Concept Formation Rules: An Executable Cognitive
Model of Knowledge Construction
Veronica Dahl
1
and Kimberly Voll
1
1
Simon Fraser University
8888 University Dr. Burnaby, BC,
Canada, V5A 1S6
Abstract. We develop a cognitive model of knowledge construction inspired
by constructivist theory as well as by recent natural language processing meth-
odologies. Our generalization of these methodologies into a new and directly
executable cognitive paradigm, concept formation, has been tested in non-
trivial applications with very promising results. Our system accommodates user
definition of properties between concepts as well as user commands to relax
their enforcement. It also accepts concept formation from concepts that violate
principles that have been declared as relaxable, and produces a list of satisfied
properties and a list of violated properties as a side effect of the normal applica-
tion of the rule.
1 Introduction
Constructivism is the idea that we construct our own world rather than it being de-
termined by an outside reality
(see [1] for an overview). Since the idea of constructing
solutions or proofs is at the heart of both traditional computing science and logic
programming, constructivism can constitute a natural bridge between the cognitive
and computing sciences.
In the field of learning, constructivist theories such as those introduced by [2] view
learning as the construction of new concepts from previously known concepts. Dur-
ing this process, “the learner selects and transforms information, constructs hypothe-
ses, and makes decisions, relying on a cognitive structure to do so”. The formation of
new concepts from known ones is also central to cognitive logic, and also to logic
programming and its recent sub-paradigm, Constraint Handling Rules (CHR) [3].
We exploit these natural connections to develop a cognitive model of knowledge
construction that can be directly executed through a specialized system implemented
using CHRs. In our model, information is selected automatically by the system as a
side effect of searching through applicable CHR rules. It is automatically trans-
formed, or simply augmented, when a rule triggers. Hypotheses are then made in the
form of assumptions and decisions follow from the normal operation of the rules.
The necessary cognitive structure is provided by those properties that must be satis-
fied based on the concepts a given rule is trying to relate. Moreover, some latitude is
provided so that the user can declare under what circumstances a given property or
Dahl V. and Voll K. (2004).
Concept Formation Rules: An Executable Cognitive Model of Knowledge Construction.
In Proceedings of the 1st International Workshop on Natural Language Understanding and Cognitive Science, pages 29-36
DOI: 10.5220/0002685700290036
Copyright
c
SciTePress
properties can be relaxed rather than having to satisfy all properties that have been
defined as necessary. Concepts formed under relaxed properties result in output
which signals not only what concepts were formed but also which of the properties
associated with that concept’s construction were satisfied and which were not. This
allows us human-like flexibility while still maintaining direct executability.
In this paper, Section 2 presents our cognitive model while Section 3 provides the
preliminary background. In Section 4 we discuss our methodological approach to
implementing our model, and present two applications in Section 5. Finally, Section
6 presents our concluding remarks.
2 A Computational Metaphor for the Mind
We can imagine the human mind as an evolving store of knowledge, which con-
stantly updates itself from new information built from previous information through
some kind of reasoning.
In this model, the problems, events, or feelings, et cetera, of which we are con-
sciously aware at any given time trigger a partly, or wholly, unconscious search
through the knowledge store for those pieces of information that relate to the prob-
lem. Once found, these pieces of information permit the inference of new knowledge.
For instance, a rule that is appropriate for modeling the formation of the new knowl-
edge might look roughly as follows:
c1, c2, …cn Æ newc .
(1)
Here ci and newc are concepts expressed as logic atoms. We call these rules concept
formation rules.
Occasionally, we may need to test some condition before being able to form a new
concept. For example, consider the knowledge that if the speed limit is X and our
car’s velocity is Y, then we will likely receive a speeding ticket if Y is greater than X.
If we also know that we are traveling at 100km/hour and the speed limit is 50km/
hour, then we might form the new concept that it is likely we will receive a speeding
ticket. In other words, using a primitive (i.e., system-defined) unary atom test, whose
argument contains the test to be performed, we can model the above concept-
formation rule as the following:
speed_limit(X), current_speed(Y) => test(Y>X), speeding_ticket(likely) .
(2)
We next examine each feature that our concept-formation paradigm must have in
order to a) constitute an adequate, constructivist metaphor for the workings of the
human mind, b) provide some of the flexibility exhibited in human reasoning, and c)
be directly executable.
2.1 Information Selection
Most of the rules in our knowledge store will be irrelevant to the situation at hand, so
the first problem in our model, and in constructivism in general, is the problem of
29
selecting the appropriate information in order to form new concept(s) (which are
actually relevant to the problem). We solve this by using a system that constantly
searches for any information that may match the new information (i.e., the initial
problem’s concepts, or the concepts formed while attempting a solution) and conse-
quently triggers those rules whose left-hand side at least partially matches some of the
new information.
If we consider our previous example, then as soon as the new information that the
speed limit is 50km/hr and the current speed is 100km/hr arrives in the knowledge
store, a search through that store is initiated which ultimately yields rule (2), whose
left hand side will be matched with unifier {X=50, Y=100}. The unifier will apply to
the rest of the rule as well, and a test to see whether 100 > 50 will be performed.
Since the test will succeed, the new information speeding_ticket(likely) will be added
to the knowledge store.
2.2 Flexible Concept Formation
Binary logic is notoriously insufficient for modeling the human mind. The main al-
ternatives are based either on probabilistic or fuzzy reasoning. As pointed out in [4],
although probability theory is appropriate for measuring randomness of information,
it is inappropriate for measuring the meaning of information. For many applications,
measuring vagueness (e.g. through a membership function that measures the “degree”
to which an element belongs to a set) is more important than measuring randomness.
While the fuzzy set approach has successfully overcome some of the problems
with probabilistic reasoning, it relies on fairly detailed comparative knowledge of the
domain being described. Not only must one say that Tom is tall, for instance, but one
must also state specifically to what degree.
Here we propose a simpler model based on properties between concepts. Vague-
ness can be expressed by relaxing the properties between concepts in accordance with
a user’s criteria. The criteria can be flexibly and modularly adjusted for experimental
purposes while still maintaining direct executability.
For our model described so far, we need only introduce the following enhance-
ment. That is, rather than inflexibly allowing a concept to be formed if a test succeeds
while disallowing its formation if that test fails, we instead denote those tests which
we wish to make flexible as properties. Properties are like any other test, except that
their failure does not necessarily result in the failure of the rule itself. Rather, the
concept will still be formed and two lists will be associated with it: a list of the prop-
erties that the concept satisfies, S, and a list of those that it violates, V. This allows us
to construct possibly incorrect or incomplete concepts while collecting the informa-
tion as to what way they are not totally justified. The user then has all the information
pertaining to the construction of a particular concept and can interpret these results in
a much more informed, holistic way. This is in contrast to the blind computation of
the degree of randomness or vagueness from those assigned a priori to each individ-
ual piece of the reasoning puzzle.
For instance, if we want to accept incorrect sentences in a parsing system we might
designate as properties all tests on rule applicability that would correspond to correct
parses, and then choose some to relax (for example, the number agreement property).
30
This might prove useful in a second-language tutoring system which allows the user
to make certain types of mistakes, while pointing out the reason why those are mis-
takes (i.e. which properties are not being satisfied).
2.3 Transforming the Knowledge Store
Just as rule selection follows automatically from the system’s normal search for ap-
plicable rules, the transformation of the knowledge store is achieved as a side effect
of the system’s normal procedure of applying a concept formation rule.
In some cases the newly formed concepts will need to coexist with the concepts
that participated in their formation, such as when forming a concept based on some
transitive property. In other cases, however, it will be more appropriate for the new
concepts to replace the concepts that led to it. For instance, imagine a rule which uses
the concept that a teacher is free at time T and that a parent is free at the same time T
in order to add the concept that a meeting between these two individuals should be
scheduled for time T. Obviously, once a meeting at time T has been arranged between
them, they can no longer be considered as free at that time.
Yet another type of rule is needed to remove redundant concepts such as “John
must meet Alice at six” and “Alice must meet John at six”.
In this paper we primarily use the first type of rule, but all three types maintain
computational counterparts within our model.
2.4 Making Decisions
It is also interesting to note that decisions also follow from the normal operation of
the rules. That is, depending on which of the currently considered concepts matches
which rule, the decision on the appropriate rule to trigger is made, determining which
new concepts will be formed. Newly added concepts then help further trigger rules,
and so on until a solution is found or no new, useful, concepts can be constructed.
2.6 Cognitive Structure
Cognitive structure (such as schemas or mental models) provides meaning and or-
ganization to experiences and allows individuals to draw conclusions beyond the
information actually given.
In our model, cognitive structure is provided by the rules themselves. The proper-
ties that are required between the concepts a given rule is trying to relate are particu-
larly important, both for the formation of new concepts and achieving human-like
flexibility
31
3 Methodological Background: CHR
Imagine an evolving store of knowledge, which represents concepts as logic atoms
(called “constraints” in the CHR literature), and uses rules of the following form to
construct new pieces of knowledge from previously known ones:
Head Guard | Body .
(3)
Here head and Body are conjunctions of atoms and Guard is a test constructed from
built-in or system-defined predicates; the variables in Guard and Body occur also in
Head. If the Guard is the constant “true” then it is omitted together with the vertical
bar. Its logical meaning is the formula (Guard Æ (Head Æ Body)) and the meaning
of a program is given by conjunction.
CHR rules are then interpreted as rewrite rules over such stores of knowledge. We
use the CHR version that is an extension of SICStus Prolog and standard Prolog nota-
tion (i.e. capital letters for variables, et cetera).
CHR have rules of three types: propagation rules, which add new information to
the store, simplification rules, which remove old information in favour of new, and
simpagation rules, which remove redundant information.
4 Concept Formation Programs/Grammars
It is clear that CHRs can provide most of the features needed for a computational
incarnation of our concept formation rules model:
- Information selection is a side effect of the CHR engine’s search over ap-
plicable rules.
- The transformation of information takes place automatically when a rule
triggers, in the way sanctioned by the rule (namely, propagation rules update
the knowledge store with the concept newly formed).
- Hypotheses can be made through assumptions.
In addition, we need to provide flexible cognitive structure through relaxable, di-
rectly executable properties between concepts. This is achieved through concept for-
mation rules, which have the same general form as CHR rules except that the guard
may include any number of property calls for properties that have been defined by
the user. These calls are automatically handled by the system, provided that the user
defines the properties as follows:
a) A property must be named and defined through the binary predicate prop/2,
whose first argument is the property’s name and whose second argument is the list of
arguments that are involved in checking the property and signaling the results. For
instance, in a grammar that needs to check for number agreement between the deter-
miner, noun and verb, and produce either the number of all three, or an indication of
mismatch, we can choose the name “agreement” for the property, and define it as
follows:
prop(agreement,[Ndet,Nn,Nv,N]):- Ndet=Nn,
Nn=Nv, !,
32
N=Nv.
prop(agreement,[Ndet,Nn,Nv],mismatch).
b) The acceptability of a thus-defined property must be checked within the rule
concerned through the binary, system-predicate acceptable/2. The first argument of
this predicate takes the prop atom just described, and the second argument evaluates
to true, false, or a continuous value indicating the degree of acceptability. For our
example, we write:
determiner(Ndet,P1,P2),noun(Nn,P2,P3),v(Nv,P3,P)=>
acceptable(prop(agreement,Ndet,Nn,Nv,N),B)
| sentence(N,P1,P).
N.B. the last two arguments of each atom, Pi, are explicit word boundaries. If left implicit, we instead
have Concept Formation Grammars, which must run under CHRG rather than CHR.
c) In order to relax a property named N (i.e. to allow the derivation of concepts that
require it but for which it is not satisfied), we simply write the following:
relax(N).
Degrees of acceptability can be defined through a binary version of the relaxing
primitive, where L is the prop atom with all its arguments and D is a measure of ac-
ceptability:
relax(L,D).
A list of satisfied and violated properties, together with the degree of violation
when appropriate, will be output for each property defined in a given CF program.
5 Two Case Studies
5.1 Scheduled Meetings
Consider the problem of scheduling meeting events between two groups of people
given the availability of each person and the specific individuals that must meet.
Regular scheduling software systems typically work in stages. The schedule is
partially assigned until it reaches a meeting that it cannot schedule. By re-shuffling
the other meetings that have already been scheduled, it is then often possible to open
up matching timeslots for the individuals in question.
Our concept formation framework allows us to produce a higher-level solution in a
concise and effective manner. We first initialize the knowledge store with the pre-
liminary concepts, namely, the available intervals of each of the participants and the
requirement of who must meet with whom. For example:
33
start:- available(snape,1,3), available(dumbledore,2,4),
available(potter,1,3), available(grange,1,5),
req(potter,dumbledore), req(potter,snape),
req(snape,grange), req(grange,dumbledore).
We can form the concept of free timeslots from the interval definitions of availability:
available(Person,T1,T) <=> T1 = T | true.
available(Person,T1,T) <=> T2 is T1+1, |
free_slot(Person,T1,T2),
available(Person,T2,T).
To schedule all required meetings we need just one rule:
free_slot(Prof,Start,End), free_slot(Par,Start,End),
req(Par,Prof) <=> meet(Prof,Par,Start,End).
To schedule all required meetings we then need only one rule:
free_slot(Prof,Start,End), free_slot(Par,Start,End),
req(Par,Prof) <=> meet(Prof,Par,Start,End).
This simplification rule says that provided that Prof and Par have the same free
timeslot, and that Par is required to meet Prof, then this information (i.e. the free
timeslots of each and the requirement to meet) will be deleted from the knowledge
store and replaced by the concept that Prof and Par must meet in that timeslot. (No-
tice that no parent will be assigned two intervals with the same professor or vice versa
since the parent’s requirement to meet that professor is removed by our simplification
rule.)
Suppose, for example, the school director wishes to see parents at random,
whether they need to meet him or not. We simply define req/2 as a property rather
than as an intervening concept, and relax it for the individual in question:
req(potter,dumbledore). req(potter,snape).
req(snape,grange). req(grange,dumbledore).
free_slot(Prof,Start,End),free_slot(Par,Start,End) <=>
acceptable(req(Par,Prof),B) | meet(Prof,Par,Start,End).
relax(req(Par,dumbledore)).
The last clause relaxes the requirement property only for Professor Dumbledore.
Duplicate meetings can be avoided using the simpagation rule:
req(X,Y) \ req(Y,X) <=> true .
(4)
This rule removes the second requirement, which is symmetric to the first if both
appear in the knowledge store.
34
Notice that any unsatisfied requirements will print automatically, since their req/2
constraint remains in the knowledge store at the end of the computation.
Similarly, free timeslots that have not been used in the solution of the problem are
reflected in leftover constraint list free/3, which also prints out automatically. Further
utilities to organize all this information in various useful ways can be provided ac-
cording to the specific application.
5.2 Error Detection/Correction
Notice that by simply treating as properties those conditions whose failure we wish
identify but tolerate, detected errors will now appear in the list of violated properties
that is automatically produced. Consider the parsing example again, if number agree-
ment has been relaxed, then in a language tutoring system a sentence disagreeing in
number will be accepted but the error will be indicated as a side effect of parsing.
Error correction is now not far away. All that remains is to decide on a heuristic.
For example, in parsing number agreement consult the user as to which of the two
mismatched elements should change in number, or alternatively, take the most com-
mon number if one exists (e.g. changing “boys” into “boy” if everything else is singu-
lar). Once the heuristic is chosen, it simply remains to effect the change, since our
concept formation rules have already done the diagnosis part of the work.
In joint work with Kimberly Voll, we are presently investigating the use of CF
rules to aid in the interpretation of free-text reports, particularly within the medical
domain. We treat semantic correctness as a property, so that nonsensical but gram-
matical sentences resulting from mistakes in input, such as those occurring through
the use of speech recognition systems, can be detected and corrected. This is an im-
portant aid because while such automatic transcription systems have accuracy rates as
high as 98% (which is a quite encouraging number in most settings), such rates can
have dangerous, even fatal consequences in the medical setting. Semantic compati-
bility between parts of a sentence is checked with the aid of semantic type hierarchies
in the case of this particular application.
6 Conclusions
We have shown a powerful yet relatively simple cognitive model for concept forma-
tion inspired by constructivism. The methodology we present for implementing di-
rectly executable concept formation paradigms generalizes parsing methods we have
developed specifically for natural language processing
[6]. We have discussed our
methodology for a directly executable, CHR-based rendition of this model, and ex-
emplified its promise through two examples: (time) resource allocation, and error
detection and correction. Other approaches need considerable more apparatus to solve
the same problem (for example, [7] uses a fairly involved abductive model based on
CHRGs for error detection and correction in string processing problem domains). We
are also currently working on applying concept formation to the problem of identify-
ing biomarkers in cancer diagnosis systems
[8].
35
Concept formation rules, as we have seen, have a grammatical counterpart (CFG),
and are applicable to many other AI and cognitive problems as well, most notably,
those involving the need to reason with incomplete or incorrect concepts.
7 Acknowledgements
Thanks are due to Alma Barranco-Mendoza for her help with typesetting, and to
her and the anonymous referees for their useful comments on this article’s first draft.
This work was supported in part by NSERC grant 31-611024 and the NSERC PGS
program. The linguistic work that inspired our cognitive model was developed with
Phillipe Blache during Veronica Dahl’s visit to Universite de Provence, invited by
CNRS as Chercheur Etranger.
References
1. Kearsley, G.: Explorations in Learning and Instruction: The Theory into Practice Database,
http://tip.psychology.org (2003)
2. Bruner, J.: Going Beyond the Information Given. New York: Norton. (1973)
3. Fruhwirth, T.: Theory and Practice of Constraint Handling Rules, J. Logic Programming,
Vol. 37 (1-3), 95—138 (1998)
4. Zadeh, L.,: Commonsense knowledge representation based on fuzzy logic, Computer, 16
(1983)
5. Christiansen, H.: CHR grammars. In Theory and Practice of Logic Programming special
issue on Constraint Handling Rules (2004)
6. Dahl, V. and Blache, P.: Directly Executable Constraint Based Grammars. TR, Univ. de
Provence (2004)
7. Christiansen, H. and Dahl, V.: Logic Grammar for Diagnosis and Repair. In’tl J. of AI
Tools, Vol. 12, No. 3, 227—248 (2003)
8. Barranco-Mendoza, A., Persaud, D. R., and Dahl, V.: A property-based model for lung
cancer diagnosis, Poster Proceedings of RECOMB 2004 (2004)
36