Department of Psychology
University of California
Los Angeles, California 96024

(net address: Jmiller @ Rand~Unix)

Dr. Feigenbaum,

This letter concerns my request for SUMEX~AIM access. I am
currently a graduate student in cognitive psychology at UCLA (as
well as a consultant at Rand~ISD), and am in the process of
completingemy dissertation [[1]]: a simulation model of semantic
memory with the goal of predicting recent and future experiments
On memory for the meaning and wording of Sentences and texts. The
work on my model, a UCI LISP program, was about one~third
completed when the computer on which I had been working was
removed to another section of UCLA. I cannot get time on the
machine at this time, and its future plans are very unclear: Lisp
may not even be provided should it later become available.
Therefore, I would like to move my dissertation work to SUMEX,
where a number of Lisp systems are of course available.

The current state of my work is outlined in the status
report enclosed, but a short summary here might be appropriate.
There have been a number of psychological experiments on the
retention of sentences and texts that indicate that people
generally have very good memory for the meaning of this material,
but very poor memory for the exact words in which it was
presented: discrimination of a presented sentence from a new, but
synonymous, sentence is rarely much above chance.
Correspondingly, the prevailing view in psychology is that
incoming linguistic material is converted from its component
words into an abstract, conceptual structure that represents the
meaning of the passage. Wording information be available for a
short period of time, but is soon lost. This position has been
advocated in psychology by the Norman and Rumelhart group at
UCSD, and in computer science by Roger Schank and his associates.

At the same time, some of the work here at Rand on text
comprehension (by Barbara and Fredrick Hayes~Roth and Perry
Thorndyke) has indicated that there are very real and observable
effects of wording variation on the recall and recognition of
sentences. For instance, subjects are much more successful at
forming inferences between two sentences when the common
components of those sentences are identical, rather than
paraphrases of each other. If there were an underlying abstract
representation, synonymous phrases should be represented
identically, and no differences in inferencing ability should be
observed. Hayes~Roth and Hayes~Roth have also found that,
despite the near~chance recognition performance of sentence
wording, subjects are reliably faster and more confident when
verifying old than paraphrase sentences. This work suggests that
words are differentially represented in memory, but that some
memory processes are occluding these differences.

My contention is that a semantic network controlled by a
limited capacity spreading activation process could account for
these results, This feature of the model encodes a sentence into
memory by spreading activation throughout those parts of the
network that are semantically related to the parts of the
Sentence. New words are incorporated into the context of the
Studied sentence as they are contacted by this spreading process.
While such a process will build a strong representation of the
sentence into memory, it will come at the cost of interference
with the wording information, since both a word ("frightened")
and one of¢its synonyms ("scared") may be encoded into the same
context. A recognition test would then falsely recognize "John
scared the babysitter" after having studied "John frightened the
babysitter", presuming that "scared" had been activated by the
model’s study of "frightened",

The model can further predict the Hayes~Roth findings of
higher confidence and faster reaction times to old sentences,
Although both old and paraphrase sentences will be "recognized"
as above, confidence ratings to old sentences should be higher as
a result of the more highly activated state of such Sentences,
The words in old sentences would be directly activated by the
full amount of activation which initially entered the sentence,
while synonymous and other related words would receive only a
portion of that original activation: the limited Capacity feature
of the model requires a division of activation entering a word
node among all the links leaving that node. Confidence ratings
will then vary as a function of the amount of activation at the
node, being greater for older, more highly activated words. A
similar manipulation could be expected to produce the finding of
shorter reaction times to old sentences.

The current state of the model is this: a set of
experimental sentences have been selected and a representation of
the words in these sentences has been constructed from their
dictionary definitions. The initial study and test routines have
also been completed, so that the model is capable of “studying” a
Series of sentences, and then determining whether or not a
particular sentence had been studied earlier. As previously
mentioned, the model is falsely "recognizing" new sentences which
are Synonymous to previously studied sentences. The initial
successes described above and in the enclosed status report
indicate that the general direction of the research is an
appropriate way to deal with the problem. What remains to be done
is to develop the aspects of the model responsible for generation
of correct test reaction times and contidence ratings, and
expansion of the model to simulate a number of related: memory
Phenomena (See the status report for instances of these).

AS I mentioned before, I think that my work here is about
One~third finished; ideally, I would like to finish my
dissertation by January. My needs at SUMEX actually seem quite
moderate: the current model was running in about 4@K in UCI LISP,
and I don’t foresee any drastic increase in that (although I will
likely move the model to INTERLISP). Disk space should also be
minimal: probably five to ten 30~4@ block files. Given my other
projects at UCLA and Rand, I would probably be using about 12
hours of terminal time per week; as much as one~half of that
would likely be during evenings and weekends.

I hope to hear from you soon, and I hope that any
affiliation between us will be productive. If you need any
further information, please let me know. Thank you,

Sincerely,

James R. Miller
({1]]

My dissertation committee is:

Robert A. Bjork, Department of Psychology, UCLA
Thomas D. Wickens, Department of Psychology, UCLA
Jackson L. Beatty, Department of Psychology, UCLA
Kenneth M. Colby, Department of Psychiatry, UCLA
Tyler Burge, Department of Philosophy, UCLA
Barbara Hayes~Roth, The Rand Corporation