INB-argument of the a~oronriate te provided 
there is only one eugn appropriate MVB-argument, 
and is not already linked to some other variahl A 
in the SR-matrix; 
(2) if Y is an Bvent-variable .I e then add - e=X, where X 
is MVB plus its modifiers and arguments; 
(3) if n is a predicate variable then substitute for 
F MVB and its arguments. 
L 
This rule says to link the "head-variable" (Lea, the 
left-moet variable) of each prefix to the PNB argument which 
matches its type. Option (2) specifies that event-variables 
are linked to the event described by the main predicate1 and 
option (3) deals with certain adverbs like slbwlw or softlv. 
The rule will be explained usine the example (25). 
(25) (a) #Yesterday the fath5r of the boy sang + #horribly 
# in the bath 4. 
(b) PSR: (THE~~:YSSTE&~~)[THEXI(TH~~:BOY~)(FATHERX~)J 
(E t2aPAST tZ) (E x*)(E P)(E e) (TIIEZ: BATH%) 
E~INc~~x~~~ & HORRIBLY(?) & INez 1 
Looking at (25b), the PSR for sentence (25a), it can be 
seen that the variables in the prefixes are of several 
different ty~es; there are the time-variables 4 and t2, 
individual-variables IC, y, 2, etc . , a predicate-variable ', 
and an event-variable The main verb (M3) is SING, and its 
arguments are a and Kow, time-variables can only be linked 
to other tkxhe-variables, and these links have the specific 
form ~EYB 
("time-of-tne-YXB is included in tirne- 
of -the-prefix" ) , in our cabe t~ c Cl. Next there ie an 
individual-variable JC, which matches only the ma-argument 
hence the link is x=x,m Then there is the predicate-variable 
F which, according to option (3) of the strategy, is replaced 
- 
by iWB plus its arguments, yielding the combined reading - 
H-(SIN G%nally, the event-variable p $s linked 
to the entire event described by the main predicate, namely 
that x2 sari(: horribly at time t21 p=tlO8RIBLY (SINGX~~) This 
exgands the matrix of (25b) into (26) , giving us, after 
simplification, the final 3R (27) for (25a). 
The entire procedure is eyntax-free, with one qualificationt 
the constituent correepond1ng to the pain verb is marked in 
the semantic repreeentation by a oonresponding feature 
That is. SING in (25b) is marked as WVlulVB This syntactic 
feature is immediately copied from the syntactic structure 
into the semantic one$ where it is preserved until the 
entire sentence has been processed. 
For many constituent types, Linking by Variable Type is 
the only Linking Strategy that is needed. This is true 
particl~larly for adverbs, temporal phrases like last Octbber , 
PrepP like into the garden, and auxiliary verbs like will. - 
In addition, if a, clause has only one nuclear noun phrase, 
as 2s the case in sentence (25a), then the entire interpretation 
fs normally taken care of by this strategy# 
The next strategy was Pattern M*tching, which is stated 
in (28). It says that if certain clause constituents match 
a pattern, then they are linked as in the pattern. 
(28) Pattern Matching: 
(1: Let (TIIPV~IA~V~)~..(TY~$~A~V~) Bvi...ve m 1 be a 
pattern called up by a PSR whose MVB is B~~~(v~~ 'V n )t 
(2 ) then for each S2-pref ix head-variable ui matching 
exactly cne desc~#iption A ov! in the pattern, add 
JJ 
a link Ui=Vj (whsre Vj occupies the same argument-place - 
in B~~ as does vf J in the predicate B of the pattern) 1 
for ui to match a description Ajvj, there must be 
an expression Ajui in the ISR. 
For instance, pqlr earlier example (20) r 
(20) This bread was sold to John by the Italian baker. 
contains the constituents BRW, JOrIN, BAKER, and SELL8 They 
are linked in accordance with the pattern (21), which stated 
fhat bakers typically sell baked goods to people. 3ere BR~Av 
in (20) matches (by implicatfon) BAKE~GOOD in (21), BAKER in 
(20) matches BAKZR Ln (21), JWN matches HUMAN, Bnd SW 
matches SELL. iiecall that the Intersection Strategy in (19) 
debmines just which patterns are called up far a giben 
sentence. Th~re will usually only be a limited number of 
activated patterns, 
Next consider the Contradiction ~limination Strategy. 
An abbreviated verEIion of it is aiven in (29) Roughly, the 
rule %ests which roles each NP could play without leading 
to a contradicption, Then it checks whether this would lead 
to a unique combined role assignment for all noun phrases, 
and if so, it accepts this combination as the interpretation. 
(29) Contratiiction Elimination Strateevr 
(1) Let vlaSav be prefix head-variablea of a single 
rn 
type T, and ul...u WB-arguments of the same type TI 
n 
(23 add each combinatorily possible non-contradictory 
link Vi- 
to a list: if there is a uniaue combinead 
3 
m-tuple of lank8 vl=ud & . & v,=uk (where no ui or 
vi OCCUF~ more than once), then add this m-tuple of 
ZLnks to the SR. 
An example is shown in (30)' 
(30) Tne cheese had been seen by the mouse. 
PSRa (THE~~CHEES~X)(THX~~MOUSE~)(E~~,Z~)~SE~~~Z~Z~ 2 
Non-contradictory linkst x=zp. y=zp y=ze (zl must be 
ANIMATE, x is INANIMATE!) 
Comb.ined 2-tuple of links; x=z2 & 
FBR : ~THEX : CHtEESEbc ) (THEY ~BIOUSE~) C SEEyx 2 
Notice that even though mouse, taken by itself, corrld 
be either logfcal subject or object of see, the combination 
mouse/cheese has a unique semantic relation to see, because 
cheese can of course only be logical object of see. 
The ~an~nicai Order Stratem is shown in (31). iioughly, 
the strategy attempts to equate the surface order of the 
major clause oonatituents with 'their "deep order I I 
(31) Canonical Order Stratenar 
(1) Let vl, .a,v rn (m>l) be prefix headvariables of 
a single type T (where Vi precedes vj; for i< j(m) , 
and let ul,e.~,u, be type T argument places of MVB 
(where ui precedes u for i<j~n) 8 
9 
(2) then each Vi not yet linked to any matru-constituent 
is linked to the MVB-argument ui, unless MVB has 
the feature PASV in the syntactic structure. 
Far example, let (33) be the preliminary SR of (32): 
(32) John gave the cat some milk. 
Here the first three prefixes are not yet linked to any 
matrix-constituent, and their head-yariablee &, y, p, are of 
the same type. Hence they are linked as in (34), yielding, 
after simplification, the find SR (35). 
Finally, 3 describes the svntax-sensitive rule. It is 
seneuive to the fdature 3?Aw of the verb and links the 
constituents marked S7Mr OBJl, 0BJ2, and AGenT to the main 
verb, insofar as this has not been done by earlier strategies. 
(36) Wtu-sensit-ive NNB-rRulet 
(1) 2he first MVB-argument is linked to the prefix head- 
~&t?ia5le pointlng to: 
{a) the SW-phrase if MVB is not marked PASV; 
(b) the AGT-phrase if there is such a phrase: 
(2) link the - prefix head-variable pointing to 8 
(a) an ORTI-phrase to the second MVB-argument1 
(b) an OBJ2-phrase to the third WB-argument; 
(3) if MVB is PASV Unk the prefix head-variable pointing 
to the SBJ-phrase to the remaining free (but not the 
f irst-) MVB-argument. 
An example is given in (37)d~ere as elswhere in this paper, 
"#" atande for initial and "$" for final clause boundary.) 
(37) # The churchSBJ had been givenPASv the moneyow2 $ 
L#Pos~c~] by the bakerACT $a 
PSRI first clalaset (T~B~:CHURCH~)(THE~~~ONEYX)(E~~X~Z~) 
tliIV~z~z~2~ 7 
Links; x pdints to OBJ2. hence x=zg; 
y pomt.8 to S0J of a PASV verb, hence 
y=z2 by option (3) of rule (36)l 
FS3 f6r first clause8 (THEY~C~~URCX~) (THEX:MONEYX) (~2~) 
CGIVE~~Y~ 1 
PSR for by-clauser (THEZIBAKERZ) 
Linkr z points to AGT-phrase, hence z=zl by option (lb)i 
VXII a DIFFEi3ENT XOD3S OF PROCESSING 
Let me interrupt here to consider the practical problem 
of constructing an automatic sentence recognizer. Some aspects 
of the thapry I have just sketched may not be optimal for a 
~omputational model, even though they seem appropriate for 
a psychological model. For instance, to a pefson engaged in 
normal conversation accuracy of understanding is not very 
crucial most of the time" Often, the goal may only be to 
'-get the essentials," and if some mistake &s made, it is 
simply corrected later on. 
This "normal mode of processingw is what the psychological 
model sets out to describe. Now, in the case of an artificial 
intelli~rtnce system, one would probably demand higher accuracy, 
so as to minimize the need for correctionsm This is, in some 
ways* similar to the situation where you put subjects in a 
psychulinguistic experiment. They usually abandon the "normal 
mode of processingw very soon and instead employ the strategies 
that guarantee best performance for the specific experimental 
task they are faced with . To give an example, consider the 
common type of experiment where a subject has to verify 
sentences like those in 138). 
(38a) 5 precedes 13. 
(38b) 5 is preceded by 13. 
(384 13 is preceded by 5. 
In order to interpret such sentences correctly, his 
knowledge that 5 in fact precedes 13 is of no help whatsoever 
to the listener, because true sentences occuf together with 
false ones in this game. Therefore, he will soon drop all 
semantic short cuts --which he normally eaploys-- and 
interpret the sentences purely on the basis of their syntax. 
In a way, therefore, such experiments do not really tell us 
anything about normal speech processing. 
Still, such conditions o'f heightened aecuracy may be just 
the onee we want to apply to the artificial intelligence 
system. Let us therefore consider how such a non-normal mode 
of processing could be sinulated in our theory. 
Now, looking back at the Linking Strategies in (22) above 
(~lnking by Variable Type, Pattern latching, Contradictinn- 
Xlimination, Canonical Order, ana Syntax-sensitive MVB-~ule), 
there is indeed ar. obvious way of simulating the "high accuracy 
procedure," namely by dropping type 2, 3, and 4 strategies. 
'Phese coflstitute the "shore cuts" which work 90% of the time, 
but sometimes lead to misinterpretations. Notice that the 
result is still not a syntax-governed model, because most 
of the linking would still be handled by the syntax-free 
ne'chod of Linking4by Variable Type, and syntactic btruc tare 
~ould still be erased in clause-intervals, As a matter of 
fact, in this last respect I think it is possible to go even 
further than I have done here, and erase syntactic structure 
after each major clause constituent (i.e., after each NP, 
adverb, or main verb), retaining only its functional feature, 
which is then simply integrated into the semantic representatbn. 
I have already done this here for the constituent ~ain~er~, 
so in a way this woula only be a logical extension of my 
proposal Looking at the syntax-sensitive MVB-Rule (36) , 
it is evident that it rerers exactly to those functional 
features, namely SBJ, OBJ1, OBJ2, MVB, and AGT. 
It would seem, then, that even for purposes of artific'ial 
in%lligence, it may be preferable to operate with a parallel 
processing model, thereby minimizing the size of the syntactic 
gltructl~re and the amount of syntactic operations. 
Apart from this, I would like to argue that the essential 
ingredients. of Pattern Matching and Contradiction-Elimination 
are still required for any adequaze tJleo~. Ws-bri~g 
back to the sixth type of Semantic Linking Strategies, 
namely Alternative Linking Strategies. 
IXsWALTERNATNE LINKING STRATEGIES" 
Alternative Linking Strategies apply if the normal 
strategies (types 1-51 have failed to produce a semantically 
acceptable reading. These strategies rely heavily on the 
Semantic Knowledge Rules. 
The most important and most general (and the only one to 
be discussed in the present paper) is the Obvious Connection 
Strategy in (39). It says, roughly, that if a variable g 
cannot be linked to the RnB then if the encyclopedia contains 
a rule conne~ting q to same other entity 11' then try and link 
this new entity yo to the MVB. 
(393 Obvious Connection Stratear 
If the head-variable u cd a prefix (QTF~UIAU) cannot be 
lhked to its appropriate IWB-argument V, and both u and 
v point to a rule in the encyclopedia such that a 
connection (~~~~u't~u')[~uu'~between u and u' is 
derivable, then t 
(a) qhange (QTF~UIP-~) to <QTF~U'Z(QTF~U~AU)(BU~ & CUU')), 
(b) link ut to v. 
For Instance, consider again sentence (9b)l 
(9b) They published Wodehouae immediatela he came over. 
s indicated in (40), the object of PUBLISH must be some 
ritten work, and Wodehoyse of course does not qualify as 
uch. Therefore, the l5nR between x  odeho house) and the 
ogical object ye of PUBL~~H is rejected. Notice that this 
s just the kind of "semantic anomaly test" which was central 
o the Contradiction-Elimifiation Strategy Its intuitive 
asis is obviousr the listener normally assumes that the 
peaker is trying to make sknse, and therefore, he will 
eject all non-sensical interpretations. 
40) They published Wodehouse. 
P3Rr (THEplz:THEYz) (THE~IWODMOUSEX) (~y* ) t PUBLISHZ~'~ 
ISR t WRITEiix & HUMANX & ANXMATEX I WRITTNWRKy' & 
-ANIMATEy' .rr 
Encyclopedic Pattern; (TYPX :WRITEXx) (TYP~;WRITINWRKJI) [(REATEX$] 
Rule (39) changes (THEX~WODMOUSEX) in the PSR to: 
{BY; (THEX r WODMOUSEx) (WRITTNWRK~ & CREATEX~)) 
Link added by Rule (391 I y'y', yielding the FSR: 
(THE~~Z~THEYZ)<E~I (THEX IIWODEHOUSEX) (WRITTNWRK~&CREATE~~)~ 
~PUBLISHZ~ ] 
Now, staying with our sentenae (40)  hey published  h ode house), 
if you know that Wodehouse was a writer you also know that he 
created written works3 and ths new entity introduced by this 
encyclopedic rule, namely "written works created by Wodeh~use,~ 
is the one which is interpreted as .logical object of the 
predicate PUBLISH. This then is the kind ofWsernantic detourw 
descrttJeb by tRe ObvTous ConnectLon strategy, and this strategy 
is applicable to the other sentences in (9) too 
10 
Now, I think the sentences in (9) exemplify something 
that happens all the time in speech: namely omission of 
the obvious. I also think it would be extremely inconvenient 
if we had to ask people to use only their 
best Sunday Grammar 
when conversing with an English-speaking robot. Me might as 
well ask them not to use pronouns, or to speak at a constant 
pitch of 4.00 Hertz. To be competitive, the robot should 
undecstand Monday Gramar as well, and that means, sentences 
like those in (9) 
Xr CLAUSE-TO-CLAUSE LINKIKG 
For the remainder, 3: would like to discuss clause-to-clause 
1Mcing. The dieti~lguishing feature of my proposal, as will 
10 
See ~chank(1972) for a different approach to the problem of 
recovering implicit information. He outlines, for instance, 
a method by which "I like books" would be expanded into the 
conceptual equivalent of "I like to read books." 
SchankDs theory is baqed on wconceptual casesw and lexical 
decomposition rather than meaning postulates. For instance, 
he claims that "John would be pleased by Il:ary8s going" is 
a proper part of the meanhg of John wants Marv to come home, 
while in my terms "John would be pleased by !daryw s going" 
is merely an allowable inference which may or may not be 
drawn. Failure to Eraw an allowable inference explains the 
possibility of holding contradictory beliefs . For instance, 
somebody might judge John's uncle left to be true, while at 
the same time believing that the brother of John's mother or 
father left was false, because he failed to apply the 
mean%% rule relating uncle and brother of mother or father. 
It is difficult to see how a theory basedon lexical 
decomposition would explain such f&c ts Certainly , it would 
be unreasonable to claim that a person has not understood 
the sentence John's uncleJaft unless he also is aware of 
the synonymy relation between this sentence and the brother 
of John's mother or father Left 
be recalled, is that syntactic structure is erased clause-by- 
clause. The problem, therefore, is to show that sentences 
containing more than one clause can in fact be interpreted 
correctly without referring to the full syntactic structure 
of any (completed) earlier clause. 
Consider first wPost-clauses" e, clause-final 
adverbs, temporal nouns, and preposition-phrases Tliey are 
already covered by the strategy of Linkillg by Variable Type 
discussed earlier. For instance, consider again sentence (25a). 
(yesterday the father of tne boy sang horribly in the bath.) 
It contains the two P08-b-clauses horribly and in the bath; 
and it was shown earlier how they are linked to the main 
clalase by Variable Type The only smtac tic information 
required for this operation was the marking of the MVB in 
the semantic representation* 
(41) indicates roughly how relative clauses are integrated. 
(~ctuall~, there are some complications here, but they are 
irrelevant to the present discussion; the main point here 
is that again no reference need be made to the syntactic 
strracture of the first clause. For details, see ~eimold (f orth- 
comina where coordinate clauqes, comp&rison-clauses, and 
various sl~bjectleos complement-clauses are treated as well.) 
(41) # The girl (if who was tired 3) giggled q. 
(TMEX:GIRLX)[--~ and (~y) [T1i(E~yl link, y-xj 
integrated structure a (T~IEX I GIRLX & TIi-iEDx) C -- 7 
Next consider circumstantial clairses like because the 
kan~aroo ium~ed in (43). The strategy for these claus~s, 
stat-ed in (42), does again not refer to the syntactic 
structure of the first clause* 
(42) Circumstantial - Clause Ruler 
Substitute the matria of the main clause for the free 
s-argument of the JUNCtor of the circumstantial clause, 
and join the prefix of the main clause before the prefix 
of the circumstantial clat~se. 
(43) # The boy wap nappy $(# lbecause JUNCJ the kangaroo jumped$) 
(THEIS : BOYX) [HAPPYX 1 and (THX~ 1 KANGA~) CBEGAUSL ( Q ,JUMPY) 3 
"xJ ?' 
The same holds true for complement-clauses like that the 
cake was ~oisoned in (45). The corresponding strategy is 
given in (44). 
(44) Com~lement Clause Ru1e.r 
Substitute the SR of the complement-clause for the 
free s-argument of the MVB of the main clause. 
(45) # J ohn btlieved $ ( LYcompll that the 2ake was pcdecncd 4) 
(THEX~JOHNX)LBELIEV~~X an41 (THE~ICAKE~)[~~EDNED~~ d 
FSRr (TH-I JOHNX )CBKLIFNE$ x, (THE~~CAKE~) (POISONED~) 3 7 
In sum, the principle of clause-by-clause erasure of 
syntactic structure seems ihdeed compatible with the 
requirements of clause-to-clause lirling. The exception is 
the feature MVB, and it was suggested earlier that this 
syntactic feature is integrated into the semantic representation 
Notice also that certain aspects of syntactic structure are 
reboverable from our semantic representations. For instance, 
the order of the prefixes in the SR reflects the surface 
order of the NP'a of a clause. Hence, if certain constructions 
require access to such syntactic aspects, this i8 still not 
incompatible with erasure of syntactic structure in clause 
intervals.  or instance, coordinate clauses and certain 
subjectless complement clauses do often require identification 
of the surface subject. For detailed discussion, see 
Reimold (forthcoming) . ) 
X-I r THE "TEMPORAL SEQUENCE STEIATEGY " 
The last strategy discussed here concerns the tense of 
consecutive clauses. When we interpret sentences S1, 52 in 
a text or sentence conjunct, where S3, and S2 have the same 
tense, we often assign a relative chrondlogy to the events 
described by these sentences. I will refer to the principle 
assigning such a chronology as the Temporal Sequence Strategy. 
For example, consider (46) (Agatha Christie, Thev came to 
Baghdad, pmlOO)t 
(46) Then his head jerkedtl back a little and he lay 
still. 
t2 
Assigning  PAST(^^) to the first clause and  PAST(^^) to the 
conjoined clause in 
(46) does not account for actual 
comprehension: the listener knows that t2 is later than tl, 
even though there is no overt sequewe marker (e sg*, before, 
after, then) 
As a first approximation, the strategy might be stated as 
f 011 ows t 
(47 ) Tem~oral Sequence stratem (preliminary) I 
Given two main or conjoined clauses C1, C2 such that 
C1 precedes C2 and the time tl of C1 has the same 
"tense predicatew (e.g., PAST, FUTURE) as the time t2 
of C2, then assume that DIRFRBC tp= 
(47) will require several modifications . First, the rule 
holds for certain nan-tensed clauses as well, as illustrated 
by (48) (~hesr came- to Baghdad 
p 044) 8 
(48) Then he was out, across the Khan, back into the Suq. . 
The full interpretation of (48) must specify that the time of 
out precedes that of acrcws, which in turn precedes that of 
- 
bacq. Since across and back had no overt tense predicate in 
7 
i48), the strategy must somehow be Liberalized to include 
SUC~ C~S~S. 
In this connection, consider also (49)(~hey came to Baghdad. 
pa711 1 
(49) Never. I thought, would the plane land. It went round 
and round*:! and roilndt,. 
t 1 
If we assumed that tpt2=t3 in (49), then the two 
conjoined phrases pnd round and round* should be redundant 
in the same sense in which Fido ia a dog and is a doe: and 
is a dog is. However, (49) can quite naturally be interpreted 
as "the plane went-round and then round and then roundN 
be.. as 45. -DIRPREC t2 & _t3 DIRPREC t3). 
Next, the Temporal Sequence Strategy seems blocked if 
some general semantic principles \in ~articular, Pattern 
Matching) suggest a chronology caR#licting with that imposed 
by the Temporal Sequence Strategy. For instance, consider 
a dialogue like (50): 
(50a) What's the matter with John? 
(50b\ Oh, he broketl his arm. He felltZ off his bike. 
Here the second sentence in (sob) is interpreted as 
preceding the first sentence in (50b) temporally, counter to 
what the Temporal Sequence Strategy would prediict. The reason 
is obviousr there is a perceived causal connection between 
the sentences, such that the second sentence describes the 
cause of the first. Since a cause must precede its effect, 
t2 must precede tl in (50b). 
To formalize this, we can make use of Pattern ittcbing. 
For instance, the encyclopedia would contain a pattern like 
(51), and there would furthermore be a meaning rule like (52). 
(52) (AU.,tl,t2) LCAUSE<A(. dl. a) ,B(. at2. 43 IMPL tl CIRPREC t2 3 
We need only make sure that the pattern (51) is activated 
by the two sentences he broke his arm and he fell off his bike 
in (~oB), which can be- done by calling up all patterns which 
are in the intersection of the main verbs of the two sentences. 
pa or instance, break and fell both point to (51).) In effect, 
we have to add the following principle to the Temporal Sequence 
Strategyr 
(53) Causal Connection Constraint: 
If C1, C2 are not conjoined by pd, call up all 
Encyclopedic Rules in the intersection of the MVBVs 
of C1 and C2. If 4 DIRPREC tl is heuristically derivable 
from a pattern, then add this link tu C2 and do not 
apply the Temporal Sequence Strategy. 
11 
Another restriction seems to be that the Temporal Sequence 
Strategy is inapplicable to proaessive tense. For instance, 
while tl must precede t, in (54), they seem to be roughly 
- 
simi~ltaneous in (59, even though liatina a ci~mette and 
leaving a room normally count as "instantaneous" events (see 
11 
The condition against conjunction by and is necessary since 
pnd can never mean "and before that..," For instance, 
a sentence like pe broke his arm and fell off his bike. 
cannot be interpreted as "he broke his arm because he 
fell off his bike. tB 
the "Short Events Principle" discussed below) 
(54) She lit a cigarette and left the room. 
tl t2 
(55) She was lightingt 
a cigarette and leavingt2 the raom. 
1 
progressive tense turns events into non-instantaneous 
events. 
Finally, I come to €he most general and important 
restriction on the Temporal Sequence Strategy, Phis restriction 
is of a subtle semantic nature: it states that the strategy 
is applicable to clauses C1, C2 on1~ if the events 
described by C1 and C2 are unlikely to he simultaneous. 
Consider, fqr instance, the following sentence: 
(56) She just stood there and looked at him. 
It is perfectly possible for someone to stand somewhere and 
at the same time to look at somebody. Hence no temporal 
sequence is imoosed on stood and looked. By contrast, in (57) 
(~heu came to Baghdad, p.76) wen3 and stood must be interpreted 
as seqyenced, since one .cannot at the same time a somewhere 
and stand somewhere elser 
(57) She wenttl out from the bar onto the terrace outside 
and stood% by the railing... 
2 
Note that it will not do to define.the condition of 
"possible simultaneity' directly for verbs. We cannot say, 
e.g., that clauses containing stand and a must be"sequencad." 
For instance, in (58) and (5% stood and went would be 
interpreted as simultaneous, the reason being, of course, 
that different agents are involved, 
which makes simultanei-ty 
c onc e ivabl e . 
(58) Jack stood by the window. Jane went to the door. 
(-59) Jane went to the door. Jack stood by the window. 
It is clear, then, that a detailed semantic analysis is 
need6d to determine "possible simultaneity of two events. 
The principle can be stated as follows: 
(60) possible Simultaneitv Constraint: 
Call u~ allEncyclopedic Rules in the intersection of 
the MVBms of. C1 and C2. Unless tl t2 is derivable 
(stricrtly or heuristically) the Temporal Sequence 
Strategy is inapplicable 
There are some cases, however, where two events are 
nocmally ihterpreted as sequenced, even though they could, 
stric-tly speaking, be sirnultaneous~ For instance, it is 
theoretically possible to light a cigarette while g&tting up, 
and one certainly would not want to add a rule to the 
encyclopedia stating that two such events are unlikely to be' 
sinnaltaneo~s~ Nonetheless, (61) is normally interpreted 
a8 sequences I 
(61) John got up and lit a cigarette. 
Aather than adding some ad hoc principle whlch somehow 
codes the intuition that it is awkward to strike a match and 
hold it to the cigarette while geteing up, we can exnlain 
this as a consequence of a much more general heuristic 
principle. Notide that both clauses of ( 61 ) describe 
instantaneous events. Now, the shorter tdo events, the lower 
the probability that they coincidel For instance, the 
probability that the radio plays at the same time when there 
is a- shot is duch higher than the probability of there being 
a shot at the same time wi~h a hiccough. As another illustration. 
consider (62)(~hev came to Baadad, p. 72): 
(62) With a fatherly smile he withdrewel, Victoria sat down, 
2 
on the Ded and passed, an experimental hand over her hair. 
3 
Although it is Legically conceivable that all three events 
el-e3 happened simultaneous~y, the normaL in~erpretation of 
these sentences takes them to be sequenced, because they are 
short and henca unlikely to coincide. 
This means, then, that we must add the following heuristic 
principlm t 
(631 Sheet Events Princur 
If C1, C2 denote $nst$uvt@neous events e, JNSTANT tl 
and  STA ANT t2 are derivable from M-rules and E-rules) 
assume that tl#t2. 
In summary, the Tem:poporal Sequence-Strategy takes on the 
followinq form: 
(64) Tem~oral Seauenee ~trate~v(revised)t 
(1) Let C1, C2 be two main or conjoined clauses such 
that CL is to the left of C2 and the time tl of the 
MVB of C1 has the same tense predicate as the time 
t2 of the MVB of C2, or t2 or both tl and t2 have 
no tense predicate; and neither tl nor t2 has the 
tense predicate PROGRESSIVE; 
(2) call up all Encyclopedia Rules in the intersection 
of the WB of C1 and the WB of C2; 
(a) if C2 is not introduced by and, and 2 DIRPREG tl 
is heuristically derivable from the active E-rules 
and i-rules, then add 42 DLSPREC ti and break off 
the application of this strategy; 
(b) if ei'ther - tl= or INSTANT t~& it? dderi-e 
(strictly or heuristically) from the active E-rules 
and M-rules, then add the link DIRPRZC t9. 

BIBLIOGRAPHY 
Abrams.K. .& Bever, ~~~~1969. Sntactic structure modifies 
attention during speech perce tion. Quarterly Journal of 
'3s~erirnefital ~ssc holo~~v 21, 2 8 0-290. 
Bever, T.Q.1970. The cognitive basis for linguistic structu~es. 
Cornition and the develonment of laneua~e, ed. by JaR~Hayes. 
New Yorkr Wile'y & Sbns. 
Caplan,D.1972. c'lause boundaries and recognition latencies for 
words in sentences* Percention & Fsvcho~hvsics 12, 73-76. 
Christie, A.I.951. They kame to Baghdad. Yew Yorkr Dell. 
CoXlins,A .Mot & Quillian,M.R.19720 How to make a lanwage user. 
Ornanization of memorrv, ed . by E l Tulving & W . Donaldson 
New York~ Academic Press 
Fodor, J .A . , Bever, T .G , 8~ Garrett, Me 1974. The ~svcholoav 
of lanmage. An introduction to ~sucholfnsruistics and 
eneratiye =ammar. New York: NIcGraw-Hill. 
GarRett , Ma, Bever , T .G ., L Fodor , J .A a1966 The acf;ius.,-use 
of grammar in speech perception. Perce~tion & Psvcho~h~sics 
1, 30-32. 
Katz,J.J., & Fodor, ~~~~19630 The structure of a semantic 
theory. Lm 39.2, 170-2100 
Kuno, S ,1965. The predictive analyzer and a path elimination 
technique. Communications of the ACM 8 e7, 453-462 
Marslen-Wtlson, W ,1973 Linguistic structure and speech 
shadowing at very short latencies. NaSure 244, 522-23. 
Nash-Webber, B.1974. Semantics and speech understanding. 
BBN Report 2896. Cambridge,, Mass.:Bolt, Beranek & Newman. 
Petrick, s .R .1966. A pro am for transformational syntactic 
analysis AFCfiL-66-69K Bedford, Mass c Air Force Cambrf dge 
Research Labs. 
-------- .1973. Transf orrnational analysis Natural languaae 
rocessinq, ede by R. Rustin. New York: Algorithmics Press. 
Qui?lian,M.R.1969. The teachable langudge comprehendera 
A simulation program and theory of language. Communications 
of the ACM 12, 459-476. 
Reirnold, P.forthcoming. An integrated system of perceptual 
strategies: Syntactic and semantic interpretation of English 
sentencess Columbia University doctoral dissertation 
Rieger, D.1973. Conceptual memory. Stanford University doctoral 
dl ssertatton. 
Schank,R.1972. Conceptual dependency: A theory of natural 
language understandine. Co~ltive Psvcholoey 3, 552-631. 
Weinreich ,U ,1966 Explorations in semantic theory. mrera4. 
trends in linm%stfcs 111, ed. by T~A~Seheok. The Hague1 
Mouton. 
Winograd,T01971. Procedures as a representation for data in 
a computer program for understanding natural language. 
M .I .T . Artificial Intelligence Laboratory Report A1 TR-17. 
Cambridge, Massla Mo1.T. 
Woods, W .A .I973 . An experimental parsing system for transition 
netyork grammars Natural lanauaae ~rocessing, e,d by 
R. Rustin . New Yorka Algorithmics ~ress'. 
