JEAN PIERRE PM~LET 
COMPUTATIONAL LINGUISTICS AND LINGUISTIC THEORY 
The present paper is an attempt to justify and explain the direction 
of present research in Ottawa (Carleton University and also University 
of Ottawa) on Computational manipulation of speech. Our actual real- 
izations are not necessarily original; rather, we are trying to make use 
of the findings of other workers, assembling them, however, in a dif- 
ferent way, on the basis of a novel conception of linguistic organization. 
1, DESIDERATA FOR COMPUTATIONAL LINGUISTICS 
The main problem facing computational linguists, when they want 
to make use of linguistic theory, is the divergence of aims. Basically, 
computational linguists attempt to design devices which will automat- 
ically manipulate instances of speech (henceforth discourses) in specified 
ways, to obtain particular results. Linguistic theories, on the other hand, 
are concerned with the overall structure of a language, and consider 
what happens in particular cases as the interaction between the overall 
linguistic competence (general structural knowledge) and many unde- 
termined factors of individual performance. Hence the computational 
linguist is left to find for himself pragmatic solutions to his particular 
problems. We examine first what sort of theory the computational 
linguist could make good use of. 
1.1. It is possible to distinguish between formal and functional 
theories according to whether the main criteria for classification and 
description of units are based on their formal properties alone or rely 
on the role they play in conveying meaning. It would appear, given the 
goals of computational linguistics, that functional theories might be 
more useful, since they give us direct access to the meat: meaning. How- 
ever, most of them are not precise enough to be used in a computer 
environment. The best attempt at using one of them was made by T. 
358 JEAN" PIERRE PAILLET 
WINOG~.AD (1972). It is significant that Winograd had to reformulate 
much of systemic grammar in order to be able to use it. 
1.2. Accordingly, most of the work in computational linguistics 
relies on formal theories of language, starting with distributional lin- 
guistics. The problem here is reversed. A good formal theory will pro- 
vide interesting possibilities for manipulating discourses, but gives no 
reason for doing so. More explicitly, the schemas of manipulation sought 
by computational linguists must have some interesting functional prop- 
erty: semantic invariance, logically oriented semantic relations, or 
the like. Hence the necessity for a theory presenting a well defined 
semantic aspect, as well as allowing manipulations of the expression 
aspect of discourses. 
1.3. There are some such theories, among which most conspicuous 
are generative semantics and Lamb's stratificational grammar. R. 
B~NNICK (1969) and S. LAMB (1973) respectively, have argued for ap- 
plying these theories to computational work on language. While widely 
different in most respects, these two theories share two interrelated 
features. First, they are theories of langue, that is, they try to account 
for wellformedness and systematic interrelationships within a set of 
objects, called sentences of the language. Second, as a consequence 
of this feature, they establish a systematic, deterministic relation be- 
tween expression and content. In the case of generative semantics, this 
systematic relation is oriented from content to expression; the under- 
standing of a particular sentence must then be viewed as the ~ undoing ~> 
of transformations, which often yields ambiguity. In the case of strati- 
ficational grammar, which is basically non-oriented, the understanding 
of a sentence appears as the activation of a static network of connections 
from the expression end. At the content end, this yields a complex of 
activated lines which, hopefully, represents the meaning of the sen- 
tence. There is here an interesting distinction, not explicit in genera- 
tive semantics, between the static set of connections and their dynamic 
activation in a particular speech act. However, both formulations rely 
heavily on wellformedness. Any speech act which is not perfect must 
be blocked somewhere in the encoding or decoding. This is normal if 
the only interaction between speaker and hearer is to take place through 
wellformed speech expressions. On the other hand, such a requirement 
is known to be unnatural in the human environment, where wellform- 
edness is patently non-essential. Thus, if realizable, generative seman- 
COMPUTATIONAL LINGUISTICS AND LINGUISTIC THEORY 359 
tics or stratificational grammar are at best an artificial simplification of 
human language, for computing purposes. Even thus, we know that 
a particular sentence may have different ~ meanings ~) in different lin- 
guistic or non linguistic contexts; therefore, it is illusory to look for 
a purely linguistic mapping between content and expression. We label 
direct those theories which posit such a mapping, and mediate those which 
allow for (and, ideally, specify the conditions of) the insertion of extra- 
linguistic information in the speaking or understanding processes. 
What is needed, thus, for computational linguistics, is a formalized, 
semantically specific, mediate theory of language. 
2. SYNTAX AND THE DICTIONARY 
2.1. Among other things, the adoption of a mediate theory of 
language forces on us an interesting consequence: it is no longer possi- 
ble to derive the wellformedness of expression from that of content, or 
conversely. Each has to be specified in its own terms. This in turn re- 
quires the creation of an independent metalanguage for content, a task 
started, with a limited scope, by symbolic logic. It also changes the 
outlook of syntax. In most formal theories of syntax, a large apparatus 
is devoted to selectional restrictions. Inasmuch as these are the man- 
ifestations of semantic wellformedness requirements, to be handled 
anyhow in the semantic description, we can now reduce syntactic de- 
scription to the specification of order structures and ~ strict subcategori- 
zation ~. As a consequence, we may ignore, in syntax, such notions as 
that of transformation, which characterizes selectional invariance. 
2.2. In order to see what kind of device might be used for such 
simplified syntactic description, a detour through morphology is in 
order. 
In a recent paper, M. HALLE (1973) offers suggestions for the treat- 
ment of morphology in a generative grammar. 1 1~. BINNICK (1973) has 
written a penetrating review of Halle's paper; within the generative 
framework, he points out a very important feature of Halle's proposal: 
the confusion between irregularity in expression and irregularity in 
1 The concern with morphology is rather new in generative grammar, but has long 
been manifested by other linguists, e.g.G. Gtr~LAUM~ (1971), or M. A. K. HaLLmAY 
(1961), in a functional approach. 
360 JEAN PIERRE PAILLET 
content. According to Binnick, the two types of structures should be 
carefully distinguished (which is precisely what a mediate theory does 
better than a direct theory). On the other hand, there are some regular- 
ities in morphology, which involve a systematic correspondence be- 
tween expression and content: for instance, in English, it is always pos- 
sible to nominalize a verb. Whenever a "strong nominalisation" (as 
in the arrival of the prime minister) is not available, it is possible to use 
a" gerund" (as in the second coming). It seems that the regular formations 
are less acceptable when a lexical possibility exists; they are therefore 
to be considered as "otherwise" solutions. We have here a suggestion 
that the description of the lexicon must make extensive use of disjunc- 
tive ordering (this, incidentally, is built into the structure of strati- 
ficational theory). 
The dictionary is "needed anyhow" either to list irregularities in 
morphology, or to account for the link between expression and con- 
tent aspects of morphemes, or probably for both. This dictionary ap- 
pears as a list of arbitrary associations between some content structures 
and some expression structures: nothing new, on this point, since Saus- 
sure. We may borrow an argument from the practice of generative 
grammarians, and argue that, since the dictionary is needed anyhow, 
we might as well use it for syntax as well. This will also remind one 
of F. Dr SAtrssu~ (1966), who thought that syntax could not entirely 
be part of langue: only "stereotyped expressions " would. Translate: 
some recurrent syntagmatic patterns of morpheme classes are part of 
the linguistic tool box, just as morphemes are. The rest is up to the 
speaker. 
2.3. Such a view would have two interesting consequences. The 
first one, from the computational point of view, is that whatever de- 
vice is needed for handling the dictionary it could do duty to handle syn- 
tagms as well. The second, from the theoretical point of view, is that 
a theory of syntactic illformedness would be naturally statable, by anal- 
ogy with morphology. To take an example from Halle's paper, arriv- 
ation is wellformed although non occurring, because of the specific prop- 
erties of-ation and arrive. Morphological wellformedness can be stat- 
ed in terms of contiguity relationships. On this model, we might want 
to state syntactic conditions in terms of contiguity relationships as 
well: that is precisely what string structure theory does. A string struc- 
ture description (A. K. Josm, 1970; z. s. HARRtS, 1965) starts with a 
vocabulary of strings, each of which has occurrence properties, stat- 
COMPUTATIONAL LINGUISTICS AND LINGUISTIC THEORY 361 
able in terms of which strings or parts of strings they may or must 
be contiguous with. 
2.4. One frequent objection to assimilating morphology and syn- 
tax can be dealt with here. It is often argued that the types of recur- 
siveness found in syntactic structures do not appear in morphology. This 
would be the reason why a finite state treatment, which is feasible for 
morphology, is not adaptable to syntax. This argument reflects a biased 
view of morphology, influenced mainly by indoeuropean languages. 
As a matter of fact, Eskimo morphology, for instance, exhibits 
instances of full recursiveness. Any solution to the problem of describ- 
ing Eskimo words will be a likely solution for describing English 
syntax, and conversely. In his axiomatic presentation of the string struc- 
ture of English, Harris offers such a solution, based on the fact that 
only very few types of syntagms are recursive: for English syntax, 
they are those labeled c, Y~, and A by Harris. One can then describe 
English order structures by a finite state diagram, where some of the 
transitions are labeled by one of these three symbols. These transitions 
imply a (recursive) call to the description of the corresponding syntagm. 
The systems developed by W. A. WOODS (1970), J. THORPE (1968), 
and D. G. BOBROW (1969) exploit a similar trick. The only theoretical 
requirement is to keep track of the depth of recursion when run- 
ning through the diagram. Practically, the depth of recursion will be 
limited by the capacity of the processor; this turns out to be an advan- 
tage when one tries to model human performance, which is notoriously 
limited in this respect. 
The transition diagrams used in this type of treatment are conspic- 
uously similar to the organization of dictionaries for computer use. 
Even if no other reason existed, this similarity should prompt the com- 
putational linguist to explore such a possibility. 
2.5. Such a scheme would not look very new to stratificationalists. 
They could argue, as pointed out earlier, that their theory distinguishes 
between the description of regularities and idiosyncracies on the one 
hand, and the use made of them on the other. They might even argue 
that what I call dictionary is nothing but the set of static relationships 
represented in a stratificational network. However, there are two basic 
differences to be dealt with. One is that I argued earlier that ira compu- 
tational device is to model linguistic performance adequately, it must 
be insensitive to noise, either in the form of occultation of the signal 
362 J~AN PIERI~ VAILL~T 
(e.g. unrecognizable characters in a printed form, or morphemes of 
unknown classification) or in the form of" mistakes " (e.g. occurrences 
of ill-formedness). Thus recognition of a message cannot rely on the 
"activation" of a stratificational network, which would be blocked in 
both cases. T. 1~. HOrM^N~ (1971) has proposed a device which he 
calls Vigilant Memory, and which is capable of " recognizing" mor- 
phemes in a string, in the presence of noise. It performs very well in 
the case of unrecognizable or substituted characters, slightly less well 
in the case of insertions or deletions, and still less well, but adequately, 
it seems, in the case of metatheses, which amount to a double substi- 
tution. We are now working on the task of combining the possibilities 
of the vigilant memory with the economy of conventional dictionary 
lookup procedures. One important point is that the output of a vigilant 
memory is a decision on the identity of some form, given a possibly 
faulty input. This output can be input to another vigilant memory, 
which will recognize other forms in terms of the previous ones (e.g. 
words or syntagms in terms of morphemes). One thus obtains the el- 
egance of formulation of stratificational networks with a useful insen- 
sitiveness to noise. 
2.6. The other important difference between our proposal and 
stratificational grammar stems from the fact that the latter is a direct, 
structural theory: it offers no way of representing an isolated construct 
of expression or content independently of the general network of rela- 
tionships describing the overall structure of the language. We propose, 
on the contrary, to have two sets ofwellformedness characterizations, one 
for expression and one for content. The link between the two is to 
be seen as a collection of procedures, called by the various recognizable 
forms of expression, which build forms of content according to the well- 
formedness schemas of content. To put it another way, units of expres- 
sion (morphemes, syntagms, sentences, etc.) do not have, or carry, or 
correspond to, a particular meaning, but induce certain computations 
whose result is some meaning structure. This view, which Winograd 
also seems to hold, is at the center of our Integrative Semantics (J. P. 
PAILt~T, T. l~. HOrMANN, 1972). 
3. INTEGRATIVE SEMANTICS 
3.1. This proposal for a semantic theory started as an effort to 
develop an adequate notation for semantic structures, which should, 
COMPUTATIONAL LINGUISTICS AND LINGUISTIC THEORY 363 
among other things, be free from syntax-induced biases (J. p. PAILLET, 
1973). By combining the insights of logicians like Frege and Tarski, 
and of linguists like Hjelmslev and Tesni~re, we come to distinguish 
first semantic forms from semantic processes. 
3.2. Semantic processes are of several kinds. The simplest kind to 
understand is the object of logical proof theory. It consists in manipulat- 
ing semantic forms to obtain other semantic forms systematically re- 
lated to the original ones. Our cognitive activity, however, does not 
consist merely in manipulation of abstract forms. We find two other 
kinds of semantic processes, also represented in Winograd's system. 
Interpretation processes have long been studied by logicians. They 
consist in putting the abstract forms which we hold into correspond- 
ence with a portion of some universe. Evaluation processes consist 
in reaching decisions as to the response which is appropriate in a partic- 
ular case. An important point to keep in mind is that such processes 
are not limited to speech activity. They are ever present in our con- 
scious life. We can look upon speech as a particularly effective way 
of nudging someone's cognitive activity in a certain direction. That 
this particularly effective tool need not always be successful is apparent 
in all cases where the hearer fails to understand, misunderstands, etc. 
3.3. In a human language, expression is contrained by wellform- 
edness conditions which limit the possible use and arrengement of mor- 
phemes. Consequently, expression is organized in such units as phrases 
and sentences, which need not correspond to "complete thoughts ", 
or, more specifically, to complete structures, of content. If a new form 
of content is to be transmitted to a hearer, it will be imparted piecemeal, 
by the use of separate (usually successive) sentences. The process of 
building a new structure of content, which, for obvious reasons, we 
call integration, has to be directed by specific devices (morphemes or 
constructions) called integration functors. Similarly, there are devices 
to direct interpretation, such as the deictics, which make a description 
definite, often with recourse to non-linguistic information. Finally 
there are also evaluation functors, which direct the evaluation of the 
appropriate response. These functors could not be said to have mean- 
ing independently of the precise conditions of use. Their action must 
therefore be described in terms of the computational procedures which 
they call for in the hearer. The simplicity argument suggests that we 
treat lexical items in the same way. 
364 JEAN PIERRE PA~LLET 
4. A MODEL OF THE HEARER'S PERFORMANCE 
4.1. It is apparent from the foregoing that our proposal is basically 
hearer-oriented. We believe, indeed, that the speaker must have recours- 
ed to a set of heuristics, taking into account his knowledge (or imagined 
knowledge) of the hearer, his status with respect to him, and many 
other variables. In most tasks of computational linguistics, however, 
the full complexity of human interactions is fortunately not present. 
There are a number of simplifying assumptions one can make as to 
subject matter, vocabulary, style, etc., so that the task of giving expres- 
sion to given semantic structures is not overwhelmingly complex. 
However, from the point of view of linguistic theory, it is much easier 
to formalize the understanding of speech (phonological perception 
excluded) than the production of sensible discourses. 
4.2. A hearer is primarily a cognitive system capable of building, 
holding and manipulating semantic forms. These forms consist of 
individual objects, descriptions for and relations between these objects, 
and various interacting modalities such as quantifiers, moods, aspects, 
and the like. The manipulation of semantic forms may yield other se- 
mantic forms as well as judgements of contradiction, consistency, im- 
plication, and the like (note that these judgments are not limited to 
linguistically induced semantic forms: optical illusions are examples of 
visual paradoxes, i.e. contradiction between two cognitive forms vis- 
ually extracted from the same object). 
4.3. It is conceivable, although not shown, that the cognitive pro- 
cedures used for building semantic forms in non linguistic situations are 
the same as those called forth by lexical items and integrative functors: 
this is a restatement of Whorf's hypothesis (B. L. WHom~, 1965). In 
any case, lexical items and integration functors are seen as subroutine 
names (to use computer terminology) whose perception usually trig- 
gers the corresponding building procedure. Similarly, integration ftmc- 
tors call forth procedures for interrelating parts of semantic forms which 
are in the building or already built. 
4.4. A hearer is also able to relate semantic forms to a universe of 
interpretation. This may be part of what we traditionally call attention. 
In this case, the interpretation functors may be said to direct the heater's 
COMPUTATIONAL LINGUISTICS AND LINGUISTIC THEORY 365 
attention. Their action can also be represented by procedures in which 
the best possible referent for a given description is computed. Some in- 
terpretation functors direct the selection of an appropriate universe of 
reference. Similarly performatives and other devices are calls for proce- 
dures of evaluation, in preparation for an appropriate response. 
4.5. Naturally, the various kinds of procedures called by elements 
of expression in a discourse are often insufficient in themselves to pro- 
duce all the results mentioned. Part of the speaker's communicative 
ability consists in leaving out much of the information needed, stricto 
sensu, to build, interpret and evaluate semantic forms, and in selecting 
only what is absolutely necessary to lead the hearer in the proper di- 
rection: hence the term "nudging " used in 3.2. The rest of the infor- 
mation, left out by the speaker, has to be provided by the hearer, either 
through autonomous manipulation of his semantic forms, or through 
recourse to his store of previous information: his "knowledge of the 
world ". In order to implement this last aspect, R.. QUILLIAN'S (1969) 
concept of a semantic network is very attractive, and is probably pref- 
erable to Lamb's version, cognitive networks, which do not to my 
knowledge incorporate any notion of semantic proximity. 
4.6. If syntax is indeed reducible, in its "order structure" aspect, 
to a set of patterns recognizable by a device like a Vigilant Memory, 
we have an explanation of why syntactic illformedness is not more de- 
structive. The only thing which counts, from the heater's point of 
view, is to be able to recognize some pattern present in his syntactic 
dictionary, which will then call some building or integrating procedure. 
Illformedness becomes destructive only when the error-correcting ca- 
pabilities of the vigilant memory are overwhelmed. On the other hand, 
it is always possible, even when recognition occurs, to detect syntactic 
illformedness through the workings of the memory. 

REFERENCES 

t~. BINNICK, Generative Semantics and 
Automatic Translation COLING, Stock- 
holm, 1969. 

IL. BINIqlCK, Epilegomena to a Theory of 
Word Formation, unpublished, 1973. 

D. G. BOBROW, J. B. FRASm~, An augmen- 
ted state transition network analysis pro- 
cedure, in Proc. Int. Joint Conf. on Art. 
int., Washington, D.C., 1969. 

G. GUIL~AUM~, Le cons de linguistique, 
vols. 1 & 2, Qu6bec, 1971. 

M. H~LE, Prolegomena to a Theory of 
Word Formation, in ~Linguistic In- 
quiry ~), IV (1973) 1. 

M. A. K. HAtUDAY, Categories of the 
theory of grammar, in ~ Word ~, XVII 
(1961). 

Z. S. HAaaIS, String Analysis of sentence 
structure, The Hague, 1965. 

T. IL. Ho~raAN~, Lexical redundancy in 
Speech Perception: Vigilant Memory, in 
VII International Congress of Phonetic 
Sciences, Montreal, 1971. 

A. K. JosHx, String adjunct grammars, in 
TDAP no. 75, Philadelphia, 1970. 

S. LArca3, Stratificational Linguistics as a 
basis for Machine translation, in MAKKAI, 
LOCKWOOD (eds.), Readings in Stratifi- 
cational Linguistics, University of Ala- 
bama Press, 1973. 

J. P. PAIl.LET, Prdrequis pour une thdorie 
semantique, in ,Cahier de linguisti- 
que ~, no. 2, Montr&l, 1973. 

J. P. PaILL~T, T. tL. HorMA~, Integrative 
Semantics 1, SILG Monograph No. 2, 
Carleton University, Ottawa, 1972. 

g.. QmLLIAN, The teachable Language 
Comprehender, in ~ Communications of 
the ACM~), (1969). 

F. DE Saussum~, Cours de linguistique gd- 
n&ale, Paris, 1966, pp. 170-171. 

J. THORPE, P. ~BRATLEY, H. DEWAa, The 
Syntactic analysis of English by Machine, 
in D. MIcHte. (ed.), Machine Intelli- 
gence 3, New York, 1968. 

B.L. WHOR~, Selected writings, Cam- 
bridge (Mass.), 1965. 

,T. WINOGaAD, Understanding Naturar 
Language, New York & London, 1972- 

W. A. WOoDs, Transition networks fol 
natural language analysis, in * Commu. 
nications of the ACM *, XIII (1970) 10. 
