NEW CONCEPT DEVELOPMENT MODEL

Explorative Study of its Usability in Intelligence Augmentation

Irena Drašković, Maikel Couwenberg

RUNMC, Radboud University, Nijmegen, The Netherlands

Janos Sarbo

ICIS, Radboud University, Nijmegen, The Netherlands

Keywords: Intelligence Augmentation, Problem Solving, Knowledge Representation, Conceptual Representation and

Interpretation, Empirical Study.

Abstract: Intelligence augmentation (IA) is used in for improving the efficiency of human problem solving by making

use of computer support. To make it work, we need a ‘human-compatible’ formal model of knowledge

representation. Here we report on empirical test of one such model. The results show that the model can be

used to analyse concept formation in human problem solving. This means that the model is congruent with

human information processing system. Being a formal model, it can be used in a wide range of IA-

applications such as computer supported tutoring systems and human-computer interfaces.

1 INTRODUCTION

The main goal of research on Intelligence

Augmentation (IA) is to develop tools and methods

to support and improve the effectiveness of human

intelligence in solving complex problems. However,

on one side, we do not have a full understanding of

human cognition and, on the other side, we do not

yet have at our disposal complex parsers to interface

with different informal representations employed in

human processing. According to Douglas C.

Engelbart (1962), the aim of IA tools is to optimize

sensory, motor, and cognitive human capacities. To

achieve better efficiency and effectiveness of

computer systems for knowledge representation,

‘human-compatible’ formal models of knowledge

representation are needed. In the present paper, we

report on an empirical study exploring the features

of such a formal model of human information

processing.

Recent empirical studies on human processing

suggest that e.g., syntactic and semantic information

processing is quasi-simultaneous (Hagoort, Hald,

Bastiaansen, & Petersson, 2004). This renders

translation between these representational formats

improbable. In addition, it underscores the

importance of developing a uniform knowledge

representation model in other modalities as well. As

far as we know, current computational models of

knowledge representation do not meet the demand of

a uniform representational format and interpretation

processes across different kinds and modalities of

information (Shadbolt, Hall & Berners-Lee, 2006).

Rather, those models use different representational

formats for the different knowledge domains. From

the computation point of view, translation between

different formats is inefficient. Moreover, the

development of translator programs is non-trivial

(Aho & Ullman. 1972).

The present paper introduces a uniform

knowledge representation model together with a

preliminary testing of its validity. So far, the

usability of this model has been tested in the

following knowledge domains: (morpho-)syntax,

naïve logic, reasoning, and mathematics (Sarbo,

Farkas & Bremen, 2006). In the present paper we

explore correspondence between the predicted and

the observed output, in different information

processing stages specified by our model.

288

Draškovi

c I., Couwenberg M. and Sarbo J. (2010).

NEW CONCEPT DEVELOPMENT MODEL - Explorative Study of its Usability in Intelligence Augmentation.

In Proceedings of the 2nd International Conference on Computer Supported Education, pages 288-293

DOI: 10.5220/0002797602880293

 SciTePress

1.1 Model Description

1.1.1 Theoretical Background

In developing our model, we considered

developmental aspects of human cognition.

According to the Social Constructivist model of

Piaget, learning and cognitive processes are adaptive

‘tools’ aiding human interaction with the

surrounding. Focusing on cognitive development,

Piaget defined different stages in child development

(Rigter, 1996):

• Sensory-motor Phase. Objects and object

characteristics are learned and recognized

through perception and motor manipulation; this

knowledge is stored as concepts or abstract

representations of object characteristics (e.g.,

sweet).

• Pre-operational Phase. Percepts are

explained through reasoning. It is assumed that

the most salient concept properties are included

into interpretations.

• Concrete Operational Phase. Learning

that different points of view are possible.

• Formal Operational Phase. Reasoning

without preceding perception, development of

abstract reasoning and hypothesizing (Delfos,

2000).

J.S. Bruner (e.g., 1966) describes cognition as a

cyclic process including hypothesizing, informative,

and confirmative phases. Similar to Piaget, Bruner

distinguishes 3 phases in child cognitive

development:

• Enactive Phase. Learning through physical

contact with the surrounding.

• Iconic Phase. Decisions based on sensory-

motor perception.

• Symbolic Phase. Meaning construction

through symbolic interpretation of information.

We will assume that a ‘fully developed’ cognitive

system incorporates an amalgam of processes and

operations specified by Piaget and Bruner and that

these are effective in concept formation. Our model

extends these notions and builds a bridge towards

IA.

1.1.2 The Model

In Farkas & Sarbo (2000) and, more recently, in

Breemen & Sarbo (2009), a novel formal

conceptualization model is presented (see Figure 1).

Conceptualisation is described through 3 processing

stages implementing 3 kinds of operations, ranging

from perceptual analysis (analysis of sensory input)

to meaningful response, such as predication or motor

response. For each stage, input and output are

specified alongside the concept types. The model is

interactive in that it takes into account both the

internal context as well as the stimulus

characteristics.

Stage 1- PERCEPTION: Sensory stimulus triggers

perceptual analysis and sorting; further processing

proceeds in two processing paths:

1. A, B: processing of the ‘raw’ perceptual data (A –

objects in focus, B – events in the world or model

associated with these objects); a distinction is made

between the background and the stimulus;

2. ¬A, ¬B (not A, not B): activation of knowledge

on both the objects in focus and the associated

events (which are not(objects), and not(events) by

themselves).

Output of the sorting operation performed in this

stage is identification of the important components

in the input and activation of knowledge on these

components. This stage is a ‘transition’ from the

‘world’ to a ‘model’ of the world. Hereby the notion

of contrast is important: perception is ‘perception of

contrast/change’.

Stage 2 - CONTEXTUALIZATION: Output from

stage 1 is abstracted into types through matching

with existing prototype concepts (Smith, Osherson,

Rips & Keane, 1988). Contextualization

incorporates: lexical access of the perceived

qualities (qualia) as types independent from each

other; completion of concept types (qualities) from

the previous stage by using internal context or

knowledge of the world per quality; matching of the

input with internal context. By sufficient

correspondence and context to relate the 2 types of

qualities with each other, the next stage is entered.

This stage is comparable to lexical access and

semantic interpretation (Margolis & Laurence,

1999).

Contextualizatio

Predication

Perception

[A]

[B]

A is B

[~A,~B]

[A,~B] [B,~A]

[A,B, ~A,~B]

Figure 1: Model of development of concept types.

Stage 3 – PREDICATION: Complements from the

two processing paths are related to each other. The

most plausible relation between the qualities is

NEW CONCEPT DEVELOPMENT MODEL - Explorative Study of its Usability in Intelligence Augmentation

289

established and the qualities are put together into a

proposition expressing a hypothesis on the state of

the world. This stage includes testing of the

hypothesis with relevant sensory information

whereupon a new cycle may start. If the hypothesis

is disconfirmed, either new complementation

context is searched for, or other focus is taken (cf.

polysemy - ‘bank’).

It is plausible to assume that with complex problems

the above outlined conceptualization process is

recursively used whereby the propositions formed at

the end of one cycle serve as input for the next

cycle. Per cycle, one proposition is generated. A

cycle is delimited by identification (naming) of a

relation (e.g., ‘Square A is larger than square B’).

The process is goal driven with the goal being the

formulation of a proposition. In solving a problem,

the number of embedded analyses (cf. recursion) can

be affected by three parameters:

1. What is in focus (always a contiguous segment of

input qualities).

2. Input complexity (number of propositions that

exhaustively describe a phenomenon in focus).

3. Internal context (relevant knowledge of the

world).

These are the sources of inter- and intra variability in

interpretation. By a well defined problem (see Plato,

380 B.C.) with a generally accepted solution, it is

possible to determine in advance the goal governing

the entire conceptualization process. Exploiting the

thinking-out-loud method in the process of solving a

complex mathematical problem, it is possible to

gather verbal reports containing utterances reflecting

the interpretation process. Utterances can be coded

as types of concepts. The degree of match with the

concept types specified by our model can be

determined.

1.2 Research Question and Hypothesis

According to our model, concepts are formed

through 3 development stages. In each stage

different types of concepts or signs are generated.

We assume that the stages are sequentially ordered.

Advance to the next stage is determined by

completion of the previous stage. Generated

concepts are expected to be classifiable into concept

types specified by our model. The following

research question will be investigated: Can verbal

utterances produced during the solution of an

abstract task be classified into concept types

specified by our model?

Hypothesis: Our model specifies 3 processing

stages: perception, contextualization, and

predication, each having specific input and output.

The model assumes an ordering of the processing

stages; output of the earlier stages serves as input for

the later stages. We assume that knowing the output

i.e., verbal reports, can be used to infer the

underlying processing stages. In coding the verbal

utterances into concept types specified by our

model, we use the terminology from a Peircean

interpretation of our model (Peirce, 1931) as shown

in Figure 2 (see also, Breemen & Sarbo, 2009).

Below, these are given in small caps.

Stage 1 – Perception:

- signal or external trigger as qualia (

QUALISIGN);

- spatial and temporal localization of the signal (

ICON,

SINSIGN

);

- knowledge on the signal (INDEX).

Stage 2 – Contextualization:

- prototype knowledge on A (

RHEME);

- prototype knowledge on B (LEGISIGN);

- prototypes in context (

DICENT, SYMBOL).

Stage 3 – Predication output:

- concepts in a relation (

ARGUMENT).

In order to test the hypothesis, we made use of a

complex task and a problem which is largely

unknown to the participants but which they should

be able to solve according to their developmental

stage, prior knowledge and familiarity with similar

tasks. In order to minimize variability in prior

knowledge a homogeneous group of participants

will be selected, namely primary school grade 8

pupils (age: 11-12 years). These pupils are normally

not familiar with the specific problem (see section

2.2) and the specific mathematical reasoning needed

to solve this problem. However, they do possess

sufficient knowledge to be able to perform the task.

The hypothesis was tested in an experiment

described below. Dependent variable is percentage

of concepts produced in order congruent with our

model.

Figure 2: Concept type in problem solving.

CSEDU 2010 - 2nd International Conference on Computer Supported Education

290

2 METHOD

To test our hypothesis we made use of a method in

which participants solve a difficult geometrical

problem while thinking-out loud. Their

verbalizations were recorded and transcribed.

Subsequently, verbal utterances were coded into the

above specified nine output concepts. The use and

the order of the three above described processing

stages were determined on the basis of the prevailing

‘output concepts’. This way, we were able to

determine whether the observed conceptualization

unfolds according to the stages specified by our

model. This method is rather coarse in that

verbalizations need not be entirely synchronous with

the actual cognitive processing. We assume that the

nine types of output concepts are ‘tied’ to the

respective processing stages of Perception,

Contextualization and Predication (e.g., qualisign,

‘icon’ etc. – Perception; ‘rheme’, ‘dicent’, etc. –

Contextualization).

2.1 Participants

Twenty-eight 8th graders from the primary school in

Nuland, The Netherlands, took part in this

experiment. Age range was 11-12 years. The

participants were assumed not to be familiar with the

problem since it is not a part of their Math course.

This was confirmed by the teachers who reported

that knowledge directly needed to solve this problem

individually has not been acquired. This sample was

chosen because many IA applications are targeting

the respective population. Regarding their cognitive

development, 8th graders are similar to the adults

(Delfos, 2000).

2.2 Task

Participants were presented a problem based on

Plato’s Meno (Plato, 1871). They were shown a

picture of a square and were to find out how to

determine the length of the sides of another square

which is half as large as the first one. We chose this

problem because its solution is straight forward, as

outlined in (Magnani, 2001). At the same time the

problem is complex enough to elicit sufficient

amount of verbal utterances to be analyzed, as

determined in a pilot study involving two

participants. The participants rated the problem as

difficult although at their school level they have

already learned to compute the area of geometrical

figures which is necessary to solve this problem.

2.3 Procedure and Materials

The experiment was conducted using a standard

protocol. All sessions were videotaped. The time

intervals needed to solve the problem were

registered by the experimenter using a stop watch.

The setting was an empty classroom; a familiar work

surrounding for 8th graders. The experiment was

conducted individually. The experimenter was

seated in an L-setting with respect to the pupil in

order to avoid a suggestion of a ‘leadership role’ to

the experimenter since this may affect the pupils’

level of commitment to solving the problem. The

experiment was conducted during regular school

time. The experimenter was instructed not to

interfere with the process of solving the problem

unless this is indicated in the protocol. Each session

started with experimenter giving an instruction about

the task and the procedure. The recordings contained

on average 75 verbal utterances.

Instruction: “First of all you will receive a card

with a drawing on it. The drawing ‘represents’ a

geometrical problem. Your task is to uncover the

problem and to find its solution. While doing this I

would like you to say loudly everything you’re

thinking about this problem.”

This is called ‘thinking-out-loud’ method.

Subsequently, the participants were handed over the

card with the drawing representing the Meno

problem (see Figure 3) and the session started. It

was determined in advance in which situations the

experimenter will interfere and how: If a participant

was stuck with a (part of) the problem

(operationalized as inactive for 20 seconds) or if

he/she made a mistake, the experimenter prompted

him/her to try again and solve the problem or to try

and correct the error. Few types of errors were

anticipated upon which were already described in

Plato’s Meno and which also occurred in the pilot

study. Additional material was developed to

facilitate the problem solving by providing clues to

shift the participants’ focus in problem

interpretation. This material was provided if needed

in 3 different orders assigned randomly. An

illustration of additional material is given in Figure

NEW CONCEPT DEVELOPMENT MODEL - Explorative Study of its Usability in Intelligence Augmentation

291

Figure 3: On the left side of the figure participants see a

square for which the method of computing the length of

the sides has to be determined. In those cases in which

participants were stuck with the problem, the square on the

right side of this figure was provided showing an

alternative way of slicing the square which allows for

refocusing and advancement to the solution.

Coding Procedure. All transcribed verbal

utterances were first assessed for their contribution

to the solution of the Meno problem. Two kinds of

codes were assigned: 1. contributes to the solution

of the problem; and 2. ‘side-tracking’ or ‘errors’ like

wrong perception/representation/interpretation of the

problem, wrong assumptions, and logical errors. For

the former kind of utterances a coding system was

developed with types of concepts and examples

specified.

Examples coding of the verbal utterances:

1. Prototype knowledge on a stimulus (

RHEME) –

“like a square.”

2. Prototypes in context (

DICENT, SYMBOL) - “The

square on the left side..”

3. Concepts in a relation (

ARGUMENT) – “This

square is twice as big as the other one.”

In order to validate the coding system two experts

independently coded a sample of verbal protocols.

The degree in which the conceptualization process in

solving the Meno problem is congruent with the

conceptualization process as specified by our model

was determined on the basis of prevalence of

‘correctly’ formed ‘argument’ concept types i.e.,

‘arguments’ preceded by concept types from any of

the preceding processing stages in order specified by

our model.

2.4 Analyses

The inter-rater reliability of the coding system was

determined using Cohen’s Kappa, and means and

standard deviations (SD) were computed using

Statistical Package for the Social Sciences, version

14 (SPSS 14).

3 RESULTS

The inter-rater reliability for the coding criteria was

high (Cohen’s Kappa = 0,924). In total, 1690 verbal

utterances were coded. Average percentage of task

related utterances was M=79% (SD=17.07). Average

percentage of utterances classifiable into our concept

types was 84 (SD=10.5). Average percentage of

congruent ‘arguments’ was 42 (SD=37.81).

4 DISCUSSION

Our preliminary results show high level of

congruence of concepts comprising verbal reports

with the concept types specified in our model.

Moreover, also the order of concept formation as

inferred from verbal reports is congruent with the

order of processing stages specified in our model.

Human conceptualization can be fast. In order to get

hold of the unfolding interpretation process we

introduced a task that, by virtue of its complexity,

forces problem solving to be split into stages. In the

first stage, subjects are typically stuck at a trivial

interpretation of their input (e.g., ‘There is a

mathematical problem’). In this stage, concept types

from the lower part of the model schema (Figure 1)

are dominantly produced. Further stages are more

difficult and often re-focussing is needed in order to

proceed in solving the problem. Alongside, concept

types from higher levels of the model schema are

generated (see Figure 1). Note that additional

material provided to participants only served the

purpose of shifting attention, thus enabling

emergence of alternative interpretations of the

problem, rather than offering information necessary

to solve the problem.

The findings suggest that

solving the problem is effective if the interpretation

process proceeds as suggested in our model.

5 CONCLUSIONS

We have shown that our model can be used to

analyse concept formation in human problem

solving. This means that the model is highly

congruent with human information processing. We

believe that our formal model can contribute to a

further development of existing IA-applications such

as computer supported instruction and human-

computer interfacing as it may serve as an explicit

basis for building such applications. Computer

CSEDU 2010 - 2nd International Conference on Computer Supported Education

292

assisted studies reviewed in Seo & Bryant (2009)

did not show conclusive effectiveness despite the

relatively large effect sizes obtained in these studies.

However, more experiments tapping into on-line

processing are needed in order to explore the

features of the model more extensively.

REFERENCES

Aho, A.V., Ullman, J.D., 1972. The Theory of Parsing,

Translation and Compiling, vol. 1, Prentice Hall.

Breemen, A.J.J.van, Sarbo, J.J., 2009. The machine in the

ghost: The syntax of mind, Signs - International

Journal of Semiotics, Royal School of Library and

Information Science. Denmark.

Bruner, J.S., 1966. Toward a Theory of Instruction,

Harvard University Press.

Delfos, M.F., 2000. Luister je wel naar míj?

Gespreksvoering met kinderen tussen vier en twaalf

jaar, Uitgeverij SWP, Amsterdam.

Engelbart, D.C., 1962. Augmenting Human Intellect: A

Conceptual Framework, AFOSR-3233, Stanford

Research Institute, Menlo Park, CA.

Farkas, J.I., Sarbo, J.J., 2000. A Logical Ontology. In:

ICCS'2000. Shaker Verlag. Darmstadt (Germany).

Hagoort, P., Hald, L., Bastiaansen, M., Petersson, K-M.,

2004. Integration of word meaning and world know-

ledge in language comprehension, Science, vol. 304.

Margolis, E. and Laurence, S. (Eds.). 1999. Concepts:

Core Readings. M.I.T. Press.

Magnani, L., 2001, Abduction, Reason, and Science,

Kluwer Academic/Plenum Publisher, New York.

Peirce, C. S., 1931-58, The Collected Papers of C. S.

Peirce, vols. 1-6, ed. Charles Hartshorne and Paul

Weiss; vols. 7-8, ed. A. W. Burks, Cambridge:

Harvard.

Plato, 1871, Meno, Translation of Benjamin Jowett, web

edition at http://classics.mit.edu/Plato/meno.html

Rigter, J., 1996. Het palet van de psychologie, Stromingen

en hun toepassingen in de hulpverlening, Uitgeverij

Coutinho, Bussum (NL).

Sarbo, J.J., Farkas, J.I., Breemen, A.J.J.van, 2006. Natural

Grammar. In: Semiotics and Intelligent System

Development. Idea Group Publishing. Hersey (PA).

Shadbolt, N., Hall, W., & Berners-Lee, T., 2006. The

Semantic Web Revisited. IEEE INTELLIGENT

SYSTEMS.

Smith, E., Osherson, D., Rips, L., Keans, M. 1988.

Combining prototypes: a selective modification

model, Cognitive Science, 12.

Seo, Y.J., Bryant D.P., 2009, Analysis of studies of the

effects of computer-assisted instruction on the

mathematics performance of students with learning

disabilities, Computers & education, 53(3), 913-928.

NEW CONCEPT DEVELOPMENT MODEL - Explorative Study of its Usability in Intelligence Augmentation

293