CONSIDERATIONSFOR 
COMPUTATIONAL THEORIES OF SPEAKING: 
SEVEN THINGS SPEAKERS DO 
John H. Clippinger, Jr. 
Teleos 
Cambridge MA 02138 
I. INTRODUCTION 
Fundamental progress has to do with 
the reinterpretation of basic 
ideas. 
Whitehead 
The opinion which is fated to be 
ultimately agreed to by all who 
investigate it is what we mean by 
the truth and the object 
represented in this opinion is the 
real. 
C. Pierce 
Any discipline in the course of its 
maturation experiences identity crises in 
deciding what it is and where it is going. 
Often an initial commonality of interests, 
methods, and even disgruntlements hides an 
eventual diversity of purposes and goals. 
It is often at this time that self-conscious 
decisions have to be made regarding the 
overall scope and directions of its 
enterprises. To a limited degree, 
computational linguistics is beginning to 
experience its own birth pains. Bedfellows 
who did not seem so strange initially are 
beginning to appear progressively more so. 
Perhaps now is a time for stock taking and 
some preliminary self reflection. 
The principal intent of this paper is 
to sort out what I feel to be some of the 
more basic purposes of computational 
linguistic research, and then to discuss 
some of the explanatory requirements for one 
aspect of such research, the modeling of the 
human speaker. The bulk of the paper will 
focus upon some of the more basic issues in 
describing human speaking by a computational 
model. 
II. THE DUAL CHARACTER OF 
LINGUISTIC RESEARCH 
COMPUTATIONAL 
Superficially, what distinguishes 
computational lingustics from other hybrid 
forms of linguistic research 
(sociolinguistics, psycholinguistics, 
ethnolinguistics, generative linguistics, 
etc.) is the concern to represent linguistic 
descriptions in terms of computer programs. 
The common bond in computational linguistic 
research is the computer, or stated somewhat 
more broadly, the information processing 
approach. However, while this common 
interest does serve to distinguish this form 
of linguistic research from a growing number 
of others, it does, I believe, masque some 
more elemental differences over the purposes 
of using computer programs to describe 
language. There seem to me to be two 
complementary but ultimately different 
directions in computational linguistics; the 
122 
first is concerned primarily with improving 
and developing the software technology 
whereby computers can use and process 
natural language; under this heading would 
be machine translation, question and 
answering systems, automated secretaries, 
and assorted text processing systems; the 
second is concerned with using the computer 
as a means of developing an accurate and 
empirically valid computational model of the 
linguistic and cognitive behaviors of a 
human speaker. To date these two interests 
have been regarded as one; the rationale 
being that human beings and machines would 
use analogous mechanisms in their 
comprehension and generation of language to 
make the system work efficiently. But if 
one's goals are descriptive, then I don't 
feel that any such latitude is justified and 
that the use of different models and 
approaches should be justified with 
reference to certain theoretical criteria 
and empirical evidence--to the extent it is 
possible. This of course, is not to say, 
that the two branches of research, the 
technological and the descriptive, will not 
or should not share their results or 
influence one another. Only that they each 
have different objectives and hence 
different methodological and theoretical 
requirements. 
Having made the distinction between 
technological and descriptive forms of 
computational linguistics, I am now going to 
have to hedge a bit. Because, whereas both 
forms of research do differ with respect to 
specific research objectives, they do share 
a common meta-language and methodology, and 
are dependent upon that language for their 
success. I am now referring to the 
dependence of computational linguistics upon 
abstract theories of programming and 
computation. Yet even this distinction 
between the abstract and the applied 
modeling is not so clearcut, as practical 
programming requirements, such as language 
comprehension, affect the design of 
programming languages, and they in turn 
affect how specific problems are represented 
and understood. However, I suspect that as 
the field of computational linguistics 
matures, the descriptive and the 
technological forms of research are going to 
exert different selective pressures on 
computational theory to produce different 
types of programming languages, one being 
more specialized and efficient, and the 
other being more generalized and probably 
less "practical". Yet here too a 
qualification is needed, for where the more 
general and human-like problem solving 
language is shown to have practical and 
technological uses, then it too will have a 
technological application. 
Therefore, any advance in improving a 
machine's facility with natural language 
would analogously constitute an advance in 
our understanding of human language 
behavior. The hedge word here, of course, 
is analogously, for it is not clear to me 
whether a program designed for very 
specialized and restricted purposes can be 
any more than superficially analagous to a 
human speaker. A simulated secretary for 
l 
! 
l 
i 
! 
I 
I 
! 
! 
t 
! 
I 
! 
i 
! 
,! 
! 
example has far fewer concerns than an 
actual secretary and therefore is designed 
and programmed for more specialized 
purposes. Moreover, and perhaps most 
significantly, the simulated secretary has 
no inherent intentionality of its own; it 
doesn't care what others think of it, it 
does not need to please, it isn't 
distracted, and it doesn't deplore 
chauvinism and resist demeaning requests. 
Certainly there are analogies between what 
the simulated and the real secretary do, as 
both perform similar clerical and linguistic 
tasks, but these similarities are quite few 
when compared to their dissimilarities. 
Moreover, it is highly questionnable whether 
the programming techniques and methods used 
to model this very restricted form of 
behavior can be easily generalized to other 
types of behavior. Most experience in 
artificial intelligence suggests the 
contrary. The point which I believe needs 
to be made is that while, yes, at some 
level, the information processing behavior 
of machines and programs is similar or 
analogous to that of human beings, that for 
the most part, this level is too general to 
be of much use in understanding human 
thought and speech at present. Hence, 
computational linguistics and artificial 
intelligence should begin to distinguish 
between those research goals that are 
technologically oriented and those that are 
intended to be descriptive of human action. 
For if one's goals are purely technological 
such as machine translation or a question 
and answering system, then one is certainly 
justified in using virtually any mechanism 
one wants. 
The problem is that in our ignorance of 
how higher order and unobservable human 
problem solving processes and mechanisms 
work, we have to rely upon experiments with 
abstract programming languages and simulated 
environments just to give us ideas about how 
such mechanisms might work. Certainly this 
is a reasonable strategy to pursue at this 
level of abstraction and experimentation. 
But when the intent then becomes the 
formulation of a computational model of some 
form of human activity, then I believe that 
there is a need to provide some theoretical 
or empirical justification for the model. 
For example, if the model is of a human 
interpreter of discourse, does the model 
make the same mistakes as people do, does it 
resolve ambiguity in the same way, can it 
make similar inferences from discourse as 
people, etc.? To answer these questions 
requires considerably more empirical 
evidence than we currently have. 
The real issue here is accountability. 
For programming languages and technological 
applications of computational linguistics 
this is relatively easy to determine. But 
when the intent is descriptive, that is, 
when a model is presented as representing 
some human intellectual skill or process, 
then the issue of accountability becomes 
murky indeed. For it is not enough to say 
that it works, or that expert judges cannot 
distinguish the model from the real thing; 
Weizenbaum's Eliza program showed how easy 
it is to attribute powers to the computer it 
123 
doesn't possess and Colby's 
indistinguishability tests have similarly 
shown the gullibility of trained 
psychiatrists. Consequently there is a need 
on the part of descriptive computational 
linguistics to both specify the types of 
tasks a model of a speaker or listener 
should perform and in addition, to the 
extent it is possible, the manner in which 
they should be performed. 
The problem then becomes knowing what 
it is one is wanting to describe and account 
for. And this is no easy matter, for as 
Wittgenstein noted (Wittgenstein, 1953), 
"the aspects of things that are most 
important for us are hidden because of their 
simplicity and familiarity (one is unable to 
notice something--because it is always 
before one s eyes)." Chomsky makes a similar 
point (Chomsky, 1970) "As native speakers we 
have a vast amount of data available to us. 
For this reason it is easy to fall into the 
trap of believing that there is nothing to 
be explained, that whatever organizing 
principles and underlying mechanisms may 
exist must be "given" as the data is given." 
Traditionally linguists -- Chomsky among 
them -- have focussed upon the grammatical 
aspects of language usebecause it was so 
"given" and assumed that in accounting for 
the generation of grammatical utterances 
that they have in some way described at 
least some of the mechanisms involved in the 
generation of human speech. But as Max 
Black (Black, 1970), and a number of 
sociolinguists have pointed out (Hymes, 
1972; Labov, 1974) to be able to 
characterize the grammaticality of a 
language with a finite set of rules does not 
mean that speakers of that language utilize 
such rules in their speaking. Therefore to 
my mind there is the real question of the 
value of such descriptions in accounting for 
either the individual or the collective use 
of language. Consequently when dealing even 
with the most obvious of linguistic "facts" 
there is a need to consider the more general 
role of language as a communicative, problem 
solving, and expressive medium. Similarly 
there is a need to characterize the type of 
explanation being sought; whether we want to 
explain language behavior in terms of 
reasons and intentions, or whether we want 
to describe it extentionally, in terms of 
causes. 
This latter distinction is one which 
Toulmin (Toulmin, 1970), Radnitsky 
(Radnitsky, 1970), Dennett (Dennett, 1975) 
and Goffman (Goffman, 1974), make in their 
varius dicussions of the types of 
explanations appropriate to the human or 
social sciences. Taking Toulmin's argument 
for the moment, causes are essentially like 
physical laws, as they are beyond our 
control, whereas "actions done for reasons 
can be regarded as applications of 
procedures (methods of calculation, 
techniques, rituals, or other formalized 
modes of behavior) that we have learned 
during our life time." While ideally it may 
be possible to explain human language 
behavior in terms of causal descriptions, 
for example by neurological models, it is 
doubtful that this is the type of 
description that computational linguistics 
is seeking. Rather the appeal of the 
computational approach is precisely in its 
capacity to characterize the symbolic 
procedures that we use and have learned. If 
this orientation is accepted, then I believe 
that the obvious conclusion to be made is 
that language behavior should be described 
and explained as a learned intentional 
activity. Going one step further, then, one 
of its goals should bethe explanation of 
the various reasons and purposes for 
different types of linguistic behavior. 
However, to answer such questions 
satisfactorily requires a twofold 
description of the reasons for a procedural 
action. For taken from a historical or 
diachronic perspective, a given procedure is 
created to solve a particular set of 
problems at a particular time; hence its 
existence or reason for being is set in 
time. However, the same procedure may then 
again be used at some later time to achieve 
some other action perhaps unrelated to the 
first, and therefore has a reason 
independent of its derivation. 
Consequently, the reason a particular 
procedure was used can be explained in terms 
of the reason for its coming into being, or 
in terms of the immediate effect or result 
it was invoked to achieve. Psychotherapy is 
full of such cases: "Why do you smoke?" 
"Because I was bottle fed" or "Because I 
like the taste?" Similarly, for example, in 
understanding a speaker's use of 
intensifiers or dubitatives it is important 
to know whether they have a specific local 
meaning or whether they are a part of some 
standard discourse style -- either personal 
or cultural. For modeling purposes the 
differences are important, as they entail 
differences in representation. 
III. CONSIDERATIONS FOR A COMPUTATIONAL 
THEORY OF SPEAKING 
So far we have established that that 
segment of computational linguistic research 
concerned with describing human speakers 
should take some initial steps towards 
saying more concretely what it is trying to 
account for. Certainly if we are going to 
evaluate a piece of research in this field, 
we should have some preliminary consensus as 
to what a successful model of speaking 
should be able to do. Likewise, there 
should be some common understanding as to 
the types of descriptions and explanations 
that are being sought. 
Taking the second question for the 
moment, I think that one of the unique 
contributions of a computationally based 
methodology to the human sciences is its 
capacity to give formal and .teleological 
descriptions to complex forms of symbolic 
behavior. Therefore one of the principal, 
if not inviolate, ingredients of a 
computational theory of speaking should be a 
computationally based methodology. This 
requirement would to my mind disqualify 
statistically based, and to a lesser extent, 
predicate calculus based, models. 
Furthermore, I feel that it is incumbent 
upon a modeler of human discourse to make 
124 
informed use of more advanced programming 
concepts, as problems resolved at this level 
eventually contribute a richer and more 
complete representation at the purely 
descriptive level. For example, although 
notions such as frames, mini-worlds, demons, 
actors, pattern matching and the like are 
essentially programming concepts, they do 
represent solutions or partial solutions to 
programming complex symbolic environments 
which in all likelihood are far simpler than 
those encountered by the everyday speaker. 
Consequently a computational model of 
speaking requires these techniques and 
methods both technically and theoretically. 
Now finally what a computational theory 
of speaking should in part, at least, 
account for. To date we seem to know very 
little about the types of mechanisms 
involved in the creation, formation, and 
regulation of speaking (here also, is 
included discourse and conversation). And 
while it would certainly be reassuring to 
compartmentalize speaking as a clearly 
bounded activity, both sociolinguists and 
phenomenologists have effectively dashed 
that hope. 
Since we can't say what speaking 
is -- if indeed it is any single thing 
(language does a disservice here), we can 
point out some of those things that people 
appear to do while speaking. It follows 
then that a model of human speaking should 
be able to do these too. 
Seven Thin~s Speakers Do 
I. People normally initiate a 
statement or a discourse out of a 
desire to be understood. They have 
some goal, or effect they want to 
achieve which they think speaking 
might facilitate. 
2. Speakers alter what they are saying 
according to the physical and 
social context in which they are 
speaking. This in turn affects 
what is said and how it is said. 
3. Speaker's have models of to whom 
they are speaking. They shape 
their remarks according to how they 
feel they are going to be 
interpreted. They apparently 
monitor what they are going to say, 
often making mistakes and changing 
their minds as they are speaking. 
4. Speakers make effective use of the 
thematic organization of their 
conversations to direct 
interpretation, to specify role 
relationships, to qualify, 
intensify, and to amuse. 
5. Speaking has a tone to it. 
Sometimes it is angry, other times 
placating. It can also be erudite, 
reverent, direct, evasive, 
sardonic, etc. 
6. Speaking can also be stylized, 
having cultural and sub-cultural 
J 
I 
I 
! 
i 
I 
,|, 
i 
S 
i, 
t 
! 
i 
! 
I , 
I 
! 
il 
i 
constraints on what is said and how 
it is said; for example, Black 
narrative styles. 
7. Speakers make effective use of 
idioms, "buzz words", vernacular 
expressions, slang words to express 
themselves. 
Very likely most all these aspects of 
speaking appear as obvious, but to 
paraphrase the previous Wittgenstein 
quotation, it is often the obvious t~at 
escapes us. The argument can be made that 
yes, indeed, we will get to these things but 
let us first write a program that first 
speaks grammatically. I think that misses 
the point. Grammaticality is only one 
aspect of speaking behavior, and one which I 
doubt is as focal as it is made out to be. 
Grammar is just one of the means that 
speakers use to communicate their thoughts, 
intentions, and feelings to others. It 
should not be that element around which all 
subsequent investigations should be built. 
I feel that future work done with empirical 
data on speaking behavior will bear this 
point out and perhaps encourage the 
development of more comprehensive 
computational theories of speaking as an 
intentional and expressive human activity. 
If these theories are going to have a 
descriptive and explanatory value, they will 
at least have to be able to do these seven 
things that human speakers "obviously" do. 
125 
REFERENCES 
Chomsky, N., "Problems of Explanation in 
Linguistics," in Explanation in the 
Behavioral Sciences, Borger, R., 
Cioffi, F., eds., Cambridge University 
Press, p. 425-470, 1970. 
Dennett, D., Whv the haw of Effect Hill qot 
~o AwaY, forthcoming, 1975. 
Goffman, I., Frame Analysis, Harper and Row, 
1974. 
Hymes, D., Towards Communicative Competence, 
University of Pennsylvania Press, 1972. 
Labov, W., personal communication. 
Radnitsky, G., 
Metascience, 
Books, 1970. 
Contemporary Schools of 
Scandinavian University 
Toulmin, S., "Reasons and Causes" in 
Explaqation in t~e Behavioral Sciences, 
Borger, R., Cioffi, F., eds., Cambridge 
University Press, p. 1-26, 1970. 
Wittgenstein, L., Philosophica~ 
ti~D_~/~?d~, Blackwell, 1953. 
