SOME APPLICATIONS OF COMPUTERS IN LINGUISTICS*

Victor H. Yngve

Computers have already had a considerable impact on linguistics, and there is every
reason to believe that the impact will be far greater and far more important in the
future. The application of computer methods in linguistics has been along several
lines: There have been applications to traditional or usual linguistic methods.
There have been applications to extensions of traditional methods made possible
by the special abilities of the computer. These are both important, but most impor-
tant of all are applications to methods that are entirely new to linguistics and that
hold the exciting promise of new and deeper insights into language phenomena.
Computer applications to traditional or usual linguistic methods are the most
straightforward. These are methods that have been useful in an unautomated form
for many years. Concordance making, text searching, and the handling and sorting
of linguistic data lend themselves to easy automation. The use of the computer can
bring speed, convenience, accuracy, and relief from a certain amount of drudgery.
Computer applications that involve straightforward extensions of the older tech-
niques hold the promise of yielding results virtually unattainable using the older
techniques. This is because the superior speed, accuracy and clerical manipulating
ability of the computer bring a new dimension to the research capabilities of the
investigator. The kinds of operations envisaged in these extensions of older tech-
niques would be entirely impractical without the computer because of the large
amount of manual labor that would be entailed. The computer thus becomes an
instrument for increasing or extending the scope and usefulness of older techniques
into areas that had previously been effectively closed to investigation.
But some of the computer applications to linguistics are entirely new and are not
just straightforward applications or extensions of older non-computer methods. The
computer is opening up exciting new vistas in linguistic research. It offers oppor-
tunities for the exploration of virgin territory and the possibility of obtaining new
and deeper insights into language, its structure and its use by the human organism.
The full scope of the future possibilities is only dimly seen, but the results already
in tend to indicate that the future will be very bright indeed.
It is thus important to understand these new techniques and to develop them and
apply them systematically in linguistics. A representative and diverse selection of
* This work was supported in part by the National Science Foundation.

172
 Some Applications of Computers in Linguistics 173
1
applications is discussed here. We do not here present a complete survey: there are
other applications that are perhaps equally important. But the ones presented here
should serve to indicate the kind of an impact that computers are already having
in linguistic research. In this new area, progress may be limited only by the bounds
of our creative imagination.
File Processing
Many existing linguistic procedures involve the handling of large files of data or
large quantities of text. In each case the application of the computer brings certain
advantages. In the case of files of data, computer handling makes possible frequent
updating involving the interfiling of new material, and then, with each updating,
the whole file can be printed out in its new form. Thus the linguist always has avail-
able the whole current file arranged and printed in a convenient format. With the
file stored in a form amenable to computer operations, the possibility is open for
easy production of complex new arrangings and sortings, with printed copies in each
new arrangement. Or it is possible to make special searches of the file for particular
types of items answering to certain specified search criteria. The ease with which
specialized searches of the data can be carried out makes possible a considerable
flexibility in research, for new searches can be planned on the results obtained from
previous searches.
The advantages that automation brings to these rather straightforward file oper-
ations are several. First of all there is the advantage of accuracy in sorting and copy-
ing. The accuracy of the computer far exceeds that of manual sorting and copying.
Another advantage is the more flexible and convenient arranging and displaying
of the data without the necessity for extensive manual operations. Then there are
the speed and flexibility in research that are gained by having the data in a tractable
form where they can be easily searched, sorted, arranged and printed.
But perhaps one of the greatest advantages of the automation of file operations
is the possibility of introducing sophisticated error checks and controls based on
known regularities in the data entries. For example, in a large file of personal names
in the ancient Semitic Amorite language,2 it is presumed that the names will con-
form to a certain known internal morphological structure. It is thus possible to pro-
gram computer checks of the accuracy of the manual copying and transliteration
based on the presumed structure of the entries. Any names that do not conform to
the posited structure are automatically located and marked by the computer so
that they can be examined in detail to determine whether the deviation is significant
or whether it is due to an inadvertent error in copying or keypunching the data.
By such means the data can easily be maintained in a state of accuracy much higher
than is generally feasible without an inordinate amount of manual checking.
Text Handling
In the case of text handling, there are several operations that become quite easy
with the availability of text in machinable form. The first thing that comes to mind,
of course, is the preparation of concordances, a task that has been a very time-
174 Victor H. Yngve

consuming one for scholars in the past. With the use of the computer and existing
concordance programs, concordances can be obtained relatively easily.
Related to the production of concordances are a number of other operations that
can easily be automated. These include the extraction of vocabulary from text and
the counting of words, morphemes, or other items of interest. These operations and
other more sophisticated ones are being extensively applied in dictionary and glos-
sary making, as well as in areas of literary and textual criticism and stylistics.
With the availability of text in the computer, certain important extensions of
concordance techniques emerge. These involve the ability of the computer to carry
out searches of the text according to complex search criteria. The limitations of
the concordance that this possibility overcomes are twofold. In the first place,
the typical concordance arranges segments of the text alphabetically according
to each of the words of the text, so that the investigator can look up any word
and find all of its contexts brought together. But typically the investigator has
neither the interest nor the time to look up every word in the concordance. But also,
typically, he could not use a partial or selective concordance because he cannot fore-
see which words he is going to want to look up, because he cannot foresee the exact
course that his research will take. The ability to carry out text searches to order
overcomes this difficulty, for a number of searches can be carried out during the
course of the research.
The second difficulty of the traditional concordance is that, although it is easy
to find all of the contexts of a given word, it is difficult to use a concordance to help
find examples of more complex patterns, for example, sentences involving inversion,
sentences involving three or more clauses, sentences involving one of a number of
negative adverbs and a progressive verb form. Searches of text for patterns such
as these become possible with the newer computer techniques. Particularly impor-
tant in achieving the flexibility and ease of programming required for specifying
and carrying out such complex searches is the use of a convenient high-level pro-
gramming language. The flexibility of the clerical tasks that can be accomplished
by the computer to order, especially if a high-level programming language like
COMIT3 is employed, is so great that the effect of this approach is to give the in-
vestigator a much more powerful tool for searching his text than he could possibly
expect by using concordances.
Bringing Structure to Light
The use of the computer with its increased speed and flexibility opens up the
possibility of entirely new operations on text and on data that promise to add new
dimensions to Linguistic research. Some of the new possibilities involve the use of
statistical techniques for revealing suspected structures in a text. These methods
include some of the cryptanalytic techniques that have been used so successfully
in deciphering texts in unknown languages. But the possibilities are much broader
than appear to be realized, for the methods have not been generally explored or ex-
ploited to the extent that they deserve.
 Some Applications of Computers in Linguistics 175
The basic idea behind some of these methods is that there is an antithesis between
structure and randomness. Randomness is lack of structure: A completely unstruc-
tured sequence of characters would exhibit all the properties of randomness that
are known to mathematicians, but if the sequence of characters is structured in
any way, the randomness is in some measure destroyed. Therefore, if statistical
measurements are made on the text and deviations from randomness are found, the
deviations are to be attributed to the influence of structure. The technique thus in-
volves a search for ways in which the text deviates from randomness and requires
the finding of statistical measures that are sensitive to the constraints of interest
to the investigator.
Now at any given point in the course of research, a certain amount of the struc-
ture of the text is understood and describable, and the remainder of the structure
is not understood and is undescribed. The statistical techniques are to be applied
to the problem of discovering some of this unknown structure. It is thus necessary,
when setting up the statistical tests, to cancel out the known and measurable
effect of the known structure. This can sometimes be done in a rather straightfor-
ward manner; sometimes it requires considerable ingenuity.
An early experiment along these lines dealt with a scheme called gap analysis,4
which was pursued as a method of looking for syntactic constraints. The experi-
ment, using an English text of about ten thousand words, was aimed at exploring
the possibilities of the technique. It was assumed that the known and measurable
structure was the different frequencies of different words in the text. The statistical
structure of the text was then examined to look for deviations from randomness
arising from the constraints of syntax. If there were no constraints due to syntax,
the occurrence of a word in a text would have no effect on the possibilities or proba-
bilities of occurrence of other words in the vicinity. The experiment, however,
showed up strong deviations from randomness, and these could be used to posit
syntactic constraints that agreed with what is known about English syntax. The
method was thus shown to be capable of giving results.
Dialect Survey
There are many other possibilities opened up by computer techniques. Let us con-
sider the following possible application to dialectology that could assist in a large-
scale survey of the phonemic structure of related dialects. A modified informant
technique would be used that would not require the presence of a trained linguist.
This, together with the use of the computer for data reduction, would make possible
the gathering and processing of phonemic dialect information on an unprecedented
scale.
A linguist with some knowledge of the dialects in question would prepare a list
of words to be presented to the informants. For each word, a punched card would
be prepared which contained the word and a sentence exemplifying its use. The
word and sentence would appear printed along the top of the card. The cards would
be made up into decks and distributed to the informants together with carefully
176 Victor H. Yngve
worked out instructions. The informants would be instructed to sort the cards into
piles according to criteria designed to reveal phonemic similarity and difference.
The informant would then check over the piles that he has made and add to the top
of each pile one of the heading cards provided. The various piles of cards, separated
by their heading cards, would then be returned to the linguist who would have a
computer program ready for reading the cards and summarizing the information
implicit in the sorting done by the informants.
In order to make use of large amounts of data of this type, programs would have
to be developed for grouping and summarizing the data according to techniques of
set theory or statistics. These operations can all be applied automatically. The re-
sult would be a rather complete catalog of informant responses arranged according
to dialect similarity. Compact statements of similarities and differences would then
be made on the basis of the processed data, and these could be used as a basis for
more thorough and detailed investigations in the field. In fact, the availability of
summarized data from a large number of individuals would allow the linguist to
select the best small set of informants for further detailed phonetic and phonemic
investigations by personal interview. The technique would thus allow a choice of
informants that would reduce the over-all effort involved in obtaining the desired
dialect descriptions.
Testing of Linguistic Statements
No linguistic statement or description can be trusted unless it has been ade-
quately tested against the linguistic data that it is supposed to cover. The difficulties
of comparing a statement with data become very great as the complexity and
wealth of detail covered by the statement increase. In this area, computer tech-
niques are finding a very important application. The utility of computer methods
in testing is particularly evident in generative morphology and syntax. However,
this is not the only area where computer techniques are relevant. Paradigmatic de-
scriptions, historical statements and all other methods of linguistic description that
are sufficiently explicit and precise are amenable to computer testing.
Linguistic statements can be tested in two ways, by synthesis and by analysis.
The testing by synthesis may involve synthesizing all forms that are predicted, as
in the case of the author's statement of the inflection of the English regular and
irregular verbs.5 A program to test this statement synthesized and printed a com-
plete paradigm for each of the irregular verbs and for representative regular verbs.6
But if the linguistic statement involves more variability, as in the area of syntax,
it may be difficult or impossible to synthesize and print out each of the described
sentences. In fact, most syntactic statements generate an infinite set of sentences,
and it is in principle impossible to synthesize and print them all. A method of ran-
dom generation thus has much to recommend it.7-10 According to this scheme, sen-
tences conforming to the grammatical constraints expressed in the grammar are
synthesized at random. In the generating of a sentence, a random choice is made at
any point where there are alternative constructions that could fulfill a given func-
tion. The resulting generated sentences can then be examined and compared with
 Some Applications of Computers in Linguistics 177
observation. If the generated structures do not conform to the language, the lin-
guist knows that his statement is inaccurate in certain respects. He is thus enabled
to make the appropriate corrections.
In the testing of generative grammars by analysis, data in the form of texts are
fed into the computer and analyzed according to the grammar. The results of the
analysis are then compared with what the linguist expects, and any deviations again
lead to improvements in the linguistic statements.
The two methods of testing a statement, synthesis and analysis, are comple-
mentary in a certain sense. The method of synthesis tests to see that the structures
conforming to the linguistic statements are legitimate expressions in the language.
The method of analysis tests to see that expressions known to be in the language
are actually covered in the statements.11, 12
The fact that our statements are currently only a partial explanation of language
phenomena shows up clearly in such testing. In the case of synthesis of sentences
from a generative grammar, the sentences may be syntactically acceptable but non-
sense. Some of the difficulties of comparing generative grammars with observation
are concerned with judging the grammatical acceptability of nonsense. This diffi-
culty is a natural consequence of the fact that generative grammars do not deal
with the distinction between sense and nonsense. Perhaps it makes no sense to ask
whether nonsense is grammatical or not. In the case of the analysis of sentences by
computer, the fact that the linguistic statement is only a partial explanation shows
up in the extreme degree to which the parsings produced involve multiple syntactic
ambiguity.13
Models of Language Users
If a computer program can analyze and synthesize sentences according to a lin-
guistic statement, the program itself can be considered a theory in the sense that it
makes predictions. Thus as our knowledge advances of how to test grammars by
means of programs, we may find that it will be reasonable to make no distinction
between the program and the linguistic statement. This practice becomes quite
feasible with the use of high-level programming languages such as COMIT,3 by
means of which a computer program can be written in a way that is convenient for
the linguist to read and comprehend.
But the most exciting implication of computers to linguistics follows from the
fact that both man and computer are symbol manipulators or information proces-
sors. For this reason, a computer simulation of linguistic behavior stands a chance
of giving us much deeper insights into language phenomena than computer simula-
tion might provide in other areas, such as the simulation of traffic flow in a city
or of material flow in a manufacturing process. In other words, a computer program
may be a model of man in his role of symbol manipulator in a much deeper sense
than a computer program may be a model of other processes, because the computer
is also a symbol manipulator.
An example of the heuristic value for linguistics that a computer model of lan-
guage behavior can provide is to be found in the work on the relation of the tem-
178 Victor H. Yngve
porary memory to linguistic structure.14,15 In this work, a computer program was de-
vised to model a certain facet of human language behavior, namely the production
of grammatical sentences. This led to a more unified understanding of a wide di-
versity of previously unconnected facts of the structure of English, and led to a
unified view of syntax that promises to be extremely important in the understanding
of language typology and language change. On the basis of this work it has become
possible to comprehend perhaps the major reason for the complexity of languages.
It may be safe to say that we will only really understand human language be-
havior when we can make working models that also exhibit language behavior. The
emergence of the computer as a tool in linguistics puts at our disposal the very
techniques that we need for making such working models, and the prospect is ex-
tremely exciting.
NOTES
1. There are no adequate surveys, since the field is new and advancing very rapidly. The
reader may find additional material in the following references: Mechanical Translation and
Computational Linguistics, a quarterly published by the University of Chicago Press for the
Association for Machine Translation and Computational Linguistics, H. H. Josselson, Sec.,
Wayne State University, Detroit, Michigan; The Finite String, a newsletter published for
the above organization—each issue includes bibliographical information; H. P. Edmundson
(ed.), Proceedings of the National Symposium on Machine Translation (Englewood Cliffs,
N. J.: Prentice-Hall, 1961); P. L. Garvin (ed.), Natural Language and the Computer (New
York: McGraw-Hill, 1963); D. G. Hays (ed.), Readings in Automatic Language Processing
(New York: American Elsevier, 1966); S. M. Lamb, “The Digital Computer as an Aid in
Linguistics,” Language, XXXVII, No. 3 (1961), 382-412; W. P. Lehmann, “Zwischen zwei
Sprachen,” Graduate Journal (University of Texas), VII, No. 1 (1965), 111-131; A. G.
Oettinger, Automatic Language Translation (Cambridge, Mass.: Harvard University Press,
1960).
2. I. J. Gelb, “On the Morpheme ān in the Amorite Language,” above (p. 45).
3. V. H. Yngve, COMIT Programming (Cambridge, Mass.: M.I.T. Press) (in prep-
aration).
4. V. H. Yngve, Gap Analysis and Syntax (“IRE Transactions on Information Theory”),
Vol. IT-2, No. 3 (1956), pp. 106-112.
5. V. H. Yngve, “MT at M.I.T., 1965,” in A. D, Booth (ed.), Machine Translation (Amster-
dam: North-Holland Publishing Company, 1967).
6. A number of others have also used computer methods to test morphological statements.
For example, Erica Reiner and James McCawley, both of the University of Chicago, have
used the computer to test generative statements of Akkadian verb morphology and of
Finnish verb and noun morphology (private communications).
7. V. H. Yngve, “Random Generation of English Sentences,” 1961 International Confer-
ence on Machine Translation and Applied Language Analysis, I (London: Her Majesty’s
Stationery Office, 1962), 66-80.
8. A. C. Satterthwait, “Computational Research in Arabic,” Mechanical Translation, VII,
No. 2 (1963), 62-70.
9. D. A. Dinneen, “The Grammar of Specifiers,” in D. G. Hays (ed.), Readings in Auto-
matic Language Processing (New York: American Elsevier, 1966), chap. 8, pp. 127-136.
 Some Applications of Computers in Linguistics 179
10. G. H. Harman, “Generative Grammars without Transformation Rules, a Defense of
Phrase Structure,” Language, XXXIX, No. 4 (1963), 597-616.
11. The author and his students are continuing to develop a computer grammar of English
that can be used for sentence synthesis and analysis. The current scheme involves a phrase-
structure grammar with discontinuous constituents and subscripts.
12. Work is also proceeding on computer grammars within the transformational-generative
framework or some modification of it at the MITRE Corporation, Lexington, Massachusetts;
under Zellig Harris at the University of Pennsylvania, Philadelphia, Pennsylvania; and
elsewhere. Unfortunately computer methods are not being widely applied by transformational
grammarians, and many transformational descriptions remain untested. It seems to be little
realized that methods of computer-program debugging (error discovery and elimination) can
be applied profitably in linguistic research.
13. See the more recent literature in the field of mechanical translation.
14. V. H. Yngve, “A Model and an Hypothesis for Language Structure,” Proceedings of
the American Philosophical Society, CIV, No. 5 (1960), 444-466.
15. V. H. Yngve, “The Depth Hypothesis,” in Proceedings of Symposia in Applied Mathe-
matics (“Structure of Language and Its Mathematical Aspects,” Vol. XII [American Mathe-
matical Society, 1961]), pp. 130-138.