1969 International Conference on Computational Linguistics 
5emantica and the Syntactic Classification of Words 
Ernst von Glasersfeld 
(Georgia Institute for Research, 
and University of Georgia) 
(*) 
In presenting some aspects of our somewhat unusual 
grammar to a gathering of qomputational linguists, I feel 
justified in taking two things for granted: first, that 
most people who have taken an active interest both in com- 
puters and in natural language have come to realise that 
computers, although impressively fast and reliable in many 
tasks, are not very brilliant when it comas to making in- 
ductive decisions on the basis of insufficient or not tho- 
roughly defineddata; and, second, that what we linguists 
know about the workings of natural ~anguage is by no means 
enough to supply computers, as they are, with a solid bas- 
is for the fully automatic handling of natural-language. 
data. 
This is not meant to be a disparaging comment on pre- 
vious efforts in linguistics, for, after all, the contin- 
gency that has opened our eyes to these shortcomings did 
not exist until a few years ago. Traditional linguists,. 
and especially grammarians, could carry on their business 
quite happily on the more or less explicit assumption that 
language - like so many things in the still wide-spread 
Platonic view of the world - could be separated into two 
levels: the ideal, uncontaminated one of pure structure 
* The research reported in this paper was sponsored by the 
U.5.Air Force Office of Scientific Research (OAR), Inform- 
ation 5ciences Uirectorate (Srant AFOSR 1319-67) and act- 
ively supported by the University of Georgia, Athens,U5A. 
- 2 - 
or form that was inherently generalisable, and the slight- 
ly messy, unsystemic, and therefore far less interesting 
one of individual content. This dichotomy was possible and 
workable as long as the use of language was restricted to 
organisms who, by the time they embarked on linguistic act- 
ivities, had necessarily absorbed a vast body of experien- 
tial and conceptuaI knowledge, on which, more or less con- 
sciously, they could draw whenevez the formulation or com- 
prehension of a linguistic expression required something 
beyond the rules of the ideal grammar. 
With the advent of computers the situation was ~edical- 
ly changed. Suddenly linguists could and would find them- 
selves committed, for one reason or another, to transmit 
their know-how to a potential language user who did not 
possess any a priori experiential or conceptual knowledge 
whatsoever. Thus there arose innumerable questions which, 
hitherto, no one had ever been compelled to answer, and it 
became painfully obvious that the application of much of 
the linguists' cherished theoretical knowledge to actual 
language material presupposed a considerable amount of as 
yet unexplored preprocessing of that very material. 
I shall not try to catalogue the types of question and 
the kinds of problem which have been thrust upon the lin- 
guist by the appearance of computers - every one of you 
is familiar with some if not all of them; instead I should 
like to present a few examples and show the direction in 
which, we believe, some solutions can be found. 
By 'linguistic activities' or 'handling language data' 
we mean, for instance, formulating a given message in e 
specific natural language; or recognising, in a given 
piece of language, the message intentionally formulated 
- 3 - 
in that way by an author; or translating a formulation giv- 
en in one natural language into a 'corresponding '(*) form- 
ulation in another natural language, etc. Since the 'pre- 
processing' in all these cases involves what, summarily, 
is called semantics, it is necessery to stress that in our 
view of language there is no unbridgeable abyss or opposi- 
tion between semantics and syntax. In fact, we speak of 
'relational' semantics (a term already used by Ullmann (I)) 
when we try to define that part of an expraseion's meaning 
that is determined by specific syntactic functions (or Co.~r- 
relators); and we call 'lexical' semantics the attempt - 
for instance the lexicographer's - to define the meaning 
of words as separate individual items. 
In traditional grammars the lexical items of a natural 
language are classified as 'parts of speech' according to 
their generic syntactic functions and/or their morphologic- 
al characteristics; in correlational grammar (**) they are 
classified exclusively according to the actual roles they 
can play in correlational structures; moreover, while tra- 
ditional grammars operate with about a dozen different ge- 
* Note that since we are interested, not in developing a ri- 
gid logically formalised theory of grammar, but in develop- 
ing a flexible operational system fwr the automatic inter- 
pretation and handling of natural-language sentences and 
text, we do not require formal instruments for the determ- 
ination of 'meaning', 'interlinguistic correspondence', 
'synonymy', etc.; for our empirical purposes the consensus 
of proficient language speakers is the relevant and suf- 
ficient criterion. 
** Although both theory and practice of Correlational Gram- 
mar have been drastically modified in our empirical appli- 
cations, much of the terminology and the basic concep~ 
derive from the pioneering work of Silvio Ceccato , whom I 
was fortunate enough to have as friend and teache~. 
- 4 - 
neric syntactic functions, our correlational English gram- 
mar distinguishes several hundred correlators (in the pre- 
sent version of our automatic parser (3) there are 350). 
The list of these correlators is the result of intuit- 
ive analysis of English tex~s, continual refinement by 
means of insights gained in the translation of English sen- 
tence structures into other languages (*), and by experi- 
ments with an automatic parsing procedure whose output, or 
lack of output, inevitably demonstrates the shortcomings 
or, as we p~efer to put it, the degree o~ completeness 
reached by the system's grammar. In our terminology a 'cot- 
relator' is a connective function which links two pieces - 
either words or word combinations - and thus forms a unit 
we call 'correlation'. Correlators are divided into sever- 
al types, the main distinction being made between implicit 
correlators, which are indicated in the sentence merely by 
the juxtaposition of the items they link, and explicit cot- 
relators, which are indicated by specific words (mostly 
prepositions and conjunctions). 
To serve as a valid tool in the parsing or interpreta- 
tion of sentences, a correlator must not only have an in- 
dividual code number ('Ic' or correlation index) but the 
particular relation it establishes, between the two items 
it correlates, must be characterised or explicated. This 
explication may be an ad hoc description, an illustrative 
paraphrase, a suitable transform ~ l a Chomsky (4) or Fill- 
more (5), or a symbolic expression devised along the lines 
(6) of function symbols in logical calculus . 
* Translation frequently helps to pinpoint relational ambi- 
guities which, although relevant in the disambiguation of 
other structures, are not immediately perceived by the 
monolingual speaker or reader. 
- 5 - 
As ~mples of correlator explication, here are two (one 
concerning an implicit correlator, the other an explicit 
one) taken from the correlator list which has been imple- 
mented in our operational system: 
Corr. No. Description 
3670 N generic type: verb phrase complemented by in- 
finitive; 
Explication: the gram. subject is the actor of the 
infinitive activity; the infinitive specifies 
the purpose of the subject's primary activity. 
e.g. "He works to live", 
"I braked to avoid the child"; 
but not: "He began to live" (3430 N), 
"I had to avoid the child" (3410 N); 
no__~r: "He was eager to live" (3350 N), 
"I was clever to avoid the child" (7012 N). 
0243 E generic type: ~Y, temporal limitation; 
Explication: the item on the right of "by" speci- 
fies the point in time after which what is as- 
serted by the item on the left of "by" is to be 
considered an accomplished fact. 
e.g. "You will be paid by Christmas", 
"He had left by 1865"; 
but no__~t: "You will be paid by cheque" (~257 E), 
"He had left by the back door" (0251 £); 
no._~r: "You will be paid by the treasurer" (0247 E), 
"He had left by sheer determination" (0255 E). 
It should be clear that the 'Explication' is intended 
to describe as univocally as possible the relation Qua re- 
(*) lation witDout regard for the particular items that 
are 
" In practice it is, of course, not alweye easy to formu- 
late the explication without reference to particular items; 
but as long as the analyst remains scrupulously aware of 
the intention, this is not a serious impediment and, in 
our experience, a sufficiently general formulation can al- 
ways be found sooner or later. 
- 6 - 
eligible to be linked by that relation. Correlator analysis 
and the explications this analytical investigations aims at 
are consequently what we may call 'the semantics of rela- 
tions'. 
The term 'correlation' refers to the word-combination a 
correlator produces, i.e. the ternary unit consisting of 
one specific correlator and the two items (words or them- 
selves word-combinations) that are linked by it; and we re- 
present correlations - for instance the ones operative in 
the first phrase sample) in this manner: 
works to live 
2230 N -j 1--2310 N ~ 
I 367L N --J 
which corresponds to a conventional tree-structure with 
labelled nodes: 
~I..(3670N)~ 
(2230N) (2310N) / \ 
He works to live 
In order to "recognise this structure in the linear in- 
put string of the sentence, first each of the four word 
items must be characterised as a possible candidate for 
linkage by correlators 2230 N and 2310 N respectively (*)-, 
and both these corzelations have then to be characterised 
as possible candidates for linkage by correlator 3670 N. 
This characterisation is achieved by correlation indices 
(or Ic's) assigned to the individual items and specifying 
(a) the code numbers of the correlaturs by means of which 
• The correlator numbers we use were o~iginai\]y significant 
as to the kinds of relation; continual correctiLns and ad- 
ditions, however, have thoroughly obliterated thJ~ signi- 
ficance and the numbers are now neither consecutive nor 
characteristic. 
- 7- 
the item can be linked to other items; and (b) whether the 
item can occupy the right-hand (RH) or the left-hand (LH) 
place in the specified correlation. 
(The process of assigning Ic's, obviously, is of one 
type when the item to which the Ic's are to be assigned is 
a pre-established vocabulary word, and of quite another 
when the item is a correlation formed in the course of the 
analysis procedure; in the first case the word's string of 
Ic's is the result of ~ priori assignation; in the second 
case it is the result of a dynamic process called 'reclas- 
sification', implemented by means of an intricate system of 
rules which take into account the correlator responsible 
for the made correlation as well as the individual charac- 
ter of the pieces correlated in the particular instance; 
for a full discussion of the operational reclassification 
procedure see ref. No.7). 
From the sample phrases given under correlator 3670 N 
it is evident that the two partial phrases constituting the 
right-hand piece of the listed correlations, i.e. "to live" 
and "to avoid the child", must bear RH-Icts not only of 
correlator 3670 N out also of currelators 3430 N, 3410 N, 
3350 N, and 7012 N, while the partial phrases constituting 
the left-hand piece of the two correlations 3670 N will 
bear only the LH-Ic of correlator 3670 N and not those re- 
ferring to the correlators operative in the remaining four 
samples. (In the sample phrases given under correlator 
0243 E, the situation is symmetrically inverted, i.e. the 
two partial phrases on the left will bear only the LH-Ic of 
correlator 0243 E, while the partial phrases on the right 
must bear RH-Ic's of ~ the other four listed correlators 
as well.) 
This assignation of Ic's automatically precludes the 
- 8 - 
formation of unacceptable phrase interpretations, such as: 
He began to live 
L2230 N ~ L2310 N ---J 
t *3670 N -j 
or I was clever to avoid the child 
i L42~0 N~ L2310 N ~ LSOiO N~ L 
2oso N-J L--4olo N~ 
I -367o N I 
which would be roughly equivalent to "de began in order 
to live" or, respectively, "I was clever in order to avoid 
the child" (Note that if we change the form of the LH-phrase 
in the second example and say "I was beinq clever to avoid 
the child", the 3670 N interpretation becomes acceptable 
while the 7012 N interpretation - equivalent to "to avoid 
the child was clever of me" - is no longer possible!). 
Traditional grammars tend to consider such interpreta- 
tional distinctions (if, indeed, they make them) as 'se- 
mantic'; for correlational grammar they are clearly rela- 
tional, beca~:se they are handled excl~sively by the mechan- 
ism of Ic-~,~tching - and thi~; mechanism is the operational- 
ly implemented syntax of the system. And there is another 
criterion as well. In dll the ~bove cases, the assignation 
of Ic's to vocabulary w~rds or correlational products emerg- 
in 9 in the course of the parsing procedure can be determin- 
ed by an examination of the indiviOual item with regard to 
the explication of the correlator whose Ic is being con- 
sidered (i.e. without considerin 9 comple~nentary items); 
this examination, in principle, is similar to the examina- 
tion - in traditional grammar - of a morphologically de- 
ficient or indefinite word with regard to the possibility 
of its being for instance a vet0 or a noun; or, to put it 
in another way, this examination is essentially different 
- 9 - 
from the one required to decide whether two given items are 
compatible as RH and LH pieces in one particular syntactic 
structure(8! 
The question of predicability - especially if elaborated 
in the way suggested in an essay by fred- 5ommers 19;-" " is a 
case in point. Whether phrases such as: 
"blue grass" or "cerise ideas", 
"the indigestion of angels" or "the wings of the morn- 
ing", 
are acceptable, acceptable only metaphorically, or not ac- 
ceptable at all, is apparently not a relational question; 
the relation - a kind of 'appurtenance' - does not seem to 
change, nor could we say that any of the four phrases is 
unacceptable because one of its items is such that it can 
never be related in that way; since we have no objection 
when the same relation is asserted in "cerise paper" o c 
"clear ideas", we can only concludethat there are certain 
items to which certain properties or thingscannot be said 
to appertain and that, therefore, it is a question of le__~x- 
ical semantics. 
If the semantic analyses initiated by Ceccato in the (-) 
195O's were pusneo further, it seems likely that a re- 
lation such as 'appurtenance' could be demonstrated to in- 
corporate (i.e. t~J have confused) a number of specifiable 
subrelations~ and, once we had isolated these subrelations, 
much if not all of what we now, for the lack of demonstrab- 
le distinctions, have to call 'lexical ambiguity' could 
perhaps be resolved relatiunally in a satisfactory way. 
* called 'operational' semantics, by which was meant the 
analysis uf the mental opezations that lead to the forma- 
tion uf the items (or concepts) designated by words (cf. 
ref. No. IO). 
- i0 - 
Thus it may indeed be so, that our need to resort to a 
static, non-relational semantic classification of items, 
in order to interpret phrases and sentences, is only the 
measure of our ignorance concerning the basic character 
and composition of relations; and that, eventually, it will 
become possible to derive a comprehensive and foolproof 
general semantics from the investigation of relational con- 
ditions as they manifest themselves in our actual use of 
language. 
Let me try to make these conjectures a little less ob- 
scure. Gne area of English grammar that has given consider- 
able trouble to analysts is that of a string of constituents 
which, in traditional terms, would have the specification: 
nominal + to be + adjective + infinitive 
A survey of contemporary text shows that this string, 
with different lexical items and in different contexts, (*) 
gives rise to ten different relational interpretations 
Explicating the relevant relations, we get the following 
listing: 
"John is easy to please" 
Paraphrase: to please John is easy 
ExpIication: the gram. subject is the object of the in- 
finitive activity; the adj. specifies an~spect (ad- 
verbial) of the infinitive activity as enacted by 
the given subject. 
B "John is eager to please" 
Paraphrase: to please is what John wants to do 
I 
* I do not wish to claim that ten is all and th~ no ot.~ezs 
are possible; but these ten ere the interpretations w,: ~ came 
across in one year's conscious scanning of everythi~g con- 
temporary we happened to read, 
- ll - 
Explication: the gram.subject is the actor of the in- 
finitive activity; the adj. specifies the subject's 
disposition towards the infinitive activity, and 
this activity is merely envisaged. 
"John was slow to understand" 
Paraphrase: John was slow abou.___~t understanding 
Explication: the gram. subject is the actor of the in- 
finitive activity; the adj. specifies an aspect of 
the subject's performance. 
"John is likely to leave" 
Paraphrase: that John leaves is likely 
Explication: the gram. subject is the actor of the in- 
finitive activity; the adj. specifies an assess- 
ment of the activity's incidence (i.e. whether or 
not it will take place). 
"John is clever to leave" 
Paraphrase: to leave is clever o~f John 
Explication: the gram.subject is the actor of the in- 
finitive activity; the adj. specifies an assess- 
ment (regarding the subject) based on the subject's 
enacting the given activity. 
"John is young to go to school" 
Paraphrase: John is young fo_~r going to school 
Explication: the gram. subject is the actor of the in- 
finitive activity; the adj. specifies an assess- 
ment of the subject's adequacy (or inadequacy) as 
actor of the given activity. 
"John is heavy to lift" 
Paraphrase: John is heavy with reqard to be~nq lifted 
Explication: the gram.subject is the object of the in- 
finitive activity; the adj. specifies an aspect 
(adjectival) of the subject as object of the given 
activity. (Note: the paraphrase given for type A 
is impossible here; and if an ambiguous adjective 
occurs in G, the construction determines its mean- 
ing; e.g. "mushrooms are good to eat" requires the 
interpretation "good"='pleasing', since it does not 
mean "to eat mushrooms is good for you"p where 
- 12 - 
"good"='beneficial'). 
H "John is sad to go away" 
Paraphrase: to go away causes John to b..._~e sad 
Explication: the gram. subject is the actor of the in- 
finitive activity; the adj. specifiesthe subject's 
state which is a reaction to the given activity. 
I "John was critical to upset the speaker" 
Paraphrase: John was critical in order to upset the 
speaker 
Explication: the gram. subject is the actor of the in- 
finitive activity; the adj. specifies a deliberate 
attitude of the subject's, and the infinitive spe- 
cifies the purpose of the subject's attitude. 
(Note: there often is an irresolvable ambiguity be- 
tween type I and type E; e.g."the dog was clever to 
get the biscuit" may be interpreted as I, 'the dog 
was ~ clever in order to get the biscuit', or 
as E, 'it was clever of the dog to get the biscuit') 
J "It is sad to go away" 
Paraphrase: to go away is sad 
Explication: the nominalised infinitive is the sub- 
ject of the sentence; the "it" functions as sub- 
ject marker; the adj. specifies an evaluation of 
the given activity as event. 
As far'as relational ~naly~is goes, this discrimination 
of types is fair\]y satisfactory (~itnough, t~) be really 
solid, it would require the detailed definition ,~n~ coher- 
ent application of the terms used in the explications , 
many of which, e.g. 'aspect', 'assessment', 'attitude',etc., 
are still rather vague). ~hen we come to a sentence-inter- 
pretive procedure, however, such a listing of relational 
possibilities does not get us anywhere, unless we are able 
to provide each type with some criterion by means of which 
we can recognise it in the input text. 
In an attempt to discover some such criterion, we as- 
- 13 - 
sembled a corpus of about i00 relatively frequent adject- 
(11) ives from a recent compendium of £nglish word frequency 
and examined their individual possibilities to function as 
acceptable constituents in the ten types of construction 
(for a complete report on our findings, see ref. No.12). 
The results of this investigation were compiled in the form 
of a matrix, showing for each adjective the types of con- 
struction in which it can occur. 
5ummarising some of the observations that could be made 
regarding that matrix, we can say: 
a) the adjectives that fit construction type D (viz. ce_._~r- 
talon, expected, known, ~, said, sure, unknow__~n, unlike- 
Iv) do not occur in any of the other constructions. 
b) the adjectives that fit construction type H (viz. co___nn- 
ten.~t, qlad, happy, ~, sa_~d, satisfied, sorry) do not oc- 
cur in any of the other constructions - with the exception 
of sad, which can occur also in type J. 
c) the adjectives that fit construction type B (viz. 9.~h;~., 
afrai____~d, anxious, careful 2, desirous, ~, fi_._!t, m~id____~, 
prepared, sea_q_~, reluctant, wild_._____~, ~, unable) do not 
occur in any of the other constructions - with the except- 
ion of fit and ~, which can occur also in type G. 
(Of the other adjectives of the corpus, approximately one 
quarter fits only one construction, one half fits two con- 
st~uctions, and one quarter fits three; but since many of 
these adjectives have more than one meaning - e.g. ~ood I = 
'pleasing', ~ = 'beneficial', qood 3 = 'moral' - the 
relational listing of the individual meanings is not an im- 
mediate help in the disambiguation of a given string. Never- 
theiess, it is a step forward from the position where every 
adjective nan to be considered a potential candidate for all 
- 14 - 
ten constructions.) 
The groups of adjectives given under (a), {b), and (c), 
on page 13, constitute extensional definitions, within the 
selected corpus, of ad.iective classes, and we can now exam- 
ine each of these groups to see whether an intensional de- 
finition can be derived from it. 
Group (a) obviously has e common semantic element which 
could be described as 'assessment of probability and/or 
actuality' (of the item to which the adjective is applied); 
Group (b) has a common semantic element which could be 
described as 'a temporary state of mind usually associated 
with a specific cause'; 
Group (c) has a common semantic element which could be 
described as 'attitude or disposition towards an event'. 
I should like to stress that we are at the beginning of 
this kind of investigation and are presumably still rather 
clumsy in formulating valid definitions of semantic ele- 
ments i what is relevant in this context, however, is not 
the efficiency or reiiability of the definitic~ns we tenta- 
tively formulate, but the fact that semantic defif, itions 
c~__~n be derived at all from word groups compiled on the 
~"rength of relational considerations. What we are. in fact, 
~ying to show, is that the analysis of the relations found 
to obtain between the items of phrases or sentences leads, 
first, to an extensional, and eventually, to an intension- 
al semantic classification of the lexical items constitut- 
ing these phrases or sentences. 
This particular sector of adjective cunstrL~ction is, of 
I 
course, not the only area of English grammar which makes it 
seem plausible that semantic classifications of lexemes c~n 
be derived from empirical grouping according to their rela- 
- 15 - 
tional behaviour. The range of relations expressed in Eng- 
lish by prepositions is extremely fertile in this regard, 
but since it is also extremely wide, we have not yet brought 
our survey of it to a definitive conclusion. Partial re- 
sults (13) however, indicate that, here too, relational se 
mantics successfully absorbs a great deal of what, hitherto, 
was considered lexical or unsystemic. 
A comprehensive study of the verb-object relations (a 
still poorly defined area in our operational system) pro- 
mises to yield the perhaps most convincing confirmation of 
our thesis. 
Even a very suPerfici~l examination of a transitive verb 
and the grammatical objects that occur with it, shows that 
the way in which the two are related may vary widely. If, 
for instance, we take the verb "to pay", we find that= 
l) in "He paid the driver" the subject gives up some- 
thing of economic value (e.g. a sum of money); and 
the object specifies the receiver; 
2) in "He paid his bill" the subject gives up something 
of econ. value; the object implies a specific amount 
and that this amount is due to some not further spe- 
cified (but specifiable) remote entity as the right- 
ful receive~; 
3) in "He paid fifty dollars" the subject gives up some- 
thing of econ. value; the object specifies the amount. 
Leaving aside the rarer and the more metaphorical uses 
of the verb (such as= to pay attention, homage, a visit, 
etc.) we can now take a corpus of nouns, try each one of 
them as object of "to pay", and determine which of the 
three described relations it fits. This Will extensionally 
define three noun classes (not necessarily mutually exclus- 
ive, because some of our nouns may fit into more than one 
of the relations) for which we can then tentatively formu- 
late intensive semantic definitions= 
- 16 - 
l') items that can act as receiver of items having an 
economic value; e.g. boy, butche___~r, colleqe, cour__.~t, 
driver, ~, tax collector, tailor, etc.; 
2') items implying a definite econ. value and the fact 
that the implied amount is due to someone; e.g. 
due.___~s, fare, fe._.~e; fin.~e, pos.taqe, ta._~x, etc.; 
3') items indicating a specific econ. value; e.g. any 
numeral followed by an indication of currency, 
and pronominal expressions such as "a lot n, "little", 
"much", etc. 
The verb "topay" is also doubly transitive, i.e. it can 
be constructed with a dative an._.~d an accusative object in 
one phrase, e.g. : 
4) "He paid the butcher ten dollars" (in which "the 
butcher" is the receiver). 
If we now test our corpus of nouns in this const:uction, 
we find, first, that only the items listed in group I' can 
be used as dative object and that they cannot be used as 
accusative object; second, that if items of group 2' or 3' 
are used as direct object, the dative object always plays 
the part of 'receiver'; third~ we find that we have to con- 
side~ su~e new items (occurring only in conjunction with a 
detive object) which are not members of the three iisted 
classes and which, moreover, change the role o ~ the dative 
object, as for instance in: 
5) "He paid his driver a holiday" (where "his driver" 
is the beneficiary of the subject's act). 
We thus get a fourth group of possible objects; their 
intensional definition iS..less obvious, but we can tenta- 
tively put down: 
5') items intended for personal ' " ; consumption e.g. 
drin____~k, a term at colleqe, a mea_______~/, a trip round the 
world, a vacation, etc. 
- 17 - 
One point that is of special interest in these still very 
crude and incomplete results of relational analysis and se- 
mantic classification, is the fact that, although the clas- 
ses of 'receiver' and 'beneficiary' are co-extensive in the 
corpus, they do not seem to create ambiguities, since the 
particular role of an item that is ambivalent in this re- 
spect, is, in any given example, determined by the classi- 
fication of the direct object; in other words, if the di- 
rect object belongs to either group 2' or 3', the dative ob- 
ject plays the part of 'receiver'; if the direct object be- 
longs to group 5', the dative object plays the part of 'be- 
neficiary'. 
This rule, incidentaIly, seems rigid enough to deal with 
at least some phrases which, on a purely experiential basis 
(i.e. 'knowledge of the world'), would have to be rated ra- 
ther odd or unlikely. If, for instance, we came across the 
sentence "5he paid he~ lover a week of clams", we might be 
uncertain how to interpret precisely "a week of clams", 
but we would have no doubt that it had to be something her 
lover could personalIy consume. - And there is consider- 
able reassurance to be got from the implication thot (at 
least in some cases) the logic of relational semantics is 
more powerful than the statistics of factual experience. 
~e could adduce many more examples of transitive verbs 
and relevant object classification, but at the present stage 
of the analysis this wouid add little: the results are in- 
variably suggestive, even indicative, but they cannot be 
considered conclusive. Therefore, I shall merely summarise 
what we expect from tnese investigations. 
On the basis of work accomplished during the last 18 
months, it appears that the semantic classes derived from 
relational analysis are recursive and that their num- 
- IB - 
bet will remain usefully smaller than the number of classi- 
fied items. There will, of course, be many more semantic 
classes than grammars have hitherto contemplated - but giv- 
en the versatility of natural language, this should not 
really surprise us; nor is there any need to be particular- 
ly pessimistic about the possibility of implementing such 
a voluminous and intricate data base in a computer system. 
We all have seen how the early computational linguists' 
worries concerning storage capacity and processing speeds 
have ~ been made to appear anachronistic by technologic- 
al progress, and this progress does not yet seem to be any- 
where near its ceiling. 
As to the theoretical implications of our kind of lan- 
guage analysis, I should like to put forward one suggest- 
ion. Assuming that we can derive (and the material present- 
ed here does imply precisely this) a satisfactory semantic 
classification of lexical items from their relational pro- 
perties defined in terms of an adequately differentiated 
syntax, it may be more profitable (and, perhaps, also more 
correct) to view syntax and semantics, not as a pair of 
mutually exclusive opposites, but rather as the axes of a 
continuum of meaning; every semantic element, particle, 
feature (or howevec we want to call itl would, in such a 
frame of reference, have both a relational and a lexical 
coordinate - which would not only make it possible for us 
to discuss one and the same item from two points of view 
without contradiction, but, I believe, it would also be a 
useful advance towards an economical representation of lin- 
guistic data in computational procedures. 
J 
- 19 - 
Semantics and the Syntactic Classification of Words 
A B5 T R ACT 
Traditional grammars classify words according to gener- 
ic syntactic functions or morphological characteristics. 
For teaching humans and for descriptive linguistics this 
seemed sufficient. The advent of computers has changed the 
situation. Since machines are devoid of experiential know- 
ledge, they need a more explicit grammar to handle natural 
language. Correlational 5rammar is an attempt £n that di- 
rection. The paper describes parts of correlational syntax 
and shows how a highly differentiated syntax can be used 
to establish word classes for which an intensional semant- 
ic defJ.nition can then be found. It exemplifies this ap- 
proach in two ares of grammar: predicative adj,~c~ives and 
transi~:ive verbs. The classification serves to eliminate 
ambigu#ty and spurious computer interpretation~ of natural 
language sentences. 
Author's address: Ernst yon Glasersfeld 
Georgia ~nstitute for Research 
711, C & 5 Bank Uldg. 
ATHENS, 5a., 30601, U.5.~. 

References

i) Stephen Ullmann, The Principles of Semantics, Blackwell, 
Cxford, 1957. 

2) 5ilvio Ceccato et al., Linguistic Analysis end Program- 
ming for Mechanical Translation, Feltrinelli, Milan, 
• 1960, and Gordon & Breach, New York, 1962. 

3) E.v.&lasersfeld and P.P.Pisani, The Multistore System 
MP-2, Georgia Inst.for Research, Athens (Georgia, 
U.5.A.), 1968, and 

E.v.&lasersfeld and P.P.Pisani, 
The Multistore Parser for Hierarchical Syntactic 
Structures, to be published in the Communications 
of the ACM, 1969. 

4) Noam Chomsky, Aspects of the Theory of Syntax, M.I.T. 
Press, Cambridge (Mass.), 1965. 

5) Charles J. Fillmore, Entailment Rules in a Semantic 
Theory, Uhio State University, Columbus (Ohio), 
1965. 

6) Hans Reichenbach, Elements of Symbolic Logic, Free Press 
Paperback, New York, 1966. 

7) E.v. Glasersfeld, P.P.Pisani, J.Burns, B.Notarmarco, B. 
Dutton, Automatic English Sentence Analysis, Re- 
ports ILR5 T-If and ILR5 T-14, IDAMI Language Re- 
search Section, ~ilan (Italy), 1965/66. 

8) Jehane Burns° A 5cneme for Semantic Controls in Auto- 
matic Sentence Analysis, Literary and Linguistic 
Computing Centre, University of Cambridge (England), 
1968. 

Fred 5ommers, Predicability, in Max Black (editor), Phi- 
losophy in America, Cornell University Press, 1965. 

5ilvio Ceccato, Un Tecnico fra i Filosofi, vol. l and If, 
Marsilio Editore, Padova (Italy), 1964/66. 

Henry Ku~era and W.N.Francis, Computational Analysis of 
Present-Day American English, Brown University, Pro- 
vidence, R.I., 1967. 

E.v. Glasersfeld and B.Notarmarco, Some Adjective Clas- 
ses derived from Correlational Grammar, to be pub- 
lished in American Speech, 1969. 

E.v. Glasersfeld, An Approach to the Semantics of Pre- 
positions, Proceedings Of the Symposium on Computer- 
related Semantic Analysis (LaB Vegas, 1965), Wayne 
State University, Detroit, Mich,, 1967. 
