The Representation of Derivable Information in Memory: 
When What Might Have Been Left Unsaid Is Said 
Rand J. Spiro, Joseph Esposito, and Richard J. Vondruska 
Center for the Study of Reading 
University of Illinois at Urbana-Champaign 
It is now widely accepted that natural lan- 
guage comprehension is a constructive process. 
Information in discourse interacts with a variety 
of impinging contextual factors (including, most 
prominently, the comprehender's pre-existing 
knowledge) in an active, creative process that 
results in understandings not derivable by any 
solely linguistic or logical analysis (c.f., 
Bransford & McCarrell, 1975; Spiro, 1977, in 
press). Acceptance of the constructive view of 
comprehension entails a concomitant delimitation 
of the range of possible theories of mental 
representation. Knowledge structures must possess 
some capability for detecting the pragmatic, as 
well as logical, implications of the incomplete 
data contained in discourse (c.f., Charniak, 
1974; Mlnsky, 1975; Rumelhart & Ortony, 1977; 
Schank & Abelson, 1977). In other words, know- 
ledge structures must contain considerable 
information about the way the world usually works. 
This characteristic of representation is useful 
and efficient because natural and social contexts 
do produce sufficient constraints on worldly 
events and ideas as to make them, to a limited 
extent, orderly and predictable. 
However,a point often overlooked is that 
these same knowledge structures, with their 
information about the world's orderliness, may 
allow for more efficient processing and memorial 
representation of explicit information in dis- 
course, in addition to their role in deriving 
implicit information. This paper will be con- 
cerned with the psychological processing of 
(imperfectly) Predictable or derivable informa- 
tion that is nevertheless explicit in discourse. 
Predictable Information in Discourse 
Despite the fact that most research on 
inferential processes in comprehension has been 
concerned with generation of implicit informa- 
tion, much inferentially related information 
is embodied explicitly in discourse. We are 
referring here primarily to pragmatic inferences, 
i.e., implications that are usually but not 
necessarily true. Language is infrequently 
characterized by absolute redundancy; semantic 
content is rarely "repeated," except for special 
purposes such as emphasis. However, pragmatic 
inferences are only imperfectly predictable. If 
you read that a karate champion hit a block, 
uncertainty is reduced by also reading that the 
block broke, despite the fact that that outcome 
is usually to be expected. Similarly, it would 
not be considered unusual when relating the 
events at a birthday party to mention that there 
was a cake with candles blown out by the 
celebrant. Many things go in stereotyped ways 
but require explicit mention because the 
stereotype does not describe all possible cases. 
Throughout this paper, "predictable" is used 
as a shorthand for "imperfectly predictable, or 
characterized by significantly less than per- 
fect uncertainty." 
How is explicit but predictable information 
processed? As was mentioned above, attention 
has been primarily devoted to the processing 
of implicit predictable information, leaving 
little guidance on the present issue. However, 
in a variety of theoretical orientations, there 
is a common implication about how predictable 
information would be dealt with: simply put, 
explicit information, whether predictable or 
not, receives sufficient processing to be 
encoded in long-term memory. For example, 
Kintsch (1974) assumes "that subjects process 
and store \[an inference\] whether or not it is 
presented explicitly" (p. 154). It is difficult 
to imagine discourse representation theorists, 
who argue for the explicit representation in 
memory of implicit inferences (e.g., 
Frederiksen, 1975; Meyer, 1974), arguing that 
explicit inferences are not represented. In 
schema theories (e.g., Rumelhart & Ortony, 
1977), explicit discourse information is used 
to bind schema variables, again suggesting 
that predictable information would receive 
explicit mental representation. If anything, 
one would expect existing theories to predict 
that explicit inferences would receive a 
stronger memorial representation than un- 
predictable information, given their greater 
contextual support. For example, in their 
associative network model, HAM, Anderson and 
Bower (1973) argued that the greater the number 
of interconnections between information, the 
greater the likelihood that information within 
the interconnected network would be recalled. 
This view will be referred to as the "storage 
of explicit inferences" (SEI) hypothesis. 
An alternative hypothesis is that predictable 
information, however central to a discourse, 
is taken for granted, processed only super- 
ficially and receives an attenuated cognitive 
representation or no enduring representation 
226 
at all. f needed subsequently, it can be de- 
rived. This view will be referred to as the 
"superficial processing of explicit inferences" 
(SPEI) hypothesis. Processing explicit inferences 
in such a manner has the advantage of a cognitive 
economy of representation (besides a likely reduc- 
tion in processing time). Most information that 
is acquired will never be used again. It would 
then seem to be more efficient to devote extra 
processing effort to the occasions when the 
information is needed (i.e., by deriving it when 
remembering) rather than exerting effort toward 
stable encoding at the time of comprehension. 
Experiments on the Representation of 
Explicit Inferences 
There are considerable problems in designing 
an empirical test of the hypothesis that explicit 
pragmatic inferences in discourse are not repre- 
sented in long-term memory. If one merely tests 
memory for the inference, failure to remember 
could be attributed to not storing the informa- 
tion or to storing and then forgetting it; if the 
inference is remembered, it could be because it 
was stored and then retrieved, or it may have 
been generated at the time of test without having 
been stored. 
Spiro and Esposito (1977) developed a para- 
digm not subject to the ambiguities of interpre- 
tation of the more simple design discussed above. 
The primary manipulation of interest involved 
subsequently vitiating the force of an earlier 
explicit inference. If the inference is not 
stored, certain predictable errors in recalling 
it should be made. 
In the first experiment, subjects were pre- 
sented stories which contained information A, B, 
and C such that B was strongly implied by A 
except in the presence of C. For example, the 
A, B, and C elements in one story (about a demon- 
stration by a karate champion) could be para- 
phrased as follows: 
A: The karate champion hit the block. 
B: The block broke. 
C: He had had a fight with his wife 
earlier. It was impairing his 
concentration. 
C was either presented prior to A and 8 (C-Before), 
after A and B (C-After), or not at all (No-C). 
When C was not included in the story, if SPEI 
is correct, the B element should be taken for 
granted, processed only superficially, and not 
stably represented. It would be derivable if 
needed. However, if C is presented after A and 
B, memory for B should be impaired since B was 
not stored and C will block its derivation from 
A at the time of test. On the other hand, if C 
occurs in the text prior to A and B, then B is 
not strongly implied by A. B cannot be taken 
for granted with the assumption that it can be 
generated later if needed. Here B should be 
stably represented and memory for B should not 
be impaired. 
However, if SEI is correct, memory for B 
should not be affected by whether C is before or 
after A and B, since B is stored whether it is 
implied by A (C-After) or not implied by A 
(C-Before). Two objections to this argument 
can be made. The information might be stored, 
but remembering C might lead to a decision 
that the memory for B must be mistaken (a 
kind of output interference). However, C is 
present whether it occurs before or after A and 
B, so such an explanation would not account 
for differential effects of C-placement. The 
other possibility is that B is represented in 
C-After, but the representation is altered 
or corrected when the C information is encoun- 
tered. This possibility was investigated in 
the second experiment. 
In the first experiment, the following 
predictions of the SPEI hypothesis were tested. 
More errors in response to questions about the 
presented predictable information (B) should 
be made in the C-After than in the C-Before 
conditions. Errors can be erroneous judgments 
that nothing about the implied information 
was presented, called B-Mention errors (e.g., 
the story did not mention whether the block 
was broken), or, when the subject believes 
that something about B was mentioned, re- 
membering incorrectly what was specifically 
said in the direction of conforming with the 
C information, called B-Incorrect errors 
(e.g., it said in the story that the block 
did not break when he'hit it). Confidence in 
errors of the latter kind were also analyzed. 
If subjects are as confident about these 
errors as they are about their accurate 
responses, it would be even more difficult to 
maintain the hypothesis that the explicit in- 
ferences were represented. 
In the No-C condition, B-Mention errors may 
occur since B would not be represented according 
to the SPEI hypothesis. The more important 
prediction regarding the No-C condition is 
that B-Incorrect errors should not occur more 
often than in the C-Before condition. Other- 
wise, the differences between C-Before and 
C-After might be attributable to heightened 
accuracy due to greater salience of the implied 
information in the former condition rather 
than greater inaccuracy due to a failure to 
store the implied information in the latter 
condition. 
College subjects read eight target 
vignettes each containing A and B information, 
and C information included or not and placed 
as a function of which of the three conditions 
subjects were randomly assigned to. C informa- 
tion was always on a separate page from the A 
and B information, and subjects were instructed 
to not look back after reading a page. After 
reading all the vignettes, the subjects were 
tested for their memory for the vignettes. 
Of particular interest were the two types of 
questions, mentioned above, concerning the B 
information (remember, B was always explicit 
in the stories). 
The results supported the hypothesis that 
pragmatic inferences presented in text are 
superficially processed and do not receive a 
227 
stable and enduring representation in memory. In 
the C-After condition, subjects tended either to 
report that the inference was not presented in the 
text or that the opposite of the inference was 
presented. Furthermore, confidence in these 
errors was as high as confidence in correct mem- 
ories. It is difficult to retain the notion that 
inferences are deeply processed and stably encoded 
when the C-After manipulation can produce errors 
like remembering the block was not broken when 
the karate champion hit it. The results cannot 
be attributed to interference produced by the 
inference-vitiating C information at output, 
since the C-Before subjects would also be subject 
to such interference. Neither can the results be 
attributed to differential availability of C at 
output, perhaps due to primacy/recency effects 
related to the position of C in the text, since 
the information was almost always recalled. Also, 
unimportance of the B informtion is not a viable 
alternative since B tended to be central to the 
story (e.g., in a story about a karate champion's 
performance, information about his success in the 
demonstration is certainly important). 
One alternative interpretation that remains 
is that subjects do deeply process and stably 
encode the presented inference, but "correct" 
their representation when the inference-vitiating 
information is presented. If subjects are storing 
B and then changing or correcting it at the time 
C is presented, errors on B'should occur in the 
C-After condition no matter how soon the test is 
administered after reading. However, if the SPEI 
hypothesis is correct, when delay intervals are 
brief enough some surface memory for the super- 
ficially processed B information may remain, 
reducing the number of B errors. Accordingly, 
in the second experiment subjects were tested 
either immediately after reading each story 
(Interspersed Questions condition) or, as in the 
first experiment, after the entire set of stories 
had been read (Questions-After condition). Again, 
the C-Before and C-After manipulations were 
employed. 
The results of the second experiment repli- 
cated those of the first one in the Questions- 
After condition. Furthermore, the C-after 
effect was largely absent in the Interspersed 
Questions condition, demonstrating that the 
effect is not due to storing and then changing 
the representation of the B information (the 
explicit inference). 
Related Issues 
The discussion of implications of the super- 
ficial processing effect will at times be limited 
to reading rather than listening. Most of the 
following is of a speculative nature. 
Representation and Underlying Mechanisms 
Assuming some compatible representation 
system, what characterizes the processes that 
produce the superficial processing effect? At 
this time, only speculations about alternative 
possibilities can be offered. There are three 
potentially beneficial aspects of superficial 
processing of explicit predictable information: 
cognitive economy (the information need not be 
specifically stored in long-term memory), speed 
of processing (you can process and understand 
such information rapidly), and automaticity of 
processing (less conscious effort and working 
memory space are required). 
Two simple, preliminary accounts of the 
first factor, cognitive economy, can be offered. 
The superficial processing phenomenon appears 
most compatible with a schema-theoretic mode 
of representation. Perhaps variable bindings 
that are default (or at least high probability) 
values are not explicitly instantiated when 
they are explicit in discourse (but see the 
discussion of Determinants of Performance 
Variability below). However, one should not 
be overly persuaded by the simplicity of such 
an account. Other types of representation 
systems could also account for the phenomenon. 
For example, a spreading activation model 
(e.g., Collins & Loftus, 1975) might predict 
that explicit information is not tagged in 
memory when it has been recently activated with 
some greater than criterion strength. This 
issue will receive further discussion in the 
next section. 
Regarding speed of processing, several 
possibilities may be offered: the information 
is actual\]y predicted, perhaps followed by a 
selective scanning for partial clues of con- 
firmation (e.g., the word "broke" in the karate 
champion example; perhaps such checks could be 
made in the visual periphery and, when posi- 
tive, result in saccades that skip the predicted 
information); or the expectation may be formed 
after beginning to read the predictable informa- 
tion followed by skipping ahead to the next 
linguistic unit ("Oh. They're talking about 
this now. Well there's no doubt how it will 
turn out. I can pass this by."); or temporary 
binding of a schema variable (essentially a 
verification of fit) may be more rapid than 
more durable instantiation; or less metacognitive 
activity (pondering, studying, rehearsing, 
etc.) may be devoted to predictable information, 
given its derivability (this also relates to 
automaticity, obviously). Regarding auto- 
maticity, it seems likely that the amount of 
conscious processing required would be nega- 
tively correlated with the goodness of fit to 
prior knowledge. Thus conscious attempts to 
make sense of predictable information would be 
expected less often. Also, related to the 
suggestions above regarding expectations and 
rapidity of processing, the operation of some 
preattentive process (in the sense of Neisser, 
1967) is a possibility. Naturally, it may be 
the case that all of these factors are con- 
tributing. However, some of the factors may 
be mutually exclusive. For example, if default 
values are processed automatically, an expecta- 
tion and confirmation process may be redundant. 
Determinants of Performance Variability 
Occurrence of superficial processing and 
failure to store information probably depends on 
more than predictability or derivability con- 
sidered in isolation. For one thing, the 
228 
derivability of other information in the dis- 
course will have an effect. The greater the 
proportion of fit to one's schemata for the dis- 
course as a whole, the more likely it is that 
conforming information will be left to be de- 
rived. If a story takes place in a restaurant, 
and all the restaurant-related information is 
typical, then that aspect of the story can be 
stored with the abstract schema node "typical 
restaurant activities." However, when the pro- 
portion of fit is poor, i.e., some atypical 
events occur, even typical, predictable events 
may have to be stored. 
Occurrence of superficial processing is 
also likely to be affected by the extent to 
which the system is taxed. When the system is 
overloaded, as when there is a large amount of 
information to be acquired or the time to 
acquire the information is limited, more super- 
ficial processing and leaving of information 
to be derived probably goes on. Perhaps the 
system has flexible criteria for derivability, 
reducing criteria under overload conditions and 
increasing them when processing load is light 
(and when demands for recall accuracy are high 
or when subsequent availability of the informa- 
tion is limited). Briefly digressing, there may 
be a temptation to confuse superficial pro- 
cessing of derivable information with skimming. 
However, skimming is a selective seeking and 
then deep processing of situationally important 
information (see FRUMP, in Schank & Abelson, 
1975) whereas superficial processing involves 
selectively not processing deeply information 
perceived as derivable, however important it 
might be. In other words, the same information 
that might receive more attention while skimming 
may receive less attention in normal situations 
if the information is derivable. This will 
happen to the extent that skimming results in 
shallow processing of earlier information that 
is the basis for the derivability of the later 
information. 
Besides context-based variability in 
derivability criteria, research in the psychology 
of prediction indicates the potential operation 
of a general bias in determining the criterion 
for derivability and superficial processing. 
For example, Fischoff (1975, 1977) has found 
that when people are told that some event has 
occurred, they increase their subjective 
probability estimate of the likelihood that 
the event was going to occur. Similarly, estima- 
tion of how much was known before being given a 
correct answer increases when the answer is 
provided. In the case of superficial processing 
of information in discourse, it is possible that 
the derivability of information is overestimated 
after it is explicitly encountered. It seems 
to be a fairly common experience, for example, 
to not write down an idea that you are sure 
will be derivable again later, only to find 
subsequent derivation impossible. What is being 
suggested here is a source of forgetting not 
usually discussed in memory theories: super- 
ficial processing of information whose deriv- 
ability has been overestimated. 
The Form of Expression of Derivable Information 
Semantic content, prior knowledge, and task 
contexts are not the only determinants of per- 
ceived derivability. The linguistic form in 
which information is expressed will sometimes 
provide signals of what information is already 
known or can be taken for granted, as when 
information is expressed near the beginning of 
a sentence (c.f., Clark & Haviland, 1977, on 
the given-new strategy). Taking an example from 
Morgan and Green (in press), compare sentences (1) 
and (2). 
(1) The government has not yet acknowledged 
that distilled water causes cancer. 
(2) That distilled water causes cancer has 
not yet been acknowledged by the 
government. 
In (2) 
of the 
tilled 
there is a stronger implied presumption 
truth of the proposition regarding dis- 
water and cancer than there is in (1). 
In general, it seems that placing informa- 
tion in a sentence-initial subordinate clause 
lowers the superficial processing criterion. 
Consider continuations (3) and (4) of "The 
karate champion hit the block." 
(3) The block broke, and then he bowed. 
(4) After the block broke, he bowed. 
The block's breaking would appear to be more 
taken for granted in (4) than in (3). 
Linguistic signals of predictability or 
derivability need not be implicit. Consider 
continuations (5), (6), and (7) of the same 
sentence as above. 
(5) Obviously, the block broke. 
(6) As you would expect, the block broke. 
(7) Naturally, the block broke. 
Words like "clearly" and phrases like "of 
course" are explicit linguistic signals that 
information to follow is predictable and can 
be superficially processed. However, one would 
expect that such signals could have their effect 
only for information within an acceptable range 
of plausibility. That is, a plausible but not 
predictable continuation may be more likely to 
be taken (erroneously) as predictable when 
preceded by a linguistic signal. However, if 
the information contains salient implausible 
aspects or something clearly irrelevant, a 
signalling phrase such as "as you would expect ~' 
might result in more attention being devoted 
to the continuation information. 
Implications for the Nature of Discourse Memory 
To the extent that discourse is super- 
ficially processed, memory must be reconstructive 
rather than reproductive. Rather than re- 
trieving traces or instantiations of past 
experience, the past must be inferred or derived. 
Just as a paleontologist reconstructs a dinosaur 
from bone fragments, the past must be recon- 
structed from the incomplete data explicitly 
stored. Evidence for such reconstructive 
229 
processes has been provided by Spiro (1977), who 
found a pervasive tendency for subjects to pro- 
duce predictable meaning-changing distortions 
and importations in text recall under certain 
conditions. In general, when subsequently en- 
countered information contradicted continuation 
expectations derived from a target story, the 
story frequently was reconstructed in such a way 
as to reconcile or cohere with the continuation 
information. This process of inferring the past 
based on the present was termed accommodative re- 
construction. After a long retention interval, 
subjects tended to be more confident that their 
accommodative recall errors had actually been 
included in the story than they were confident 
about the accurate aspects of their recall. Why 
should such gross errors occur and then be 
assigned such high confidence? Part of the answer 
surely involves their function in producing co- 
herence. Still, it is somewhat surprising that 
subjects should be so sure they read information 
that bore not even a distant inferential relation- 
ship to what they actually did read. 
Spiro suggested that the basis for such an 
effect may be in the way information is treated 
at the time of comprehension; namely, it is 
superficially processed and not stored in long- 
term memory. Then, when remembering, individuals 
should know (at least tacitly) that considerable 
amounts of predictable or derivable information 
they have encountered will not be available in 
memory. In that case, recall would typically 
involve deriving a lot of missing information. 
Accordingly, it would not be surprising that 
subjects faced with memories that lack coherence 
would assume that missing reconciling information 
was presented but only superficially processed 
at comprehension. The information could then 
be derived at recall with high confidence. Hence 
the capacity for restructuring the past based on 
the present. 
Individual Differences 
A final caveat should be offered regarding 
the superficial processing effect, but also 
applicable to all research on schema-based pro- 
cesses in comprehension and memory. The assump- 
tion is usually made that there are no qualita- 
tive differences between individuals in the 
manner in which discourse is processed. How- 
ever, Spiro and his colleagues have recently 
found that reliable style differences can be 
predicted in children (Spiro & Smith, 1978) and 
in college students (Spiro & Tirre, in prepara- 
tion). Some individuals appear to be more dis- 
course bound, tending toward over-reliance on 
bottom-up processes. Others are more prior 
knowledge bound, tending toward over-reliance 
on top-down processes. For the adult bottom- 
up readers, prior knowledge obviously must be 
used to a certain extent in comprehension. How- 
ever, where use of prior knowledge is more 
optional, e.g., in providing a scaffolding for 
remembering information (Anderson, Spiro, & 
Anderson, 1978), the bottom-up readers capitalize 
less. Whether the latter type of individual will 
evince less knowledge-based superficial pro- 
cessing (again an optional use of prior know- 
ledge) is a question currently under investiga- 
tion. 
References 
Anderson, J. R., & Bower, G. H. Human associa- 
tive memory. New York: Wiley, 1973. 
Anderson, R. C., Spiro, R. J., & Anderson, M. C. 
Schemata as scaffolding for information 
in text. American Educational Research 
Journal, 1978, in press. 
Bransford, J. D., & McCarrell, N. S. A sketch 
of a cognitive approach to comprehension. 
In W. B. Weimer and D. S. Palermo (Eds.), 
Cognition and the symbolic processes. 
Hillsdale, N.J.: Erlbaum, 1975. 
Charniak, E. Organization and inference in 
a frame-like system of common sense 
knowledge. In proceedings of Theoretical 
issues in natural language processing. 
Cambridge, Mass.: Bolt Beranek & Newman 
Inc., 1975. 
Clark, H. H., & Haviland, S. E. Comprehension 
and the given-new contract. In R. 
Freedle (Ed.), Discourse processing. 
Hillsdale, N.J.: Erlbaum, 1978. 
Collins, A. M., & Loftus, E. F. A spreading 
activation theory of semantic processing. 
Psychological Review, 1975, 8._2_2, 407-428. 
Fischoff, B. Hindsight # foresight: The 
effect of outcome knowledge on judgment 
under uncertainty. Journal of Experi- 
mental Psychology: Human Perception 
and Performance, 1975, i, 288-299. 
Kintsch, W. The representation of meaning in 
memory. Hillsdale, N.J.: Erlbaum, 1974. 
Minsky, M. A framework for representing know- 
ledge. In P. H. Winston (Ed.), The 
psychology of computer vision. New York: 
McGraw-Hill, 1975. 
Morgan, J. L., & Green, G. M. Pragmatics and 
reading comprehension. In R. J. Spiro, 
B. C. Bruce, and W. F. Brewer (Eds.), 
Theoretical issues in reading comprehen- 
sion: Perspectives from cognitive 
psychology, linguistics, artificial 
intelligence, and education. Hillsdale, 
N.J.: Erlbaum, in press. 
Neisser, U. Cognitive psychology. New York: 
Appleton-Century-Crofts, 1967. 
Rumelhart, D. E., & Ortony, A. The representa- 
tion of knowledge in memory. In R. C. 
Anderson, R. J. Spiro, and W. E. Montague 
(Eds.), Schoolin 9 and the acquisition 
of knowledge. Hillsdale, N.J.: Erlbaum, 
1977. 
Schank, R. C., & Abelson, R. P. Scripts, 
plans, goals, and understanding. 
Hillsdale, N.J.: Erlbaum, 1977. 
230 
Spiro, R. J. Remembering information from text: 
The "State of Schema" approach. In R. C. 
Anderson, R. J. Spiro, and W. E. Montague 
(Eds.), Schooling and the acquisition of 
knowledge. Hillsdale, N.J.: Erlbaum, 1977. 
Spiro, R. J. Constructive processes in text 
comprehension and recall. In R. J. Spiro, 
B. C. Bruce, and W. F. Brewer (Eds.), 
Theoretical issues in reading comprehen- 
sion: Perspectives from cognitive psy- 
chology, linguistics, artificial intelli- 
gence, and education. Hillsdale, N.J.: 
Erlbaum, in press. 
Spiro, R. J., & Esposito, J. Superficial pro- 
cessing of explicit inferences in text 
(Tech. Rep. No. 60). Urbana, Ill.: Center 
for the Study of Reading, University of 
Illinois, 1977. 
Spiro, R. J., & Smith, D. Distinguishing sub- 
types of poor comprehenders: Patterns of 
over-reliance on conceptual- vs. data- 
driven processes (Tech. Rep. No. 61). 
Urbana, Ill.: Center for the Study of 
Reading, University of Illinois, 1978. 
Frederiksen, C. H. Representing logical and 
semantic structure of knowledge acquired 
from discourse. Cognitive Psychology, 
1975, ~, 371-458. 
Meyer, B. J. F. The organization of prose and 
its effects on memory. Amsterdam: North 
Holland, 1975. 
Footnote 
This research was supported by the National 
Institute of Education under Contract No. US-NIE- 
C-400-76-0116. 
231 
