International Journal of Artificial Intelligence in Education (2001), 12, 280-292

Scaffolding, contingent tutoring and computer-supported

learning
David Wood
Centre for Research in Development, Instruction and Training
School of Psychology, University of Nottingham, Nottingham, NG7 2RD, U.K.

(CREDIT),

Abstract. This paper provides an overview of the application of principles of tutoring, derived
from studies of face-to-face tutoring, to the design of computer-based tutoring environments.
One of the main theoretical challenges facing work on tutoring is to develop a conceptual
framework to explain how the dynamics and the consequences of learner-tutor exchanges arise
out of the joint regulation of interaction. Ways in which computer-based tutoring have helped to
meet this challenge are illustrated. A primary focus for the empirical work presented is the
impact of individual differences in learners regulation of the tutor on learning outcomes. The
findings demonstrate the potential of contingent, computer-based tutoring for the dynamic
assessment of individual differences in prior attainment and learning. Ways in which computerbased systems might help to develop aspects of learners help seeking skills and their use of
time on task are identified. Important points of convergence between the findings of this work
and research into knowledge-based models of tutoring and learning are discussed, and the
advantages of synthesising the two approaches explored.
SCAFFOLDING AND CONTINGENT TUTORING
The main aim of this paper is to provide an overview of the theory, empirical findings and
issues that motivated the development of computer-based implementations of principles of
scaffolding and contingent tutoring derived from studies of face to face tutoring. It will also
identify and discuss implications of the ensuing research for our understanding of the tutorial
process and its impact on learning.
Early research into scaffolding and tutoring grew out of investigations into the acquisition
of rule governed activity by children under the guidance of an experimenter-cum-tutor (Wood,
Bruner and Ross, 1976). The children were confronted with a construction task that, unaided,
they were unable to master. The main aim of the investigation was to discover whether and
under what conditions 3-5 year olds might be able to master the task and to induce regularities
and patterns in their task activity when an adult helped them.
Although the rules for experimenter engagement with the learners were determined ahead
of time, the forms of interaction that emerged in learner-tutor interaction changed systematically
as a function of tutee age. The three-year-olds drove the experimenter/tutor to provide
qualitatively and quantitatively different patterns of activity than the four-year-olds, and these,
in turn, varied from those solicited by the five-year-olds. It would only be necessary to examine
the behaviour of the tutor to make confident inferences about the age, and task competence, of
the learner. The main features of tutorial support were identified in an analysis of what was
termed scaffolding functions, and the findings of the investigation were employed to illustrate
as to how and why these different functions might allow learners to master and learn with help
that which they could not achieve alone.
Further investigations, including a study of mothers tutoring their 3-4 year olds with the
same task (Wood and Middleton, 1975), demonstrated that learning outcomes for tutees varied
with specific features of tutorial activity. Successful post-tutoring task activity correlated highly
with the extent to which each of the tutors actions reflected the nature of the learners
immediately preceding activity. For instance, if the tutor had made a verbal suggestion with

280

Scaffolding, contingent tutoring and computer-supported learning

which the child did not comply, they provided more help perhaps by pointing out critical
features, or by demonstrating what they meant. Conversely, if the child managed to follow a
suggestion, they would leave him or her with more scope for autonomy perhaps by simply
acknowledging the childs efforts, or remaining silent. Thus, the tutorial interventions of the
contingent tutors were tied to the status of the moment to moment activity of the child in the
problem-solving context (a process subsequently termed instructional contingency see
below).
Several studies have built upon and extended these observations into contexts such as
parental support for mathematics problem solving, picture-book reading, classification tasks and
practical problem solving activities (e.g. DeLoache, 1984; Rogoff, Ellis and Gardner, 1984;
Pratt, Jerig, Cowan & Cowan,1988). Intervention studies have also been employed to
demonstrate that the correlations found between tutorial contingency and learning outcomes are
indicative of a causal influence of tutoring on learning (Wood, Wood and Middleton, 1978).
TESTING AND TUTORING: DYNAMIC ASSESSMENT AND LEARNING
A second line of research that was influenced by work on contingent tutoring, and developed by
Brown and her colleagues (e.g. Brown and Ferrara, 1985; Campiogne, Brown and Ferrara,
1982), has led to applications in the field of dynamic assessment. Brown and her team
compared the performance of children taking IQ test items under two conditions. One, static
assessment, involved children trying to solve problems under conventional test conditions
where they did not receive any help or guidance. The same children were also tested on similar
items under dynamic conditions of contingent help provision, using a series of graded hints.
The investigation demonstrated that dynamic assessment provided a stronger basis for
predicting learning outcomes than the static measures. The greatest learning gains tended to be
achieved by children who needed only minimal levels of guidance. The magnitude of the gap
between assisted and unassisted performance indicated by the amount of help needed was thus
prognostic of individual differences in learning outcomes. Assessing how much help a learner
needed to succeed provided more decisive information about readiness for learning than
determining how often they failed on similar, untutored tasks. This work indicated a diagnostic
potential for principles of contingent tutoring and suggested procedures by which it might be
possible to integrate the assessment of learners prior knowledge with the task of helping them
to learn.
THE LEARNERS ROLE IN REGULATING TUTORIAL INTERACTIONS
Objections have been raised against this conceptualisation of the nature of tutoring and learning
on the grounds that it ignores the learners perspective on the learning task, casts the learner into
an essentially passive role and implies a transmission model of teaching and learning
(Leseman and Sijsling, 1996). A counter to these objections is that by making their actions
contingent upon the activities of the individual learner, tutors are accommodating to individual
differences in the learners task conceptualisation. Since the tutors actions and reactions occur
in response to the learner, the theory also promotes an active view of the learner and an
account of learning as a collaborative, constructive process (Wood and Wood, 1996).
This debate is symptomatic of a longstanding but unresolved conceptual issue that
permeates research on tutoring. There is no accepted scheme within which we can adequately
capture and conceptualise the interplay between the actions of the learner and the tutor as the
processes observed in their interactions emerge. Intervention studies into the impact of different
tutorial strategies on learning outcomes (such as that mentioned above) may establish a causal
influence of specific features of tutoring on learning. However, they do not provide a framework
for understanding how, when and why such features of tutorial interactions do or do not emerge
in a given learner-tutor interaction.

281

Wood
Aspects of the role of the learner in regulating the process of tutorial interaction have been
explored in studies of individual differences in help seeking (Nelson-Le Gall, 1985; Nelson-Le
Gall, Kratzer, Jones & DeCooke, 1990; Puustinen, 1998; Winnykamen, 1993). Here, the claim
is that, by seeking help, the learner both participates in the construction of their own learning
environment and creates an opportunity to develop and practice didactic and auto-didactic skills;
learning how to regulate the learning of self and others.
One result of this research is the recurring finding that lower achieving learners are less
effective help seekers. This may imply, as Puustinen and Winnykamen suggest, that skills in
help seeking are symptomatic of individual differences in meta-cognitive abilities which, in
turn, contribute to the development of differential achievement. However, the interpretation of
such findings is problematic because the potential causal role of the tutor has not been explored.
There is no evidence advanced to rule out the possibility that any differences in help seeking
only emerge because tutorial activity is not contingently adapted to the needs of the lower
achieving learner. For instance, the level of task demands, the timing, specificity and
accessibility of instructional help have not been factored into the analyses as potential
differentiating influences on learner activity.
The findings from the work on face to face tutoring, contingency and assessment exhibit a
similar problem of interpretation. Is contingent tutoring only a way of identifying individual
differences in readiness for learning, or is it also a means for helping to optimise the conditions
under which that learning might take place?
CONTINGENT, COMPUTER-BASED TUTORING
The main reason for extending the work on scaffolding and contingent tutoring to computerbased applications was to provide a strategy for exploring the impact of individual differences
between learners on the processes and the learning outcomes of tutorial interaction. It could also
provide an informative strategy for investigating cause and effect in tutoring more directly,
providing a way of addressing issues that have proved extremely difficult to progress
empirically through studies of face to face tutoring.
EXPLAIN (EXperiments in PLanning and INstruction)
The first attempt to formalise and implement principles of contingent tutoring involved the
construction and evaluation of system designed to support teaching of 3-8 year olds with the
construction task used in the early studies of face-to-face tutoring (Wood, Shadbolt, Reichgelt,
Wood and Paskiewitcz, 1992).
One result of the implementation was a more articulated formulation of the demands
involved. The goal of tutoring is to provide instruction and support that is contingent upon the
learners (potentially changing) level of domain knowledge in contexts where the tutor is
challenging them to master tasks that present manageable problems; problems whose mastery
promises to enhance their domain knowledge. This involves tutorial decisions about what
challenges to set for the learner, if and when to intervene to support them as they attempt tasks,
and how much help to provide if they appear to need support.
Decisions about what tasks to set, or about how to decompose a difficult task down into
potentially easier sub-tasks, are characterised as attempts to realise domain contingency by
which the tutor seeks to engage the learner in levels of challenge that are contingent upon their
current levels of domain knowledge.
When the learner is involved in attempts at problem solving, tutorial decisions about if and
when to intervene are characterised as efforts to achieve temporal contingency. This involves
interpretations of situations in which, for example, the learner is performing operations that are
unlikely to progress problem solving, or in which they are currently inactive. The tutor needs to
decide if and when to treat task inappropriate operations as errors to which the learner appears
committed, or as solution attempts that they are likely to reject as inappropriate for themselves.
Similarly, a tutor may have to decide whether inactivity should be interpreted as a result of

282

Scaffolding, contingent tutoring and computer-supported learning

uncertainty, confusion or disengagement with the task as opposed to a sign of deliberation.
Without access to non-verbal and other situational information about the learner, it is difficult to
envisage how computer-based systems could currently be engineered to gain and interpret such
information. Consequently, no attempt was made to implement temporal contingency in any of
the tutors outlined in this paper. Responsibility for decisions about help seeking was left to the
learner.
Having decided to intervene, or following a request for help, the tutor has to decide how
specific their instructions should be; instructional contingency. Instructions may range from
general feedback about the task appropriateness of a current action, through progressively more
specific verbal instructions to a demonstration of an appropriate next step in problem solving.
The specificity of the help that is provided in any given intervention is contingent upon the
status of the learners preceding action, as outlined above.
The impact of computer-assisted tutoring strategies implemented in EXPLAIN on learning
outcomes closely paralleled those found in face-to-face tutoring. Analyses of the performance
of a contingent tutoring regime demonstrated that near perfect levels of contingent interaction
were achieved with the system. Although uniformly high levels of contingent support were
achieved with all learners, there were marked individual differences in childrens ability to
complete the task for themselves after tuition. Since the rules of tutoring and the degree of
contingency had been controlled, this variability was taken as an index of individual differences
in learning aptitude. A comparison of childrens performance after tutoring with the structure of
their interactions with the tutor revealed, as expected, a close connection between tutorial
interactions and learning outcomes. Children whose activity drove the tutoring environment
into frequent and deep levels of help performed less well after tutoring than those whose activity
prompted the tutor to fade quickly. This suggests that ease of learning and rate of fading are in a
reciprocal relationship, and that the shape of the fading curve exhibited by a contingent tutor is
a mirror image of the learning curve evidenced by a learner.
EXPLAIN was also exploited in an investigation of peer tutoring to establish connections
between measures of learner-tutor interaction and the subsequent performance of the computertaught learners as peer tutors. Learners who were taught by children who had driven the tutor to
provide more frequent and specific help learned less of the task from their tutoring than learners
taught by children who had completed the learning task with relatively little help (Wood, Wood,
Ainsworth and OMalley, 1995).
These findings illustrate in more detail the duality of contingency interaction, which
provides both a means for assessing individual differences in learning aptitude and a measure of
tutorial efficacy. The tutor accommodates systematically to the activity of the learner. The same
learner activities are also an index of ease of learning. Consequently, measures based on the
depth, frequency and time course of tutorial interactions provide (a) a means of (dynamic)
assessment and (b) a way of predicting the likely outcomes of tutoring. This makes it difficult to
assess the extent to which the measures of the interaction between learner and tutor are also an
index of teaching. It is not clear from such evidence whether tutoring activity is simply a
response to extent individual differences or an integral part of the learning-support process
itself.
The QUADRATIC tutor
Quadratic is a contingent tutoring environment designed to teach aspects of the quadratic
function and related polynomial expressions to 14-15 year olds with little or no previous
experience of algebra. The design and structure of the domain material is based on previous
theoretical and empirical work by Dienes (1960) and Bruner and Kenny (1965). The system has
been evaluated in quasi-experimental classroom investigations (Wood and Wood, 1999).
QUADRATIC provides contingent, on-line help at the learners request. The tutor continually
monitors and logs learner activity and, in response to requests for help, exploits principles of
instructional contingency to determine what help to provide. Learners are left to work through
problems at their own pace and, although QUADRATIC does not exhibit domain contingency,

283

Wood
learners are free to abandon problems at will. Thus, the structure and contingencies in learnertutor interaction are dictated by the interplay between system design and learner decisions.
Before taking sessions on the tutor, each learner was pre-tested on a standardised test of
mathematical attainment. The measures used to assess learning gains included tests of each
learners ability to expand quadratic expressions. These were assessed, off-line, before and after
tutoring sessions. On-line measures designed to explore each learners ability to generalise what
they had learned from tutoring in quadratic expressions to support their learning about other
polynomial expressions were also developed.
Data analyses were designed to model the structure of relations between performance on
the pre-tests of mathematical attainment, measures of learner-tutor interaction collected on-line,
and the indexes of learning gains. A specific aim was to test the hypothesis that the measures of
learner-tutor interaction would provide a basis for reliable and valid (dynamic) assessments of
individual differences in mathematical attainment. This was done by comparing pre-test
attainment scores with on-line performance to provide powerful support for the prediction. The
most powerful statistical model based on the learner-tutor interaction measures accounted for
well over 70% of the variance in pre-test attainment scores.
The investigation also assessed the claim, following the work of Brown and her colleagues
(op cit), that the dynamic, on-line measures would provide information about learning gains
independently of that provided by the off-line, static test measures.
PRIOR ACHIEVEMENT, LEARNER-TUTOR INTERACTION AND LEARNING
OUTCOMES
When compared on the same learning gain measures, learners with the higher attainment scores
showed greater evidence of learning that their lower scoring peers. The correlations between the
attainment scores and the magnitude of gains in handling quadratic functions, and with the online measures of the learners ability to generalise learning to related polynomial expressions,
were significant. The strength of the relationship between attainment and gains showed an
increase over the period of tutoring and persisted at post-test.
The on-line interaction measures displayed a similar pattern of association with the gain
scores. Learners who worked through the tutor more quickly, made fewer errors and requested
less help showed most evidence of gains. Because the prior attainment scores and on-line
interaction measures were so highly correlated they accounted for a great deal of shared
variance in learning gains. Learners with higher attainment who showed greatest gains worked
more quickly, accurately and autonomously with the tutor. However, the interaction measures
also provided information about learning outcomes independently of any association with the
off-line test scores. The time that learners spent before seeking help was not related to prior
attainment, but it did predict gains.
More informative were the results that emerged when the variance in learning gains
associated with prior achievement were partialled out of the relation between the learner-tutor
interaction measures and learning gains. For instance, because the lower attaining learners
worked more slowly on the tutor and also did less well in learning to expand quadratic
expressions, a slow rate of working with the tutor emerged as a negative indicator of likely
gains. With individual differences in prior achievement partialled out, however, this relation
moves from negative to positive. In other words, when differences due to prior achievement
were controlled for, it emerged that learners who work more slowly showed most gain. The
frequency of errors followed a different pattern. Even after controlling for prior attainment,
error frequency remained a negative correlate of learning outcomes, as some theories of learning
would predict (e.g. Anderson, 1987; Anderson, Boyle, Farrel, & Reiser, 1987). Thus, it is not
the case that the relations between learner-tutor interaction and learning gains can be explained
away simply as evidence of the assessment of individual differences in learner knowledge or
attainment: An analysis of the process of interaction between learner and tutor adds predictive
and explanatory power.

284

Scaffolding, contingent tutoring and computer-supported learning

HELP SEEKING AND APTITUDE X PROCESS INTERACTIONS IN LEARNING
Learners who achieved lower attainment scores on the off-line test sought help from the tutor
more often than their higher scoring peers. On first sight, this finding appears to be at odds with
those cited earlier in relation to help seeking and face-to-face tutoring. However, the low scorers
with QUADRATIC also made more errors and were less likely to self-correct these than their
higher scoring peers. Thus, their objective need for help was also greater. Individual differences
in the conditional probability of help seeking after errors were consistent with the findings from
human tutoring in that higher attainment correlated with greater conditional probabilities. Thus,
the lower attainers appeared to create a less contingent learning environment for themselves
because they were less likely to seek help when in trouble. Further analyses also showed that
this could not be explained away on the basis any assumption than the help provided by the
tutor was less helpful for the lower scorers. If fact, help had a more positive effect on the
performance and the learning gains of the lower attainers.
In a discussion of the potential value of help provision on learning, Anderson and his
colleagues (Anderson, Corbett, Koedinger & Pelletier, 1995) express the concern that learners
might be predisposed to abuse help, using it to avoid effort. The findings from QUADRATIC
point in the opposite direction: Those learners who should have benefited most from help were
more likely to be help refusers than help abusers. One of the benefits to be claimed for
computer based tutoring is the fact that it is relatively easy to find out who these help refusers
are and to assess the possible impact of any under use of on-line help on learning. Implications
of this observation are considered later in this paper.
DATA (Dynamic Assessment and Tutoring in Arithmetic)
The results from the QUADRATIC tutor, and from the work on help seeking in face-to-face
tutoring reviewed earlier, suggest that learners who are relatively low achievers contribute to the
creation of less-than-contingent tutoring environments for themselves by a failure to recognise
when they need help. There are at least two possible explanations for such findings. One is that
some learners are low achievers, at least in part, because they have relatively less knowledge
and weaker skills in meta-cognition and self-regulation. The second is that they only appear to
have such weaknesses because they are observed in situations that confront them with problems
that they find harder than do their peers. In none of the investigations reviewed was the nature
of the learning task made contingent upon the prior knowledge of the learner. No study has
controlled domain contingency in tutoring, so it seems reasonable to suppose that low achievers
are faced with relatively harder problems than their peers, perhaps explaining why they appear
less competent in help seeking.
DATA has been developed to incorporate and evaluate principles of both domain and
instructional contingency in a system designed to assess and support learning in elementary
arithmetic (Wood, Wood and Marston, 1998). It has been evaluated with a representative
sample of 7-9 year old primary school children, and with children aged 11+ who are
significantly behind in their maths achievement (Wood, 2000).
The structure of the learning domain in DATA is based on a comprehensive and evidencebased account of learning in elementary addition and subtraction which has grown out of several
decades of international research (Fuson, 1992). Fuson offers a taxonomy and developmental
model of the addition and subtraction domain with reference to arithmetical word problems. She
identifies more than 20 different sub-classes of problem type and fits research findings into a
theoretical account of how and why children learn and understand different sub-classes at
different periods of development.
The taxonomy rests on the identification of different types of situation (e.g. problems
which demand quantification involved in increasing or decreasing the size of a set; to quantify
the differences between two different sets; or to equalise two sets) together with specification of
the unknown term in the expression of the problem (e.g. missing start or missing end). DATA
exploits this taxonomy in on-line assessments designed to order performance into different
categories of development.

285

Wood
In order to identify classes of problem that were contingent upon each learners knowledge
of the learning domain, DATA was designed to undertake on-line assessment prior to tutoring.
Briefly, each learner was first asked to solve 6 problems of intermediate levels of difficulty
using Fusons framework. These, classified by Fuson as missing start problems were of the
type John had some marbles. His aunt gave him three more. John now has eight. How many
marbles did he have to begin with?. In such cases, it is the first term in the statement of the
problem that represents the unknown. For the initial test set, only integers less than ten were
used in the statement of the problem. Learners who succeeded with these problems were
presented with missing start problems involving multi-digit addition and subtraction. Those who
made no errors with these problems were not offered any tutoring. Those learners who had
succeeded in solving the single digit missing start problems but made errors with multi-digit
ones were offered tuition with multi-digit problems.
Learners who did not succeed in passing the initial, missing start problems were confronted
with the simplest class of missing end problems of the form Mary had three sweets. She
bought four more. How many sweets does she have now?. Any learner who failed on these
items was tested on similar items with the addition of voice over presentation to test for
possible difficulties in reading. Those who could not solve such items were presented with
similar problems expressed in arithmetical notation to investigate possible difficulties in relating
verbal to arithmetical expressions.
In all, DATA classified learners into one of eight categories ranging from the most
advanced. who succeeded with multi-digit, missing start problems, to ones who failed to solve
missing start problems under any form of presentation (oral, written text or arithmetical
notation). After the on-line testing, all learners were offered instructionally contingent tutoring
in the classes of problems with which they had shown evidence of error.
In order to assess the validity of the on-line testing, the classification produced by DATA
was compared to learners scores on attainment tests presented before tutoring and off-line.
Significant correlations between the two sets of scores provided support for the validity of
Fusons taxonomy and for the assumption that DATA succeeded in differentiating tutorial
demands contingent upon learners prior mastery of the domain.
One motivation for the evaluation was to help adjudicate between the conflicting
interpretations of relations between learners prior attainment and the probability of help
seeking. If low achievers are less likely than higher achieving peers to seek help when they
make an error, the findings from DATA should parallel those from QUADRATIC and from
investigations of face-to-face tutoring. Conversely, if the relations between achievement and
help seeking with the QUADRATIC tutor are attributable to differences in problem difficulty,
performance in tutoring with DATA should exhibit an attenuation or elimination of the
association between attainment and help seeking.
The evidence from DATA supported the second interpretation. In fact, there was no
evidence of any significant associations between scores derived from off-line tests of prior
attainment with individual differences in the probability of correct solutions, errors or help
seeking under tutoring.
Exploring and explaining individual differences in learning outcomes with domain
contingent environments is problematic because each learner is provided with tutoring in
problems that are selected on an individual basis. One potentially useful metric was derived
with DATA by determining the number of problems that each learner experienced before they
achieved autonomous, successful solutions on the problems they initially failed to solve during
testing (i.e. before they drove the tutor to fade completely). On this measure, there were no
correlations between measures of prior attainment and the outcomes of the learner-tutor
interactions. However, learners who exhibited longer time intervals between actions on the tutor
did achieve successful, autonomous performance after fewer problems, mainly because they
also made fewest errors. In contrast to findings from QUADRATIC, however, there was no
association between prior attainment and speed of working with the tutor indicating that
differences in rate of working are dependent upon the relative difficulty of the problems being
attempted. As with help seeking, any correlations between prior attainment and measures of

286

Scaffolding, contingent tutoring and computer-supported learning

assisted problem solving performance may only hold under regimes of tutoring that fail to
achieve domain contingency.
There is clearly a need for more focused evidence than that currently available to inform
the development of an adequate and compelling account of the relations between learners use
of time on task, their help seeking, errors, problem solving and their learning outcomes.
However, the current evidence makes it clear that drawing inferences from studies of (nondomain-contingent) tutoring about the abilities of relatively low achievers is problematic, and
will only be compelling after we have better models of these relations. Computer-based
investigations offer an important and perhaps the best available methodology for developing
these.
The findings concerning relations between time on task, errors and help seeking on
learning have implications that are independent of issues concerning relations with prior
achievement. The findings suggest that learners who produce error upon error, failing either to
self-correct or to seek help, inhibit their own learning. Others try to work their way through
problems at too fast a rate: Their speed of operation (relative to other learners) indicates a level
of task knowledge/achievement that it out of alignment with other evidence of achievement.
Basically, these are learners whose speed suggests that they are more competent than is
indicated by the evidence from independent tests of achievement and/or on-line performance in
problem solving.
If these inferences are sound, then successful tutoring in help seeking, or in the regulation
of time on task, should lead to better learning outcomes for such learners. Currently, we know
of no evidence to tests these claims. It is possible that human tutors acknowledge and
compensate for such individual differences by taking over responsibility for providing help to
learners who show evidence of help refusal and by encouraging learners who are too hasty
(given their level of domain knowledge) to slow down. These possibilities need empirical
investigation. One prediction for test is that tutors who are contingent upon such aspects of
learners self-regulatory activity should be more successful in promoting progress in learning. In
relation to computer-based tutoring environments, the challenge is to design systems able to
detect and respond to on-line evidence of sub-optimal help seeking and use of time on task.
Again, the prediction is that such tutors would be more successful in promoting learning than
the current generation of contingent tutors. Such research would also deepen our understanding
of the nature of causes and effects in learner-tutor interaction and the impact of tutoring on
learning.
LEARNING THEORY AND PRINCIPLES OF TUTORING
Unlike most previous work on computer-based tutoring, the tutoring systems discussed in this
paper arose directly out of empirical work on tutoring, rather than implementations of theories
of learning. Despite such differing points of departure, a degree of convergence onto some
common principles of tutoring has emerged from work in the two traditions (Wood and Wood,
1996b). For example, principles of tutoring identified by Anderson and his colleagues
(Anderson, Boyle, Farrel & Reiser, op cit) and the approach based on scaffolding and
contingency share a common emphasis on the theoretical importance of
providing feedback and guidance in the context of problem solving
supporting successive approximations to competent performance through tutorial fading
varying the grain size of instruction as learners make progress
engineering problems of controlled complexity, contingent upon the learners
developing knowledge and skills.
One major difference between the two approaches rests on the status of student modelling.
Of crucial theoretical significance in Andersons approach is the commitment to represent both
learner knowledge and the knowledge to be learnt as production rule models. The current

287

Wood
versions of contingent tutors build no model of the learner at all; their performance is local and
situated - constrained by contingencies in the learners current activity. Although the systems do
aggregate data for purposes of evaluation and assessment, this information currently plays no
part in driving any tutorial action by the systems, thought it is currently being used to specify
and to evaluate when tutoring in help seeking might enhance learning.
As argued elsewhere (Wood and Wood, 1999), the two approaches could benefit from a
degree of integration. Student modelling could be extended to exploit information derived from
learner help seeking and from indexes of the impact of tutorial help provision as a means of
informing the construction of the learner model (in line with the approach taken by Luckin and
Du Boulay, 1999, for example). This should help to improve the assessment of the learners
grasp of domain knowledge and extend the scope of the models of learning to integrate
information about aspects of how learners learn with the assessment of what they have learnt.
The results from QUADRATIC and DATA suggest ways in which it would be possible to
monitor and identify learners who could benefit from tuition in help seeking, or from advice
about their use of time on task, on the basis of their performance data. If this information was
integrated with a model of learner knowledge, it should be possible to enhance the detection any
problems in the learners self-regulation and to generate and test hypotheses about their origins.
For example, help abuse would be signalled when a learner seeks help to solve problems
demanding knowledge that the learner model suggests they already possess; help refusal when
they persist in attempts to escape from an impasse without help, despite inadequate knowledge.
Allied to a theory that models the connection between domain knowledge with help seeking and
the distribution of time on task, it should also be possible to explain when and how any changes
brought about through tutoring, and which impact on subsequent assessments of learning gains,
are mediated.
MULTIPLE REPRESENTATIONS AND INTEGRATED KNOWLEDGE
STRUCTURES
Both QUADRATIC and DATA make use of multiple representations (diagrammatic,
arithmetical and, in QUADRATIC, algebraic) to support learning. The rationale underpinning
their use rests on the claim that conceptual understanding resides in the learners construction of
integrated knowledge structures (Baxter and Glaser, in press). A key element in the
construction of such knowledge is an appreciation of the equivalencies and correspondences
across superficially different but theoretically related representations. An important mark of
domain expertise is the ability to select a representational system which is most likely to fit the
demands of particular tasks (e.g. Kaput, 1992) or ones own cognitive strengths (Nelson-Le
Gall, Kratzer, Johnes & DeCooke, 1990).
Findings from QUADRATIC illustrate how changes in the processes observed in learnertutor interaction provide potential measures for tracking the construction of such structures
(Wood and Wood, 1999). This approach exploits changes in demands for help by learners and
in the level of help provided by the tutor as a means of exploring learning generalisation. For
instance, the amount of help that learners requested (and made successful use of) to solve
specific quadratic expressions (e.g. (x+2)2 etc..) was compared with the amount of help
requested to expand the more general case (x + n)2. Similar measures were derived as, later in
the tutorial sessions, the learners solved specific and more general cases of cubic expressions
(e.g. (x + n)3). Learners requested more help to solve the initial cases of cubic functions than
they had asked for in solving the initial cases of the quadratic ones, implying that they found the
cubic cases more difficult initially. However, tutorial fading was more rapid when they moved
on to the general case of the cubic, which was solved with less help than that requested to solve
equivalent examples of the quadratic case. This may indicate that learning about a generalisation
from specific instances to a general case proved easier with the cubic function, and that this
resulted from the transfer of experience from previous problem solving with the quadratic. If so,
the would provide evidence that the learners were acquiring a degree of abstract and conceptual
understanding, not only specific procedures for solving particular types of problem. Here too, an

288

Scaffolding, contingent tutoring and computer-supported learning

integration of the interaction process data with a model of the learners knowledge would
provide a means of detailing and evaluating such inferences (Gobet and Wood, 1999).
The potential benefits (and a number of possible disbenefits) arising from the use of
multiple representations to support computer-based learning are a focus for considerable
contemporary research. As Ainsworth (1999) argues, we stand in need of new approaches to the
assessment of performance in these new contexts in order to evaluate claims about the supposed
beneficial effects of multiple representations on learning. We need such evidence both to enrich
our theoretical understanding of the nature of conceptual understanding, and to develop
enhanced models of knowledge and learning. Dynamic, on-line assessment provides one
promising approach.
MULTIPLE TUTORIAL AGENTS TO SUPPORT LEARNING
The use of on-line, interaction process measures to investigate tutoring and learning that I have
considered has been focused on the use of help to support on-going performance. However, the
strategy of investigation developed has wider application, and we have been extending it to
explore other aspects of the issue of how learners influence the construction of their own
interactive learning environments.
The general approach is to design environments so that the control of key functions or tools
is left in the hands of learners. As a first step, this helps us to assess the extent to which there is
evidence for any connection between variations in the patterns of tool use and selected
outcomes of learning. Where connections are discovered, it is possible to assess how far the
decision to leave the learner with responsibility for regulating the environment without any
tutorial guidance was a wise one. If individuals are identified who might benefit from tuition in
tool use, attempts can be made to help them to enhance their regulation of the tutor. Any
predictable and beneficial effects on their learning which follow then provides some confidence
in the inference that the processes involved in tool use mediate learning. It also enhances our
understanding of self-regulation and learning.
DATA has been designed to offer the learner access to another tutorial agent which
enables them to access, review and analyse their own immediate problem solving at will. There
are strong theoretical grounds for the assumption that encouraging such reflective activity plays
an important role in learning and understanding . As I have tried to demonstrate in the study of
self regulation, help seeking and learning, the exploitation of computer-based tutoring systems
should provide a novel and powerful means of exploring these issues and of developing
principles for the design of tools to support learning.
CONCLUDING COMMENTS: SOME LIMITATIONS ON CONTINGENCY THEORY
The overarching goal of the research just outlined is to develop a better understanding of the
nature of tutoring and learning. The main aim of the present paper was to assess what the design
and evaluation of computer-based tutoring environments might add to our understanding of how
individual differences in a learners regulation of both their own problem solving activity and
the actions of the tutor relate to the outcomes of learning under tutoring.
The exploitation of computer-based environments was not designed to emulate human
tutoring but (a) to try to determine what methodological benefits the use of computer-based
models of tutoring have to offer and (b) to provide a means of exploring the limits of
contingency theory. One example of (a) is the description and analysis of learner-tutor
interaction to provide a means of assessing individual differences in prior attainment and
learning outcomes. An illustration of (b) is the unmet challenge presented by a need to identify
and formalise the processes which underpin temporal contingency.
In a critique of models of scaffolding in application to literacy teaching, Hobsbaum, Peters
and Sylva (1996) argue that the specific goals for tutorial action often arise out of the process of
tutorial interactions and do not appear to follow a pre-arranged programme. There is

289

Wood
considerable evidence in the research literature that, even in well-structured domains like
elementary school mathematics, learners often develop their own problem solving strategies
which differ from those taught. If our intuitions about the nature of domain contingency are
sound, then the competent tutor should make any help or guidance they provide contingent upon
such learner conceptions, even if the ultimate goal is to help them to learn more powerful
conceptions and methods. None of our current tutors are able to reliably diagnose such
idiosyncratic conceptions and, hence, cannot offer any guidance that is contingent upon them. In
this sense, critics of contingency theory are correct in the claim that current versions of the
theory have no compelling explanation as to how and when a human tutor might accommodate
to the learners task perspective when this differs from their own.
The current limitations of the approach, which underpin weaknesses in both temporal and
domain contingency, are, of course, reflections of more pervasive constraints on our ability to
describe, formalise and implement the process of communication. As du Boulay and Luckin
(1999) observe, the limiting factor in the development of effective intelligent learning and
teaching environments will continue to be, for some time, more their impoverished ability to
deal with language input and output than their ability to model. And language subsumes nonverbal aspects of communication whose interpretation underpins the capacity to offer
temporally contingent guidance to the learner.
These theoretical limitations on our ability to develop adaptive learning environments seem
likely to be with us for some time and, this implies, as du Boulay and Luckin also argue, that we
need to design systems for educational application in a way which respects and exploits the
classroom as a social forum in which discussion about what is being learned is a central aspect
of learning. However, the creation of environments which help to inform such discussions
with well founded knowledge about the nature of what and how individuals learn, provides
ample scope for useful contributions to educational practice as we continue to try to overcome
the current theoretical limitations.
Acknowledgements
CREDIT is supported by the Economic and Social Research Council, whose support is
gratefully acknowledged. The main architect of much of the work reviewed in this paper is Dr
Heather Wood, whose contribution is acknowledged with gratitude.
References
Ainsworth, S. E. (1999). A functional taxonomy of multiple representations. Computers and
Education, 33(2/3), 131-152.
Anderson, J. R. (1987) Production systems, learning and tutoring. In D.Klahr, P. Langley & R.
Neches (Eds) Production system models of learning and development. Cambridge, Mass:
The MIT Press.
Anderson J.R., Boyle, C.F., Farrel, R. and Reiser, B J.(1987) Cognitive principles in the design
of computer tutors. In P. Morris (Ed), Modelling cognition (pp. 93-133). Oxford: John
Wiley.
Anderson, J.R., Corbett, A.T., Koedinger, K.R. & Pelletier, R. (1995) Cognitive tutors: lessons
learned. The Journal of the Learning Sciences, 4(2), 167-202.
Baxter G. P. and Glaser, R. (in press) Investigating the cognitive complexity of science
assessments. Educational measurements: Issues and practice.
Brown A.L. and Ferrara, R.A. (1985) Diagnosing zones of proximal development. In J.V.
Wertsch, Culture, communication and cognition: Vygotskian perspectives. New York:
Academic Press.
Bruner J.S. and Kenny, H.J. (1965). Representation and mathematics learning. Monographs of
the Society for Research in Child Development, (30),1, 50-59.
Campiogne, J.C., Brown A.L. and Ferrara, R.A.(1982) Mental retardation and intelligence. In
R.J. Sternberg, Handbook of Human Intelligence. Cambridge: Cambridge Univetrsity
Press, pp 392-490.

290

Scaffolding, contingent tutoring and computer-supported learning

DeLoache, J.S. (1984) Whats this? Maternal questions in joint picture book reading with
toddlers. Quarterly Newsletter of the Laboratory of Human Cognition, 6, 87-95.
Dienes, Z.P. (1960) Building up Mathematics. London: Hutchinson.
du Boulay B. and Luckin. R. (1999) It aint what you learn but the way that you learn it.
Computers and Education, 33, 209-215,
Elbers, E. (1996), Cooperation and social context in adult-child interaction. Learning and
Instruction, Vol 6, No 4, 281-286.
Fuson (1992) . In D.A. Grouws (Ed) Handbook of research on mathematics teaching and
learning. New York: Macmillan.
Gobet, F. & Wood, D. (1999). Expertise, models of learning and computer-based tutoring.
Computers and Education, 33(2/3), 189-207.
Hobsbaum, A., Peters, S. and Sylva, K. (1996) Scaffolding in reading recovery. Oxford Review
of Education, 22 (1), 17-35.
Kaput, J.J. (1992) Linking representations in the symbol systems of algebra. In D.A. Grouws,
Handbook of research on mathematics teaching and learning (pp 167-294). New York:
Macmillan.
Leseman P.P.M, and Sijsling, F.F. (1996) Cooperation and instruction in practical problem
solving. Differences in interaction styles of mother-child dyads as related to socioeconomic background and cognitive development. Learning and Instruction, Vol 6, No 4,
307-324.
Luckin R. and Du Boulay, B. (1999) Ecolab: The development and evaluation of a Vygotskian
design framework. International Journal of Artificial Intelligence in Education, Vol 10,
198-220.
Nelson-Le Gall, S. (1985) Help-seeking behavior and learning. Review of Research in
Education, 12, 55-90.
Nelson-Le Gall, S., Kratzer, L., Jones, E. & DeCooke, P, (1990) Childrens self-assessment of
performance and task-related help-seeking. Journal of Experimental Child Psychology, 49,
245-263.
Pratt, M,W., Jerig, P., Cowan, P.A. and Cowan, C.P. (1988) Mothers and fathers teaching threeyear-olds: authoratitive parenting and adults use of the zone of proximal development.
Developmental Psychology, (24) 6, 832-839.
Puustinen, M. (1998) Help-seeking behaviour in a problem-solving situation: Development of
self-regulation. European Journal of Psychology of Education, XIII(2), 271-282.
Rogoff, B., Ellis, S. and Gardner, W. (1984) Adjustment of adult-child instruction according to
childs age and task. Developmental Psychology, 20, 193-199.
Winnykamen, F. (1993) Gestion socio-cognitive du recours a laide dautrui chez lenfant.
Journal International de Psychologie, 28, 645-659.
Wood, D. (1999) Learning, problem solving and tutoring: Learning from computer-based
systems. Paper to the London Conference of the British Psychological Society. Dec 19-20th,
Intitute of Education, London.
Wood, D., Bruner, J. S., & Ross, G. (1976). The role of tutoring in problem solving. Journal of
Child Psychology and Psychiatry, 17(2), 89-100.
Wood, D. J., & Middleton, D. J. (1975). A Study of assisted problem solving. British Journal of
Psychology, 66, 181-191.
Wood, D. and Wood, H. (1996a) Contintgency in tutoring and learning. Learning and
Instruction, Vol 6, No 4, 391-398
Wood, D. & Wood, H. (1996b) Vygotsky, tutoring and Learning. Oxford Review of Education,
22 (1) 5-16..
Wood, D. J., Wood, H. A., & Middleton, D. J. (1978). An experimental evaluation of four faceto-face teaching strategies. International Journal of Behavioral Development, 1, 131-147.
Wood, H, Wood, D, & Marston, L. (1998) A Computer-Based Assessment Approach to Whole
Number Addition and Subtraction. Technical Report No. 56. Centre for Research in
Development, Instruction & Training, University of Nottingham.
Wood, D., Shadbolt, N., Reichgelt, H., Wood, H., & Paskiewicz, T. (1992). EXPLAIN :
Experiments in Planning and Instruction. AISB Quarterly(81), 13-16.

291

Wood
Wood, D.J., Wood, H., Ainsworth, S.E. & OMalley, C. 1995 On becoming a tutor : towards an
ontogenetic model. Cognition and Instruction 13(4), 565-581
Wood, H. A., & Wood, D. J. (1999). Help seeking, learning and contingent tutoring. Computers
and Education, 33(2/3), 153-170.

292