KNOWLEOGE ANDLANGUAGE I. Kurcz, G.W. Shugar and J .H.Danks (editors) ' ElsevierScience Publishers B.V. (North-Holland), 1986
189
CONSTRUCTION OF INDIVIDUALIZED TEXTS FOR THE TRANSMISSION OF KNOWLEDGE THROUGH DISCOURSE Jean-Francois Le Ny, Luc Carite and Sebastien Poitrenaud
A cognitive system for computer-assisted instruction, called CINACC (for Computation of Information Necessary for Acquisition or Correction of Knowled~e), is present› ed. It is designed to create in students memory a pre› determined "final state of knowledge", which is de› scribed as a set of elementary mental propositions. Knowledge is transmitted through text items which are stored in the computer’s memory and can be presented to the students on a video screen. Questions are put to the students during the study, and their responses are evaluated by the system. The outputs of this evaluation are used to determine the presentation of subsequent text items, or questions, according to predetermined rules. So during a particular session of instruction, each student is presented with his or her own individ› ualized text and set of questions, constructed by the system. Questions belong to several types: some are conceptual questions, such as, "What is an X?", in which X designates a concept. An automaton, using pro› cedures from Artificial Intelligence and special cog› nitive schemata, enables the analysis and evaluation by the system of written responses to such conceptual questions. The system in general, and the analyzer-ev~ uator in particular, help psychologists to make explic› it the links between language and declarative knowledge during acquisition of knowledge. This paper is based on more extensive
ress,
whose main ideas have been presented
Denhiere, 1982). Here we discuss
now in prog›
elsewhere
(Le Ny
&
theoretical
assumptions
The task we have undertaken is the development
of what may
underlying our research encountered.
be termed
~
some
research
and a few of the new problems
cotnitive system
of computer-assisted
The first version of this system viates the French for something
mation Necessary
for Acquisition
we have
teaching."
was named CINNA, which abbre›
like
"Computation
of Knowledge."
of New Infor›
The second
ver-
J.-F. Le Ny. L. Carite and S. Poitrenaud
190
sion is now completed;
it is named CINACC, i.e.,
of Information Necessary
edge."
for Acquisition
CINNA/CINACC was conceived
as a didactic
creating in the memory of the student some
of knowledge.
In order to achieve
endowed with the combined son’s
signed
Micromega),
specifically
resources
a specialized
for analysis
system
of
state
of a microcomputer
program, and procedures of cognition.
(Thom› de›
Students are pre›
and work in inter›
that this approach will provide new data and fresh
about acquisition
pecially
capable
predetermined
sented with written material in this context We hope
of Knowl›
this, CINNA/CINACC has been
action with the system. evidence
"Computation
or Correction
texts.
of knowledge
through discourse,
es›
SOME THEORETICAL ASSUMPTIONS The A in CINNA/CINACC refers
assimilation, for example
of knowledge
to a student’s acquisition,
about some
domain or subdomain,
a part of a chapter in meteorology,
psychology.
This knowledge
hending texts.
physics,
or ~
or
is acquired by reading and compre›
It must be brought about in such a way and to
such a standard that the student can later use it effectively
to perform various cognitive
identify concepts,
discourse.
similation"
solve
tasks:
problems,
From the perspective
analyze specific
and comprehend
of cognitive
here amounts to "deep
situations,
and produce
psychology,
memorization",
"as›
i.e., acqui›
sition of highly retrievable knowledge.
So CINNA/CINACC may be defined as a didactic
system
whose
aim is to create, in the student’s memory, some complex Jtstate of knowledge Jt regarding the domain ~. The first question, there› fore,
is how to characterize
tem considers called
or describe
Jttarget states of knowledge." ~.
the latter being
In the case
target state is defined relative to some
ing domain
such a state. The sys›
both actual and intended states,
of CINACC the
future time tf’ regard›
191
Transmission of Knowledge through Discourse
The general assumptions
edge
we use for the acquisition of knowl›
by reading a text are the following:
1. Before
a teaching session,
state of knowledge,
the student is in an initial
SKi’ regarding D. He or she may know
little, or much about
nothin~
Q.
2. As the student reads the text, he or she processes
presumably
comprehends
it, and stores
in his or her long-term memory.
many related successive
the knowledge
thus gained
As the student reads,
states of knowledge,
then,
one for each mo›
ment of time, follow
one upon another within the memory.
in empirical studies
may vary with the specific
time interval we have to take into consideration the research
or of the actual application.
3. After a session,
knowledge,
SKf,
Of course,
sessions
regarding
requirements
will presumably
be modi›
by forgetting and other phenomena.
this into consideration. concerning
Q.
Let us suppose
In this case,
that two
the student’s
initial state of knowledge
in session
between the end of session
1 and the beginning of session
uct of knowledge
2 will be the joint prod›
acquired during session
1 and of its evolution
we are reset to point 1 above. 4. For any Q, any session,
target state of knowledge.
and any student, the system
The goal of the system
the actual final state of the student’s knowledge possible
to the target state. The problem,
any teacher,
is to calculate
the best,
shortest,
as close
in particular, in psychology,
For present purposes,
and most effi›
Although
in cognitive
to a
this ques›
science,
and,
we can give a tentative answer.
a state of knowledge
as a set of psychological
as
How can a state of knowl›
(of a given domain) be characterized?
tion is still a matter of research
has a
for CINACC as for
predetermined target state.
To return to our initial question:
2; so,
is to bring
cient path from an initial state of a student’s knowledge
edge
of
Q and the session.
this state of knowledge
are held
The
in theory or
the student is in a given final state of
fied after the session, CINACC takes
it,
propositions,
will be characterized
each
containing a predi-
192
J.-F. Le Ny, L. Carite and S. Poitrenaud
cate and one or more arguments.
here to psychological knowledge
proposition
Several types
Propositions
of connectives,
as and, or, because,
Eefore,
etc.,
may also
refers
components
within the student’s mind (we prefer to say,
the student’s memory).
ments.
Psychological
entities taken as cognitive
within
be taken as argu›
logical
or semantic,
such
are used to relate these
positions.
Thus the set of propositions
chological
propositions
of
has a structure and
pro›
might alternatively be presented as a graph. This use of psy› of text, especially
is fairly common
of stories
in cognitive
(for example.
Denhisre,
analysis 1982;
& Le Ny. 1979; van Dijk & Kintsch. 1983; Kintsch & van
Denhisre
Dijk, 1978; Le Ny, 1979).
with texts that convey important differences
CINACC deals
a didactic
in the same general way
content,
although there are
as regards the content of the concepts
involved. We shall return later to this question of concepts. Below is a sample
mentary meteorology, 1.
HEAT
(cx,fJ)
3.
SEND
(ex,B.
2.
BECAUSE
where
0: =
set of propositions
comprising
((1.),
in the domain of ele›
fairly simple
(3.))
sunbeams)
the sun, B = the earth
The set might correspond to texts such as:
earth because sends
concepts.
sunbeams
it sends
sunbeams
"The
to the earth,"
to the earth and, therefore.
sun heats the
or "The
sun
it heats the earth,"
or "The heating of the earth (by the sun) is due to the fact that the sun sends
sunbeams
Standard logicians,
chologists,
postulate
to the earth."
and, following
the existence
their example,
of a constituent
many psy› of dis›
course.
the proposition, which has the same predicative form or
cording
to this assumption,
structure as that just used
with the syntactic
in the discussion
sentences
and semantic rules
bring about adequate processing
knowledge
in the students’
In this view, a sentence
produced
of knowledge.
in accordance
Ac›
of language are able to
and, thereafter, adequate
minds. ~
fied on two different grounds.
is something
that can be identi›
First, on physical
grounds
(or
Transmission of Knowledge through Discourse
193
on grounds that can be analyzed as such), ~ is a certain bundle of acoustic or visual stimuli or cues. Second, on a functional level, the communicative role played by a sentence consists, under specific linguistic and pragmatic conditions, in creating some psychological proposition in the working memory of a per› son, and later on in his or her long-term memory. Theoretically, t.nI s comes down to assuming that at least some of the knowledge stored in the mind is present in a form isomorphic to, or at least compatible with, discourse. This seems to be close to Vygotsky’s (1962) view. A similar point of view has been presented, though in a rather different theoretical con› text, by J. Fodor (1975), who called such a form "the language of thought." This does not come down to saying that all knowl› edge necessarily takes on such form: perceptual, imaginal, and procedural knowledge may also be considered to exist in a simi› lar or a different form. As a matter of fact, the notions con› nected with the expressions propositional and imaginal are in no way mutually exclusive (Denis, 1979; Denis & Le Ny, 1984; Le Ny, 1979). By definition, CINACC handles only those types of knowledge that can be created or modified by discourse, or more precisely, by reading. Let us summarize how reading is assumed to operate on two different states of knowledge. We shall call one of these the current state of knowledge (about D) for a given student at a given time. We shall call the other the next target state of knowledge, to be reached by this student at some near future time. Observe that it is in the teacher’s mind that this target exists as a cognitive content of his or her intentional task of teaching. For instance, at some time, a particulax student does not yet know that a fraction of the energy emitted by the sun as sunbeams is absorbed by the earth’s atmosphere, that a sec› ond fraction is reflected by the earth’s surface, and that a third fraction is absorbed, and is stored, by the earth. A tea› cher’s next target may be to bring about a particular state of knowledge in which the new propositions corresponding to this
J.-F. Le Ny, L. Carite and S. Poitrenaud
194
knowledge
are contained in the student’s memory.
of CINACC is that such a target may be included
as a subtarget of some
specific
program.
Ihemilln tasks of CINACC consist
rent state of knowledge sessing
the difference
The principle
into the system
in evaluating a student’s
concerning
some
cur›
item of knowledge,
as›
between the current state and the next
target state, and in light of this difference,
next item of information to be presented
deciding
on the
to the student, so as
to allow him or her to attain the next target state. These
may be summarized briefly possible
steps:
1) student h does 2) but student
3) therefore,
and simply as involving the following
not know that £i’
h should know that £i’
let us present ~ with an item of text tt
where the item of text tt of an item of knowledge,
These
tasks are performed
get states.
All these
targets
structures
text and so progressively
edge
by a final state
i, to be an expression
is assumed
i £i’
for successive
ward achieving the final goal: their cognitive
tasks
current and next tar›
(or subtargets)
by processing
successive
items of
to the final tar›
get state determined by the teacher.
states of knowledge
to›
replacing the initial state of knowl›
as similar as possible
As pointed out before,
are directed
to have the students transform
the way in which the intermediate
are taken into account
depends
on the choice
of time interval or unit and on the size of the information
bundles adopted.
In the first version,
the bundle that CINNA
usually dealt with was the short paragraph, but CINACC can as
well deal with complete or concepts.
As the system
or elementary
provides
sentences,
computer-composed
word meanings,
texts on which stu›
dents work individually, its main advantage is to individualize the choice
and order of the sentences
or paragraphs that con›
stitute a text so that each particular student can be presented
with his or her own particular text or sequence accordance with predetermined cognitive rules.
of texts,
in
Transmission of Knowledge through Discourse
195
STRUCTURE AND OPERATION OF THE SYSTEM We shall now briefly describe how the system is constructed and how it works. For convenience we present this description so as to correspond with the preceding description of the sys› tem’s three main tasks. These tasks may be analyzed into three components: 1) does student g know that Ei? -- the two alternative re›
sponses being y or n (for "yes" and "no"); 2) given two alternative items of text, tt 1 and tt 2, decide: if 1) = n, then tt otherwise tt2; 1, 3) present g with the chosen item of text. Thus, the three components are: (1) questions (about states of knowledge), (2) decisions (about presentations), (3) presentations (of text items). Accordingly, the system works from three main sets of data, which have been set up in advance in three different files in the computer’s memory. Let us begin with the file corresponding to step (3). This file contains a set of text items. In the first applications, these are, as has been said, paragraphs, sentences, and parts or skeletons of sentences concerning the domain; these text items, or texts, convey all the information needed to instruct any student -- from a predetermined population -- in the domain in question. In other words, they convey all the information that the students need to process as their previous knowledge is changed in direction of the target state of knowledge. Of course, the content of this file is determined by an expert teacher of the particular subject matter of the domain. The file corresponding to step ~) contains a set of questions the students are asked at predetermined times, in order to evaluate their successive states of knowledge. As shown in the general "question" schema outlined above as task 1, possible responses have to be explicitly stored in the file in addition to the questions, since evaluation of the actual responses is a necessary part of the task of questioning in the system. In
J.-F. t» Ny, L. Carite and S. Poitrenaud
196
general,
a student’s actual response
it with one of the predicted complex
process:
a very simple
controlling
and we shall introduce
further below.
The second
questions,
Such matching can be a
the alternative we presented
case,
possibilities
is evaluated by matching
responses.
file
in the system
contains the decision
the presentation of text items,
in accordance rules.
(1) If subject
£
rules
or of subsequent
with the evaluations obtained as the
output from the questioning process.
as production
(y or n) is only
more sophisticated
These rules are organized
The simplest schema
E~
has given response
for them is:
~i,
to question
present him with text tt and afterwards with question w Of course,
more
schemata
~omplex
can be used,
then
~k
such as,
(2) If subject h has given response r i or response r i to ques› . - . -a. -c. and . tion ~l, and response ~ or response £~ to question ~J,
etc.,
the present him or her with text tt followed by text v’ I by text tt and afterwards with question ~ , etc. z’
ttw’ followed So,
in this respect,
based
system
designed
the architecture
The contenre of the text, question,
have to be conceived
didactic
texts,
practice.
or as a field for cognitive
and decision
ing a large set of psychological processing,
rules are considered hypotheses
as involv›
concerning
question answering, structure of concepts,
One kind of current research
re›
point of view,
vance of rules assigning texts or further questions ses.
all
itself may be viewed as a matrer
From the present experimental cognitive
questions,
rule files
and general didactic
expertise
of pure empirical experience search.
and decision
and organized by an expert with regard to
both the domain of knowledge Actually,
of CINACC is a knowledge›
for the solution of teaching problems.
text
and rele›
to respon›
based on this system
consists
in studying the various routes that various students follow
through the network of knowledge
nothing more about such research
enough data to reach conclusions. on the particular types
so provided.
here,
We shall say
as we have not collected
We shall, however,
of questions
elaborate
and their evaluation.
Transmission of Knowledge through Discourse
197
TYPES OF QUESTIONS All in all, we use three categories of questions, with three corresponding degrees of complexity involved in their evalua› tion. One category contains conventional multiple-choice gues› tions. The students respond to these by pressing on the key› board, in accordance with previous instructions, the key they judge appropriate for their response. The system’s evaluation of these responses and the decision on the next presentation are then direct. Examples of such questions are presumably not needed. This type of question, with the possible responses and the corresponding decision rules, is not necessarily trivial. In addition, such questions are often used as the final step after a series of more complex questions. A second category consists of open guestions with a limited range of possible responses. Only the teacher or experimenter knows that this range is limited, so that they appear to the students to be true open questions. Responses are evaluated by the identification of key words. Thus, in the program, there is a fourth file which contains labeled classes of words. Let us illustrate their role by an example taken from classroom teach› ing of psychology. After students have read a chapter of text taken from an introductory handbook in the psychology of memory. they can be asked: What are the successive phases that psycholo› gists have, as a rule, distinguished in the study of memory? The correct reponse is of course some combination of words including at least three words such as fixation, storage, and retrieval. But many other words may be considered acceptable as they pertain to the same classes as the words just mentioned. So we can state that a correct response includes at least one word of the class "fixation", plus one of the class "storage", plus one of the class "retrieval". In the third class, for in› stance, we can decide to accept, in addition to retrieval, such words as recall, recollection, reactivation, access, recogni› tion, etc., as well as the corresponding verbs. The file will contain only word skeletons, such as "reactiv-", so as to per› mit matching with "reactivates", "reactivated", etc., as well
198
I.-F. Le Ny, L. Carite and S. Poitrenaud
as with "reactivation".
Everything in the file
rclevant decision
Such evaluation by key words may be ap›
pert in the subject
matter, too,
rules.
plied to a large range of open questions,
kinds of wh- questions:
which,
including various
when. where, etc.
But let us now examine our third category
shall call them conceptual "What is an X?",
concept
guestions.
in the domain under study.
are not intended to elicit
they are not intended to elicit a response conditions
ther a text expressing
mind which are related to the concept not consider A special
the elicited
(Woods,
1975) which produces
sponses
consists
interpretation and evalua› This subprogram
Transition Network type
is a
a definite abstract representation
given by the student. Evaluation of such re›
in trying to match this abstract representation
of the current response
previously
The teacher mayor may
to do this job.
semantic analyzer of the Augmented
of the response
being an X, but ra›
units in the subject’s
text appropriate.
subprogram of discoure
tion is being implemented
X.
true definitions
In other words,
stating the necessa›
for something
a set of semantic
We
usually an important
from a student but rather characterizations. ry and sufficient
of questions.
They are of the type:
with X being a concept,
These questions
is up to the ex›
and so is the statement of the
introduced
the expert teacher.
ral process.
with an abstract target representation
into the machine under instructions
Figure 1 presents
of
an outline of this gene›
The machine program of analysis and interpretation extracts
the relevant information from the subject’s
vides the system
of this response.
archized couples,
response
with an intermediate semantic
and pro›
representation
This representation is in the form of hier›
each of which consists
of one attribute and
one value of this attribute. This state is intended to mimic
human comprehension.
The evaluation program compares
this intermediate representa›
tion with a target representation of the concept
under consider-
199
Transmission of Knowledge through Discourse
ation. (Such type representations
are exemplified
in Tables 2
and 3.) This stage is intended to mimic evaluation by a human judge.
The evaluated representation of the response.
i.e., the
output of the program’s evaluation, is in a form usable
left part of the decision
ditional questions)
c rules.
follows
Presentation of texts
from their application.
SUbject’s response
~ncerning
as the
(or ad›
X
I I I I I
AUTOMATON I
[ -
. ----Intermediate semantic ,t; representation of the response
~
Evaluation program Target type representation
_ _CJ.f
~
_
Evaluated representation of the response
I
) IText presentRtion
Figure 1. General procedure
for response
evaluation.
I
200
J.-F. Le Ny. L. Carite and S. Poitrenaud
EVALUATION OF RESPONSES In this paper we shall not describe the subprogram used as a semantic analyzer and interpreter of texts. Instead, we shall sketch the information structure used to evaluate students’ re› sponses to conceptual questions. We shall exemplify this sketch with two psychological concepts, "f’or-ge t t i.ng" and "interference". We assume that the concept "forgetting!! is present, with an appropriate basic content, in the initial knowledge of any stu› dent belonging to a normal population. In contrast, we assume that initially the uninstructed subjects have no concept of "interference I! with a meaning relating to scientific psychology. But this concept may be considered an important item of knowl› edge to be included in a target state of knowledge for a teach› ing session devoted to the subdomain of !!forgetting". The two concepts should be present in a student’s final state in a way that can be expressed as FACTOR OF (interference, forgetting). "Interference I! (and I!forgettingl!) can be used here in either a causative or a merely descriptive sense, depending on the gen› ~ru epistemological view of the user. Our illustrative use of these concepts here does not commit us to either view. The teacher must specify in the outline of the target state of knowledge which items of meaning he or she considers to be essential to characterize any new concept introduced (in this case, I!interference"). Items of meaning here refers to items of knowledge that are assumed to be basic to the target concepts, and to the interrelationships of these items. They are consid› ered basic here in that they would already be known by the stu› dents and are so unambiguous that there is no need to verify their meaning for any student. In other cases, they may be iwms of meaning that have been already verified during this or some previous session with a given student. Let us illustrate this with the concept I!interference!! by stating what the target re› presentation of this concept should contain. First, it should refer in this context to I!learning!!. In the same way as forl!forgettingl!, we have good reason to believe that
201
Transmission of Knowledge through Discourse
an average student knows what learning is, at least well enough for our purpose. So "learning" will be considered a basic item of knowledge in our program , and will not have to be verified. In addition, the target representation should state that "inter› ference" involves (at least) two situations and processes of learning, and that the two processes interact so that perform› ance is worse for one of them if the other has taken place than if it has not. But we must describe this interaction in a pre› cise way in the target representation, in particular as con› cerns time relationships between learning situations and proc› esses. Either the interfering situation occurs before the other, in which case proactive interference is produced, or else it occurs after, in which case retroactive interferenoe is produred. Furthermore, we must take into account the possibility of improvement in performance (i.e., of transfer) rather than im› pairment (i.e., interference) after the two processes have in› teracted. Accordingly, we may state the target representation of the two interrelated concepts of "interference" and "transfer" as an equivalence formula (in quasi-natural language, as proposed by the authors; various wordings may express it in the teach› er’s or student’s language): For any couple, ~, ~, of learning situations in a subject (given an appropriate description of "learning situation" and "performance due to the previous learning situation"): 1- IFF performance on :2 WHEN 1l, PRECEDES £ IS INFERIOR TO performance on £ WHEN 1l, DOES NOT PRECEDE £ THEN this ifj proactive interference (of ~ with ~) 2. IFF
[replacing with and with in rule 1]
performance performance PRECEDES FOLLOWS
.
on £ on g (at an appropriate time)
202
J.-F. Le Nv. L. Carite and S. Poitrenaud
performance on a (subsequent to 2) WHEN ~ FOLLOWS !2: IS INFERIOR TO performance on a (at a corresponding time) WHEN E DOES NOT FOLLOW ~ THEN interference (of b with !!) is retroactive this 3. and 4. IFF [substitute IS SUPERIOR TO for IS INFERIOR TO in rules 1 and 2J THEN this is proactive transfer or retroactive transfer, respectively. Thus we have systematically characterized four interrelated concept~ among which "retroactive transfer" is a possibility, though barely mentioned in published reports. This type of char› acterization is used in GINACC in three ways: as part of the target state of knowledge, as a description of the basic seman› tic information the students are presented with during the ses› sion (about "interference"), and as a machine representation to be matched by the student’s responses to open questions of the type, "What is I?". In this case the question would be, "What is interference?". Since this triple use of the same representation might be confusing when we describe how responses to open questions are evaluated, we shall now use the second example, namely "forget› ting". As we have said, at least the crude concept "forgetting" is assumed to be present in the initial state of knowledge of an average student. Of course, it will also be present, but en› riched -- i.e., be broader and more precise -- in the final state of knowledge. But let us for the moment consider just the initial state of knowledge. The concept "forgetting" can be characterized in exactly same general way as "interference" was.
the
The assumption that the student is familiar with "forgetting~ no matter how the concept may have been acquired, leads to the following consequence: for our characterization to be correct,
Transmission of Knowledge through Discourse
203
it must be able to match a set of students’ actual responses (made before any lecture or learning session) to the question. "What is forgetting?". So we first set up a formula characterizing "forgetting". Then we asked twenty subjects to give a short descriptive re› sponse to the question. "What is forgetting?", and we analyzed their response texts with respect to our characterization. Be› low is our initial formula characterizing "forgetting" as unde~ stood by the average student at the outset. Observe that "for› getting" refers to a state ("to have forgotten") rather than a process ("to forget"); only focal items of knowledge are repre› sented (cf. Figure 2). For any individual x and for any content y. IFF at time t5, x NO LONGER KNOWS Y or x CAN NO LONGER REACTIVATE Y WHEREAS at some previous time t3. x KNEW y or x COULD REACTIVATE Y THENt h i s is forgetting In this context "forgetting" involves two main contrasted items of knowledge. or components of meaning, which are them› selves composed of several more basic components. The focal item of knowledge is that some individual. x. no longer knows or can no longer reactivate (recall. remember. etc~ some content, y. at some time. called t5. A second item of
knowledge may be said to be "presupposed" by the words "no long› er (knows)". used above. But we think it is more correct to say that the second item of knowledge is also a part of the con› cept "forgetting". or of the meaning of the word forgetting: that x knew y. or could remember y. at some time, t3. prior to the focal time. t5. A third item of knowledge
is implicit in this characteriza›
tion: that the state at t3 ("x knew y") had been acquired. i.e., had been achieved through some learning or memorization that
Figure 2.
Analysis
~
I
y
x knows y
~
t5
t4
to successive
Contrast
Y’
J
of time.
I
Focal period
periods
,
x no longer knows y (x has forgotten
Forgetting (state of)
Forgetting (process of)
x comes not to know y (x forgets y)
of forgetting with reference
x comes to know y
t3
t2
t1
x does not know y
IVJemory
IVJemorization (learning)
Ignorance
y)
~
is
~.
cl’
~
’l:l."
’";::
~.
D
t:-"
’" ~
t-o
~
~
~
N
Transmission of Knowledge through Discourse
205
had taken place at a still earlier time, t2. Before t2, there was an initial state of ignorance, at t1. For the sake of com› pleteness, one more time should be considered, t4, at which a process formally symmetrical to learning takes place: forget› ting. During this process, x partially knows y, or sometimes can reactivate y and sometimes cannot. So "forgetting" is char› acterized in our formula as complete forgetting; i.e., as the end product of the process of forgetting. Our assumption is that this is the dominant (prototypical) sense of "forgetting" in the average student’s mind. Figure 2 presents this analysis diagrammatically. All times mentioned ~ this characterization are periods. Ho~ ever, it would be feasible to consider any instant of time ti, such as ti5, as being an instant of time in the period t5, etc. We said that the focal knowledge in our formula is "x no longer knows y", £E. "x can no longer reactivate y". We could add "reactivate y either in an external (behavioral) or in an internal (mental) way". So such a characterization may contain disjunctions. It must be emphasized that our forula is assumed to charac› terize an initial concept for uninstructed students. We assume that for such students the following relationships hold -- at least as a correct first approximation: if x knows y, then x is able to reactivate y; if y is able to reactivate y, then x knows y. A sentence like the second one is used by cognitive psychologists too, as a tool for inference. But the first sen› tence needs to be elaborated further. One goal of the instruc› tion may be to convey information about notions such as "acces› sibility" in memory. In that case, the characterization of the final target concept of forgetting should contain some kind of distinction between "x no longer knows y" and "x can no longer reactivate y". We do not address this distinction in this paper. In order to validate the formula of "forgetting" we have to look at every written reponse to the question, "What is forget› ting?" for words or phrases that can be confidently taken to ex› press linguistically the existence in the student’s mind of the
J.-F. Le Ny, L. Cariteand S. Poitrenaud
206
following
items of knowledge:
1. an individual, X;
2. a mnesic content, y; (both X and y must appear in an undetermined form, some
universal quantifier);
or with
3, a relation R between x and y, which may be either 3.1. x knows y or
3.2. x can reactivate y
4.
as discussed
above;
a negation;
5, a time contrast associated
with the negation, such as:
t5), the relation R has a
5.1. at a later time (called negative value,
5.2. whereas
5.3. previously either
the relation had a positive
value, i.e.,
5.3.1. x knew y (at t3)
or
5.3.2. x could
reactivate y (at t3)
or
5.3.3. x acquired y (at t2)
(which entails 5.3.1. or 5.3.2).
These
items of the characterization are then used as the ele›
menw of a conceptual or phrases example,
schema.
For each element,
can be listed a priori as possible
students may express
or gQ more.
R may be expressed
as various forms
to recall,
x can be expressed
££ --
to recollect,
the lists are different in French. For example, as
sometimes
~
appear as some
student, etc.
of the verbs
etc.
Of course,
the undetermined
more frequent in French than one in
(someone,
somebody).
Moreover,
x may
determined name or noun, such as John,
It may not be expressed
is infinitive. If some
For
the negation as not, no longer,
to know, to remember,
English -- or as guelgu’un
a set of words
instances.
word is used,
at all if the verb form
it must alwBys refer to
the same individual who does not know at t5 but who did know at 0 .
207
Transmission of Knowledge through Discourse
The time contrast is normally expressed
time difference
ample,
pressed
ly.
"x does
not know y, x new y".
But some
In French this is best ex›
(passe
compose
the aspectual
The preceding
example
at t2 can be expressed
information conveyed the previous
with it.
also
shows
indirectly.
to lead by inference
particular statements
of the conceptual
every subject
or an event (learn›
how an event taking place
to "to
and inferences
schema
is presumably
have put in one’s
memory".
are not included
to represent
can later be extended
by ob›
of fresh data. In the present case,
may be assumed
produced
by subjects
be interpreted as due to awkwardness in expression This will not be so for more complex
The texts of 20 responses
presented
above.
undergraduate students who had received
about memory,
"forgetting"
The subjects
and in particular about forgetting. framework.
with in
were 20
no direct instruction
es were acquainted with the notion of matching,
an Artificial Intelligence
can
rather
concepts.
were analyzed by two judges
regard to the characterization of the concept the form of the schema
since
to know the crude meaning of for›
getting, any bizarre words or phrases than to ignorance.
in
Naturally, any cur›
wordings used in discourse
servation and analysis
safely
bv the
time may be either
"To have seen"
and will be treated separately.
rent set of possible elements
for instance
or plus gue parfait), as
t3 or t2, and so either a state (knowledge)
ing) is accordingly’associated
the schema
respective›
is different in French from that in English. As is
mentioned in 5.3 of the schema,
supposed
tenses,
is no longer able to remember what he had once
Of course,
past tenses
of R: for ex›
other past tense may be used as well,
or pluperfect
in, "Somebody
These
by a grammatical
instances
by use of the present and imperfect
perfect seen."
between two successive
In a sense,
The two judg›
considered
what they
in
tried to do was to simulate in their own minds what a machine
program would do when matching an input text to a predeter›
mined schema.
People
mimicking machines
best thing in the psychological Matching was directly
world!
successful
may be the worst or the
for 17 of the 20 response
208
J.-F. Le
texts: cue
in everyone
there was a word, phrase,
-- for example,
a predetermined ed above.
Some
formation;
»s. L. Cariteand S. Poitrenaud
instance
be matched to
of each item of the schema
responses
as present›
contained redundant or irrelevant in›
the rule for this aruDysis had been stated so that
any extra information was diregarded. information was therefore
problem
or morphological
a sign of tense -- that could
discarded,
Redundant or irrelevant
as there was no important
in applying this rule to these
For twc other response
most parts of the schema,
texts,
matching was succecsful
but some
problems
to the kind of problem exemplified text could
be considered
saw, heard, and experienced" committed
hear, to experience of saying,
"to
some
for
They belonged
words in the
of an element
of the sche›
"what someone
be interpreted as implying
to memory".
in the case
arose.
-- for instance,
could
deal with each particular case of the lexicon:
above:
as instances
ma only via a one-step inference "what someone
texts.
It would be possible
to
of this kind by an accommodation
mentioned,
we could
put to see,
into the list of words considered
learn at t2."
But this is a rather ad hoc
tion. A better one, allowing interpretation by inference,
be constructed
in the future;
be knowledge-based cific
information that could
no longer remembering
hardly be matched
has to
that it was necessary
This example
to aJ.low a lexical
solution:
For such
such a large lexicon
to the predeter›
Here is ilia rosponse:
bread for lunch."
overspecification.
solu›
shoul~
text contained a large amount of very spe›
mined wordings of the schema.
=
such a solution obviously
and not purely logical.
One last reponse
to
as ways
clearly
cases,
too,
we could
"Forgetting
to buy a loaf of
provides an example
of
it would be conceivable
provide the system
that the words in it could
with
be matched by
very many token words in the texts,
if not by all. But this
What CINACC would do
would be to apply a simple
woule be uneconomical special
rule:
(according
response
reSDonse
and clearly
could
in such a case
If the program fails
not be exhaustive.
to construct
a satisfactory
to determined criteria) representation of a subject’s
text, ask the sUbject: to this question?".
"Would
you please
rephrase your
209
Transmission of Knowledge through Discourse
In accordance
interpreter, some
with the present limitations of the system’s constraints have to be imposed
texts it can take as input. These (1) Reponse
length:
(2) Response
type:
concern
for the time being,
cept responses
longer than 4 lines
with questions
of the type,
deal with incorrect
or complex
of analysis
could
of the response
possibly
not ac› only
text: More
be set up in order to
response
texts.
But it is not
our main goal at present to interpret every single since
does
4 x 80 characters).
(i.e.,
"What is an I?".
and simplicity
forms
the system
as has been said, the interpreter deals
(3) The correctness
complex
on the response
the following:
response,
the restatement rule mentioned earlier can always be used.
Moreover,
the requirement made by this rule may be considered
appropriate in an instructional session. The same kind of analysis
as for "forgetting"
either to an initial, pre-session
tained during a session, similar to those
reponse,
or to a response
presented
in our two formulas
with the text presented
has been set up.
questions
schemata.
component
fore,
Some
makes
concepts
hypothetical
false
the program could
some
response
from cognitive
schema
schema
some
There›
a corresponding
from mismatches,
with. From such
diagnosis
because
to be used as a means of teaching:
identify which incorrect
matches
conceptual
conceptual
is added.
not only distinguish matches
students have acquired
rules
or a component
to enter into the system
to the correct
the latter case,
to open,
to set up false
are often misunderstood
in them is missing,
it is possible
responses
it possible
has to be made
during the session.
In addition, the evaluation of responses
"wh-"
ob›
provided an appropriate representation
Naturally, evaluation of non-initial responses in accordance
can be applied
about concept
but also,
conceptual
schema
identifications,
of what is incorrect
of presentation in the system,
of students’
in
i.e.,
in the knowledge
I, we can, by appropriate
decide
which items of in›
formation have to be transmitted to each particular student in order to improve
In this section,
his or her knowledge we described
of I.
in some
detail the third cate-
210
J.-F. Le Ny. L. Carite and S' Poitrenaud
gory of evaluation in CINACC. All three categories of evaluation involve, as is now clear from tbese descriptions, inferences by the system -- based on rules given by the conceiver -- that lead to a representation of what the student’s current state of knowledge is. In all these evaluations, two more main psycho› gical concepts are used in connection with student knowledge: first, the accessibility to tr.e student of his or her own knowledge, and second, the degree of confidence the student bas in tbis knowledge. These concepts, which are widely used in psychology, will not ce elaborated on in this paper. CONCLUSION We may now summarize how the whole system works, and what is postulated in it from a cognitive point of view. CINNA/CINACC has been devised as a tool that closely language and knowledge.
links
First, it uses language in order to crea~ knowledge. So, this knowledge bears no direct relationship to the world in the course of its acquisition. In this respect, it may be said that such knowledge, once acquired, is different from "natural" knowledge, which is directly extracted from or related to the perceptible world. Recent psychological studies on semantic memory (e.g., Anderson, 1976; Collins & Quillian, 1969; Dubois, 1982; Johnson-Laird, 1983; Klix, Hoffmann, & van der Meer. 1982; Le Ny, 1979, 1982; Norman & Bobrow, 1979; Rips. Shoben, & Smith, 1973; Rosch, 1978; Rumelhart & Ortony, 1977; Smith & Medin, 1981; Tversky, 1977) have been primarily concerned with natural terms, or related ones, i.e., with concepts or word meanings such as "dog", "table", "under", "to move", which are presumably both acquired and maintained by frequent interactions between discourse and perception of the world. It might be argued that the word-meaning II forgetting" belongs to this same category of knowledge in that it is rooted in every subject’s experience, even though not in perception proper. By contrast. a concept such as "interference" can be created only by the use of lan› guage.
211
Transmission of Knowledge through Discourse
However, in instructional discourse,
in general are always assumed
to refer to some of reference
concepts
to be about the real world, and
entity or relation in it. So whatever theory
is adopted,
be related to some
the knowledge
parts or aspects
imparted
takes into consideration
data.
or on words,
consider
cally
special
methodology,
phrases,
therefore,
which
acquired
students rely
and sentence
in or related to perception.
knowledge
to
both observational and inferential
When reading instructional texts,
are rooted
is assumed
of the world via percep›
tion. This relation is granted by scientific
on concepts
and knowledge
meanings that
This is why we cannot
by language processing
kind of knowledge.
as a radi›
However. the conditions
of its
acquisition -- reading a text -- certainly raise psychological
problems
that are somewhat different from those
recent years crete
in cognitive
knowledge. system
A second
and language are linked in a computer›
aspect
of the relationship between language and
in CINACC is worth emphasizing.
assessing
their knowledge.
ments made by the students in repsonse theoretical
The system
to questions
evaluation refers In summary.
views about the
CINACC is clearly (Newell,
both on knowledge
a knowledge-based
in
use the
1982). This means that it has to rely
and on knowledge
Firstly, it requires
between two such experts.
Secondly.
managed by an expert
interactive communication it provides
data that allow interactive communication system)
system
specialists
provided by an expert in the domain or subdo›
main under consideration
in cognition.
in which
to "gaining external knowledge
in which Artificial Intelligence
word knowledge
state›
In the latter part of this paper we
and semi-technical
specifically
of internal knowledge".
uses
as a way of
interpretation and evaluation programs of the system,
the sense
is one
such as CINACC.
knowledge
have presented
in
with regard to more con›
This method of acquisition of knowledge
way in which knowledge assisted
psychology
discussed
procedures
between itself
and
(the
and a student, that is, an indirect interactive commu-
212
I.-F Le Ny, L. Came and S. Poitrenaud
nication between the expert in the domain, the expert in cogni› tion and the student. To make this possible, the separate ap› proaches of Cognitive Psychology and Artificial Intelligence are conjoined from a point of view that we believe may illus› trate a more comprehensive approach to Cognitive Science than is permitted by either discipline on its own. In doing so we try to make the system CINACC both a practical tool for compu› ter-assisted instruction and a research tool for the investiga› tion of the interrelations between knowledge and language. Acknowledgements The first part of the entire research project was supported by a grant from the Agence De l’Informatique (France). Valuable help was provided by the Service d’Etudes pour Ie Developpement et l’Animation (Paris). REFERENCES Anderson, J. R. (1976). Language, memory and thought. Hillsdale, NJ: Erlbaum. Collins, A. M., & Quillian, M. R. (1969). Retrieval time from semantic memory. Journal of Verbal Learning and Verbal Beha› vior, ~, 240-247. Denhiere, G. (1982). Do we really mean schemata? In J. F. Le Ny & W. Kintsch (Eds.), Language and comprehension (pp. 219› 238). Amsterdam: North-Holland. Denhiere, G., & Le Ny, J. F. (1980). Relative importance of meaningful units in comprehension and recall of narratives by children and adults. Poetics, ~, 147-161. Denis, M. Les images mentales (1979). Paris: Presses Universi› taires de France. Denis, M., & Le Ny, J. F. (1984). Centering on figurative fea› tures during the comprehension of sentences describing scenes. Paris: ERA 235. Mimeo No. 220. Dijk, T. van, & Kintsch, W. (1983). Strategies of discourse comprehension. New York: Academic Press. Dubois, D. (1982). The nature of mental representation and sen› tence comprehension. In F. Klix, J. Hoffmann, & E. van der Meer (Eds.), Cognitive research in psychology. Berlin: Verlag der Wissenschaften. -Fodor, J. A. (1975). The language of thought. Hassocks, Sussex: Harvester Press. Johnson-Laird, P. N. (1983). Mental models. Cambridge, England: Cambridge University Press.
Transmission of Knowledge through Discourse
213
Kintsch, W., & Dijk, T.A. van (1978). Toward a model of text comprehension and production. Psychological Review, ~, 363› 394. Klix, F., Hoffmann, J., & Meer, E. van der (1982). The storage of concepts and their cognitive application. In J. F. Le Ny & W. Kintsch (Eds.), Language and comprehension. Amsterdam: North-Holland. Le Ny, J. F. (1979). La semantique psychologiguo. Paris: Presses Universitaires de France. Le Ny, J. F. (1982). Knowledge, meaning, and representation: Some current problems. In F. Klix, J. Hoffmann, & E. van der Meer (Eds.), Cognitive research in psychology. Berlin: Verlag der Wissenschaften. Le Ny, J. F., & Denhiere, C. (1982). Profile of CINNA: Con› struction of individualized texts. Text, ~, 193-210. Newell, A. (1982) The knowledge level. Artificial Intelligence, l.§., 87-127. Norman, D. A., & Bobrow, D. G. (1979). Descriptions: An inter› mediate stage in memory retrieval. Cognitive Psychology, 11, 107-123. Rips, L. J., Shoben, E. J., & Smith, E. E. (1973). Semantic dis› tance and the verification of semantic relations. Journal of Verbal Learning and Verbal Behavior, ~, 1-20. -Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B. Lloyd (Eds.), Cognition and categorization. Hillsdale, NJ: Erlbaum. Rumelhart, D. E., & Ortony, A. (1977). The representation of knowledge in memory. In R. C. Anderson, R. J. Spiro, & W. E. Montague (Eds.), Schooling and the acquisition of knowledge. Hillsdale, NJ: Erlbaum. Smith, E. E., & Medin, D. L. (1981). Categories and concepts. Cambridge, MA: HRrvard University Press. Tversky, A. (1977). Features of similarity. Psychological Review, g, 327-352. -Vygotsky, L. S. (1982). Thought and language. Cambridge, MA: M.I.T. Press. Woods, W. A. (1975). Syntax, semantics and speech. In D. R. Reddy (Ed.), Natural language processing (pp. 345-400). Englewood Cliffs, NJ: Prentice-Hall.