VERBAL CONTEXT AND THE RECALL OF

MEANINGFUL MATERIAL

By GEORGEA. MILLERand JENNIFER A. SELFRIDGE,HarvardUniversity

Communicativebehavior,perhapsmore than any of man's other activi-

ties, depends upon patterning for its significance and usefulness. An
accidentalinversion of words or letters or sounds can producegrotesque
alterationsof a sentence, and to scramblethe elements at random is to
turn a sensible message into gibberish. No attackupon the problems of
verbal behaviorwill be satisfactoryif it does not take quantitativeaccount
of the patternsof verbalelements.
We can dependablyproduce and distinguish only a small number of
differentletters or speech sounds. We must use these few elements to talk
about millions of different things and situations. To stretch these few
elementsto cover these many needs, we are forced to combinethe elements
into patterns and to assign a differentsignificanceto each pattern. Since
the numberof possible patternsincreasesexponentiallyas the length of the
pattern increases, this proves to be an efficient method of solving the
problem.
Not all the possible patternsof elements are used in' any particularlanguage.
In English, for example,the sequenceof lettersqke does not occur.It is reasonable
to ask, therefore,why we do not exploit the availablepatterns more effectively.
Is it not inefficientto ignore some patternswhile others are greatly overworked?
The preferencefor some patternsat the expense of others forces us to produce
more elements-letters, sounds, words, etc.-in order to make the same number
of distinctionsthat we could makewith the same elementsif we used all possible
patterns.To illustrate:imaginea languagewith 10 elementarysymbolsthat is used
to refer to 100 differentthings, events or situations.If we used all possible pairs
of elements,we could refer to every one of the 100 things with one of the 100
pairs of 10 symbols.If, however,we refuse to use some of the pairs and so rule
them out of the language,it is necessaryto makeup the differenceby using triads.
Thus the languageuses patternsof three elementsto make distinctionsthat could
be madewith patternsof two elements.
On further consideration,however, this kind of inefficiencydoes not appear a
complete waste of time. By favoring some patterns rather than others the language
is protected against error.1 More specifically, in Engish we recognize immediately

* Accepted for publication August 30, 1949.

1C. E. Shannon, A mathematical theory of communication, Bell Syst. Tech. J.,
27, 1948, 379-423, 623-656.
176
 VERBAL CONTEXT AND RECALL 177

thatan errorhas occurredif we readin our newspaper,"Manbites dxg." The pattern

dxg is not admittedin English, and so we catch the error.If, however,all patterns
of elements were admissable,dxg would have some semanticrule and we would
not be able to catchthe mistake.The numbersystemis an exampleof the efficient
use of ten symbols,but it is highly susceptibleto mechanicalerrors.If a man says
that his telephonenumberis 9236 we have no way of recognizingthat he has or has
not madean error.
Patternsare unavoidable,and a preferencefor some patternsprovidesinsurance
against errors. Thus it seems reasonablethat the statistical studies of different
languages all show that some patternsof elements are greatly overworkedwhile
others occur rarely or not at all.2 The present interest in verbal patterning,how-
ever, is in the light these observationscan throw upon the psychologicalproblem
of verbal context.
Verbal context. Psychologists use the word context to refer to the totality
of conditions influencing a behavioral event. For the present discussion
we want to restrict this broad definition and to consider only the antecedent
verbal conditions. When a man talks, his choice of words depends upon
his training, his needs and intentions, the situation and audience. These
factors comprise the total context in which his words must be studied. By
verbal context, as opposed to total context, we mean only the extent to
which the prior occurrence of certain verbal elements influences the talker's
present choice. If the talker has said "children like to," his choice for the
next word in this pattern is considerably limited-elephant, punished,
loud, Bill, and many other words are highly unlikely continuations.
By verbal context, therefore, we mean the extent to which the choice
of a particular word depends upon the words that precede it. In the
statistical' sense, this definition of verbal context is given in terms of de-
pendent probabilities.3 The probability that event C will occur is not the
same after A as it is after B. The statistical dependencies between successive
units form the basis for a study of verbal context.
To illustrate the operationof conditionalprobabilitiesin our verbal behavior,
considerthe set of all possible sequences10 letterslong. We could constructa table
listing them. The first row of the table would be the patternaaaaaaaaaa,10 con-
secutive a's. The second would be aaaaaaaaab, then aaaaaaaaba, aaaaaaaabb, aaaaa-
aabaa, etc., until all possible arrangementsof letters, spaces, commas, periods,
hyphens, quotes, colons, numbers,etc., were exhausted.Altogetherthere would be
about 50 different symbols, and the table would contain 5010,or about 100,0,00000,-
000,000,000 differentpatterns.Then we would examinesomeEnglishwritingand try
to determinethe relative frequenciesof occurrenceof the patterns.Only a small
fraction of the 5010alternativesactually occur in English. The table would show
2G. K. Zipf, HumanBehaviorand the Principleof LeastEffort,1949, 1-343.
aG. A. Miller and F. C. Frick,Statisticalbehavioristicsand sequencesof responses,
Psychol. Rev., 56, 1949, 311-324.
178 MILLER AND SELFRIDGE

strong dependencies. For example, the letter q is always followed in English by

the letter u, and so all those entries in the table that contained a q followed by
anything but u would not occur in English. It is not possible to predict the relative
frequency of qe, for instance, by multiplying the relative frequencies of q and of e.
If such a table existed, along with the relative frequencies of occurrence, it
would be possible to construct sequences of letters that reflected the statistical
dependencies of English verbal patterns. We can imagine similar tables constructed
for shorter or longer sequences of letters. A table for all patterns of 2 symbols
would represent the relative frequencies of pairs; for 3 symbols, triads, etc. The
longer the sequence, the more information the table contains about the pattern of
dependencies in our molar verbalizations.
To illustrate the use of such information we shall borrow a device used by
Shannon. Suppose we have no knowledge at all of the relative frequencies of oc-
currence, but only a list of the 50 different symbols. Then, for all we know, any
sequence of symbols might be permissible. If we tried to construct a message in
the language, the best we could do would be to draw at random from the 50
different symbols. We have no reason to think that one sequence of symbols is
more likely than another. Proceeding in ignorance to construct a message, we might
produce something like this: cplp'rzu(p".:k!)"ntegznqO?i6vlaur .8h, etc.
Now suppose that we have a reliable tabulation of the relative frequencies of
'patterns' of one symbol. We know, therefore, that e and the space between words
are more likely to occur than are ? and z. With this information we can increase
the chance of constructing a meaningful message, although our chances are still
very small. If we draw successive symbols according to their relative frequencies of
occurrence in English, we might produce something like this: wli hnrooye Iricocnri
mae c zg 2eaya, etc.
The next step is to imagine that we have a tabulation of relative frequencies of
occurrence of patterns of two symbols. Now it is possible to improve the statistical
approximation to English by drawing in the following way. Begin by drawing any
likely pair. Suppose the pair is au. Now look at all the pairs starting with u and
draw from them according to their relative frequencies of occurrence. Suppose the
result is ud. Now look at all pairs starting with d and draw one of those, and so
proceed to build up the message. Notice that each draw depends upon the preceding
draw-the preceding draw determines from which set the present draw is to be
made. Drawing in this way reflects the conditional probabilities of successive
symbols. A message constructed in this way might read aud ren stiofivo omerk.
thed thes bllale, etc.
If we have a tabulation of sequences of three letters, we can construct a message
that reflects the conditional probabilities of English triads. First we draw a likely
triplet, say ann, then draw next from the triplets starting with nn and obtain nna,
then from the triplets beginning na, etc. The preceding two symbols determine from
which set the next triplet is drawn. In this way a message might be produced that
would read: annation ef to the acticas. Oth rested, etc.
With a tabulation of sequences of four letters we might produce: influst
intradio be decay, the condive, etc. By tabulating the relative frequencies of longer
sequences and drawing successive items so as to reflect these frequencies, we can
construct messages that reflect the statistical dependencies of English as extensively
as we please.
 VERBAL CONTEXT AND RECALL 179

For convenience, we shall refer to these different ways of constructing a sta-

tistical English as orders of approximation, and shall number them from 0 to n.
At the zero order we have no knowledge of relative frequencies, at the first order
we know the relative frequencies of individual symbols, at the second order we
know the relative frequencies of pairs, at the nth order we know the relative
frequencies of n.
Consider this statistical English now in terms of verbal context. With a zero-
order approximation to English there are no contextual influences whatsoever on
the choice of successive symbols. At the nth-order approximation, however, each
symbol is selected in the context of the preceding n-1 symbols. As the order of
approximation is increased, the amount of context for each symbol is increased, and
the contextual constraints (dependent probabilities) have a chance to operate. As
the order of approximation is increased, the messages we can construct become
more and more familiar, reasonable, meaningful. The more we permit contextual
restraints to operate, the better are our chances of producing a message that might
actually occur in English.
We have, therefore, a scale for what can be loosely called 'meaningfulness.' At
one end are the random jumbles of symbols we customarily call nonsense, and at
the other end are patterns of symbols that could easily appear in our daily dis-
course. Equipped with this quantitative estimate of 'the degree of nonsense' or
'amount of contextual constraint,' we can proceed to study certain psychological
problems that have been phrased in terms of meaningfulness.
An experimental illustration. Briefly stated, the problem to which this
concept of verbal context has been applied is, How well can people remem-
ber sequences of symbols that have various degrees of contextual constraint
in their composition ? The experimental literature contains considerable evi-
dence to support the reasonable belief that nonsense is harder to remember
than sense. This evidence has suffered, however, from a necessarily sub-
jective interpretation of what was sensible.
In the present experiment, the learning materials were constructed at
several orders of approximation to English. These materials were presented
to Ss whose recall scores were then plotted as a function of the order of
approximation.4
Learning materials. In the preceding examples we have used patterns of letters
to illustrate the effects of contextualconstraints.There is, of course, no necessity
to limit the argumentto letters. It is possible to use words or even sentencesas
the componentelements that are arrangedaccordingto the statisticalstructureof
English. In the present experimentalillustration the materialswere constructed
with words as the units of analysis.
In theory, the constructionof materialsto incorporatethe statistical structure

4The
experimentwas carriedout by the junior authorand is presentedin detail
in her honor'sthesis, Investigationsinto the Structureof VerbalContext,1949. The
thesis is on file in the Library of the Psychological Laboratory, Memorial Hall.
Harvard University.
180 MILLER AND SELFRIDGE

of English over sequences of several words requires a tabulation of the relative

frequencies of such sequences. Such a tabulation would be exceedingly long and
tedious to compile. An alternative method of construction is available, however,
which makes the procedure practicable. Instead of drawing each successive word
from a different statistical distribution indicated by the preceding words, we draw
the word from a different person who has seen the preceding words.
At the second order, for example, a common word, such as he, it or the, is
presented to a person who is instructed to use the word in a sentence. The word
he uses directly after the one given him is then noted and later presented to another
person who has not heard the sentence given by the first person, and he, in turn,
is asked to use that word in a sentence. The word he uses directly after the one
given him is then noted and later given to yet another person. This procedure is
repeated until the total sequence of words is of the desired length. Each successive
pair of words could go together in a sentence. Each word is determined in the
context of only one preceding word.
For higher orders of approximation the person would see a sequence of words
and would use the sequence in a sentence. Then the word he used directly after
the sequence would be added, the first word of the sequence would be dropped,
and the new (but overlapping) sequence would be presented to the next person.
By this procedure we constructed sequences of words at the second, third, fourth,
fifth and seventh orders of approximation.
For the first order approximation to English a scrambling of the words in the
higher orders was used. By drawing words at random from the contextually deter-
mined lists, we obtained as good an approximation to the relative frequencies of
individual words in English as these higher order lists provided. The alternative
method of selecting words at random from a newspaper might have given a sample
quite different in difficulty (familiarity).
A zero order approximation to English could be obtained by drawing at random
from a dictionary. Most dictionaries contain too many rare words, however, so we
drew from the 30,000 commonest words listed by Thorndike and Lorge.5 This
source had the additional advantage that it listed separately all forms of the word,
whereas the dictionary lists only the lexical units. Words drawn at random from
this list of 30,000 words are selected independently and without any constraints due
to adjacent words or the relative frequencies of appearance of the words in English.
A final set of words was taken directly from current fiction or biography. These
lists represent a full contextual determination.
By these devices we constructed sequences of words with eight different degrees
of contextual constraint. In the following discussion we shall refer to these lists as
0, 1, 2, 3, 4, 5, 7 and text-orders of approximation. At each order four lists of
different length-10, 20, 30 and 50 words-were constructed. Thus the experi-
mental design called for 32 different lists. Two such sets of 32 lists were con-
structed. Since the lists require considerable time to compile and since they may be
some general interest, one of the sets of 32 is reproduced in full in the appendix to
this paper.
Experimental procedure. Each set of 32 lists was read aloud and recorded on a
wire recorder. A man's voice was used. The words were read slowly and distinctly

5E. L. Thorndike and I. Lorge, The Teacher's Wordbook of 30,000 Words, 1944.
 VERBAL CONTEXT AND RECALL 181

in a near monotone,with a short pause between words. At the beginning of the

recordingthe instructionswere given and a single practicelist was presentedto
make sure the Ss understoodtheir task. They were to listen until a list was
finished,at which time a bell soundedsignalling them to begin writing what they
had just heard. The Ss were instructedto write the words they rememberedas
nearly in their correct order as possible. Order was not used, however, as a
criterionfor scoringtheir responses.All eight of the 10-wordlists were given first,
proceedingfrom least to greatest contextualdetermination,then the 20-word lists
in the same order, then the 30-wordlists, and finally the 50-wordlists. Short rest

Ll00 I I I _&__o_ I / =

Q:80
so - f / o
vu
% ./ p

40
4:
'>
/ d 50-word passage

20
Lu

t I I i ?I r 1 1 I I
O 2 3 4 5 6 7 TEXT
ORDEROF APPROXIMATION
TO STATISTICAL STRUCTURE OF ENGLISH
FIG. 1. Percentageof Words of the Lists of Different Lengthsthat were Correctly
Recalled at the Various Orders of Approximationto the StatisticalStructureof
English.
periods (5 min.) were given between the 20- and 30-word lists and betweenthe
30- and 50-wordlists.
Two groups of 10 Ss were used. One group heard and recalledone of the sets
of 32 lists, the other group heard and recalled the second set. The Ss were prin-
cipally studentsat Harvardand Radcliffe.It was E's impressionthat a largernum-
ber of Ss would not have reducedthe irregularitiesin the results, for most of the
variability seemed attributableto sampling peculiarities in the lists themselves.
Several more sets of 32 lists would be needed before an accurateestimateof the
functionalrelationscould be made. The results are adequate,however,to indicate
the approximatemagnitudesand general trends of the functions.
Ss' answerswere scoredfor the numberof words that they had written that had
occurredin the test material.The number recalled, regardlessof order, was ex-
pressedas a percentageof the total numberpresented.
Results and discussion.The experimentaldata are summarizedin Figs.
1 and 2. In Fig. 1 the recall-score,expressedas a percentage,is plotted as
182 MILLER AND SELFRIDGE

a function of the order of approximationto the statisticalstructureof

English, with the length of the lists as the parameter.In Fig. 2 the same
data are replottedto show the relationof the recall-scoreto the length of
the list, with the order of approximationas the parameter.In both figures
the functions representthe mean scores for all 20 Ss. It is clear from the
results that percentagerecalled increasesas the order of approximationis
increasedand decreasesas the length of the list is increased.Inspectionof
Fig. 1 leads to a reasonablesuspicion that the two variables,length and

e too i i l I

60 -

40 -

U 20-
20- ppro i =0
Orderof approxim.tion

t O '; ' I I J
0
1O 10 2 30 40 50 60
LENGTHOF PASSAGE
FIG. 2. Percentageof Words of the Various Ordersof Approximationto the Sta-
tistical Structureof Englishthat were CorrectlyRecalledat the DifferentLengthsof
PassagesLearned.
order of approximation,interact. With the short, 10-word lists there is
little to be gained from contextual bonds extending over more than two
words. With the 20-wordlists the Ss rememberedas well at the third order
of approximationas they did for the textual material.With the 50-word
lists, however, only orders 5 and 7 are comparableto the textual material
in termsof percentagerecalled.It would seem,therefore,that the longer the
passage the greater is the usefulness of contextual associations extending
over long sequences of items.
By a strict interpretation of the word 'nonsense,' one is forced to con-
clude that all orders of approximation less than the full text are nonsense.
Consider an example from Order 5:
house to ask for is to earn our living by working towardsa goal for his team in
 VERBAL CONTEXT AND RECALL 183

old New-York was a wonderful place wasn't it even pleasant to talk about and
laugh hard when he tells lies he should not tell me the reason why you are is
evident.
The experimental results show that this kind of gibberish is as easily
recalled as a passage lifted from a novel. Thus there are kinds of non-
sense that are as easy to recall as are meaningful passages.The significant
distinctionis not to be drawnbetweenmeaning and nonsense,but between
materials that utilize previous learning and permit positive transfer and
materialsthat do not. If the nonsensepreservesthe short rangeassociations
of the English language that are so familiar to us, the nonsenseis easy to
learn.
The experimentshows, therefore,that the problemof meaningvs. non-
sense in verbal learning need not be approachedin terms of a qualitative
dichotomy,but can be studiedas a functionalrelationbetweenquantitative
variables.The resultsindicatethat meaningfulmaterialis easy to learn, not
because it is meaningful per se, but becauseit preservesthe short range
associationsthat are familiar to the Ss. Nonsense materialsthat retainthese
short rangeassociationsare also easy to learn. By shifting the problemfrom
'meariing'to 'degree of contextualconstraint'the whole area is reopened
to experimentalinvestigation.
Psychologistsfamiliar with the problemsof verbal learning will recog-
nize the usefulness of the kind of material employed in this illustrative
experiment.For example, is retroactiveinhibition affectedby interpolating
differentordersof approximationto English betweenthe original learning
and the recall? What is the effect of using original and interpolatedma-
terials of the same or of different orders of approximationto English?
Do the higher approximationsto English show the same differencesbe-
tween recall after sleep and recall after waking activity that the lower
approximationsshow? Is it possible to show a continuumfrom the short-
term reminiscencethat can be demonstratedwith syllablesto the long-term
reminiscencethat can be shown with poetry? How does the span of im-
mediate memoryvary with the order of approximation?Is the superiority
of distributedover massedpracticea function of the order of approxima-
tion of the materialsto the statisticalstructureof English? Can differences
in learningand recallingdifferentordersof approximationbe demonstrated
as a functionof age?
The operational analysis of meaningfulness makes it possible to ask
such questions and to see how one would proceed to answer them. The
problemnow is to collectthe experimentaldata.
184 MILLER AND SELFRIDGE

SUMMARY

A quantitativedefinitionfor verbalcontext is given in termsof depend-

ent probabilities.The definition is used to constructlists of words with
varyingdegreesof contextualdetermination.When short range contextual
dependenciesare preservedin nonsensematerial,the nonsenseis as readily
recalled as is meaningful material.From this result it is argued that con-
textual dependencies extending over five or six words permit positive
transfer,and that it is these familiar dependencies,ratherthan the mean-
ing per se, that facilitatelearning.

APPENDIX
LISTS USED IN RECALL EXPERIMENT
O-order approximation
10: byway consequence handsomely financier bent flux cavalry swiftness weather-
beaten extent
20: betwixt trumpeter pebbly complication vigorous tipple careen obscure attrac-
tive consequence expedition pane unpunished prominence chest sweetly basin
awoke photographer ungrateful
30: crane therewith egg journey applied crept burnish pound precipice king eat
sinister descend cab Idaho baron alcohol inequality Illinois benefactor forget
lethargy fluted watchtower attendance obeisance cordiality dip prolong
bedraggle
50: hammer neatly unearned ill-treat earldom turkey that valve outpost broaden
isolation solemnity lurk far-sighted Britain latitude task pub excessively chafe
competence doubtless tether backward query exponent prose resourcefulness
intermittently auburn Hawaii unhabit topsail nestle raisin liner communist
Canada debauchery engulf appraise mirage loop referendum dowager abso-
lutely towering aqueous lunatic problem
1-order approximation
10: abilities with that beside I for waltz you the sewing
20: tea realizing most so the together home and for were wanted to concert I
posted he her it the walked
30: house reins women brought screaming especially much was said cake love
that school to a they in is the home think with are his before want square
of the wants
50: especially is eat objections are covering seemed the family I that substance
dinner raining into black the see for will passionately and so I after is
window to down hold to boy appearance think with again room the beat go
in there beside some is was after women dinner chorus
2-order approximation
10: was he went to the newspaper is in deep and
20: sun was nice dormitory is I like chocolate cake but I think that book is he
wants to school there
30: the book was going home life is on the wall of you are ready to the waltz
is I know much ado about it was a dog when it was
50: you come through my appetite is that game since he lives in school is jumping
and wanted help call him well and substance was a piano is a mistake on
this is warm glow in and girl went to write four turtledoves in my book is
fine appearance of the
 VERBAL CONTEXT AND RECALL 185

3-order approximation
10: tall and thin boy is a biped is the beat
20: family was large dark animal came roaringdown the middle of my friends
love books passionatelyevery kiss is fine
30: happenedto see Europeagain is that trip to the end is cominghere tomorrow
after the packagesarrivedyesterdaybroughtgood cheer at Christmasit is
rainingoutsideas
50: came from the beginningand end this here is the top spins in a house by the
libraryis full of happinessand love is very nice of her that fell from the
window she went home from work to pass the cigarettesdown to earth he
pickedan apple
4-order approximation
10: saw the football game will end at midnighton January
20: went to the movies with a man I used to go toward HarvardSquare in
Cambridgeis mad fun for
30: the first list was posted on the bulletin he broughthome a turkeywill die
on my rug is deep with snow and sleet are destructiveand playful students
always
50: the next room to mine silver in Pennsylvaniais late in getting home on time
my date was tremendousfun going there skiing this day would end and
have no more objectionsto his speech on the radio last night played the
viola in the orchestraand chorus performedthe
5-order approximation
10: they saw the play Saturdayand sat down beside him
20: road in the countrywas insane especiallyin drearyrooms where they have
some books to buy for studyingGreek
30: go it will be pleasantto you when I am near the table in the dining room
was crowdedwith people it crashedinto were screamingthat they had been
50: house to ask for is to earn our living by working towardsa goal for his
team in old New-York was a wonderful place wasn't it even pleasant to
talk about and laugh hard when he tells lies he should not tell me the
reason why you are is evident
7-order approximation
10: recognizeher abilities in music after he scolded him before
20: easy if you know how to crochetyou can make a simple scarf if they knew
the color that it
30: won't do for the memberswhat they most wantedin the coursean interesting
professorgave I went to at one o'clockstoppedat his front door and rangthe
50: then go ahead and do it if possible while I make an appointmentI want to
skip very much aroundthe tree and backhome again to eat dinnerafter the
movie early so that we could get lunch becausewe liked her method for
sewing blouses and skirts is
Text
10: the history of Californiais largelythat of a railroad
20: more attentionhas been paid to diet but mostly in relationto diseaseand to
the growth of young children
30: Archimedeswas a lonely sort of eagle as a young man he had studied for
a short time at AlexandriaEgypt where he made a life-long friend a gifted
mathematician
50: the old professor'sseventiethbirthdaywas made a great occasionfor public
honors and a gathering of his disciples and former pupils from all over
Europe thereafterhe lecturedpublicly less and less often and for ten years
receiveda few of his studentsat his house near the university