ArticlePDF Available

Dynamic course generation on the www

Authors:

Abstract and Figures

This paper presents an approach and architecture for Dynamic Course Generation, based on applying AI planning techniques to a structured representation of the domain knowledge and allowing explicit representation of teaching expertise. An individual course is generated automatically for a given teaching goal and is dynamically adapted at run time to the student's individual progress and preferences according to the teaching expertise. The separate representation of the teaching materials from the domain structure allows an easier updating and re-use of ready CAL materials. In this way our approach provides an alternative to traditional CAL-authoring. An implementation in a simple engineering domain is described An evaluation of the benefits of this approach in terms of cost-effectiveness for authoring is shown.
Content may be subject to copyright.
1
Dynamic Course Generation
Julita Vassileva
Universität der Bundeswehr München
85577 Neubiberg, Germany
jiv@informatik.unibw-muenchen.de
Abstract
This paper presents an approach and architecture for Dynamic Course Generation, based on
applying AI planning techniques to a structured representation of the domain knowledge and
allowing explicit representation of teaching expertise. An individual course is generated
automatically for a given teaching goal and is dynamically adapted at run time to the student's
individual progress and preferences according to the teaching expertise. The separate representation
of the teaching materials from the domain structure allows an easier updating and re-use of ready
CAL materials. In this way our approach provides an alternative to traditional CAL-authoring. An
implementation in a simple engineering domain is described An evaluation of the benefits of this
approach in terms of cost-effectiveness for authoring is shown.
Keywords: authoring, adaptive CAL, discourse strategies, domain ontology, instructional tasks and
methods, Intelligent Tutoring Systems, ITS, ITS authoring, tutoring systems, teaching strategies.
1. Introduction
The need of bringing together the fields of CAL and Intelligent Tutoring Systems (ITS) has been recognized
(Larkin & Chabay, 1992) and there have been attempts for "intellectualizing" CAL. One possible approach is to start
from a set of teaching primitives defined at different levels of generality and to manage their sequencing. Schemes
for controlling the dialogue and presentation of teaching materials have been introduced borrowing from the
formalisms for representing natural language dialogues, for example augmented transition networks, like in (Woolf,
1987) and (Murray, 1992) or by taking a task-based perspective - defining instructional task hierarchies and
modeling instruction as planning a sequence of tasks (Van Marcke, 1991). Several approaches take the other
direction - of "de-intellectualizing" ITS - applying ITS-shell architectures (Elsom Cook & O'Malley, 1989),
(Leonhardt et al, 1991), (Linard & Zeiliger, 1995), (Brusilovsky, 1992). Our approach for Dynamic Courseware
Generation (DCG) (Vassileva, 1992) falls into this stream. It is based on an ITS-shell architecture (Vassileva, 1990),
whose main idea is applying AI planning techniques to determine the content of instruction, which was originally
proposed in (Peachey & McCalla, 1986). Based on an explicit representation of the structure of the concepts / topics
in the domain and a library of teaching materials, the system dynamically generates instructional courses. The
course-plan is created individually for a given student with a given teaching goal; the plan is substantiated with
teaching materials and can be changed at run-time according to the changing learning needs of the student. The main
advantage of this approach is that it allows automatically building goal-directed adaptive CAL courses which is
impossible within the traditional CAL concept of courseware. Because of the separation between the structure of
knowledge from its presentation, it supports an easier authoring: extending and modification of the teaching
materials as well as re-using existing CAL materials. By organizing the Domain Structure so as to distinguish
between various domain aspects, it is possible to generate content plans representing different viewpoints of the
material. The partially generic pedagogical knowledge, which is explicitly represented by means of instructional task
hierarchies and teaching rules, as proposed by Van Marcke (1991), provides for ensuring pedagogical consistency of
2
the automatically generated courses. In this paper we describe the architecture and functioning of the DCG, the DCG
as an authoring tool and an application of the tool to an engineering domain.
2. Characteristics of Dynamic Courseware Generation
A course generated by a DCG looks for the student like a CAL course. However, this course is generated
individually for every student to achieve a certain teaching content goal (a topic or concept that has to be learned),
and it takes into account the already existing knowledge and preferences of the individual student. In addition, the
course is dynamic, i.e. it changes at run time and adapts to the progress of the student, his/her learning style and
preferences. The main differences between the two approaches are summarized in Table 1.
Table 1. Differences between CAL and DCG.
Between:
Differences in:
CAL
Dynamic Courseware Generator
Course Definition /Selection Pre-defined course with a fixed
teaching goal
A goal-oriented course is generated
at run time
Starting Point Fixed starting point or a choice of
several starting points
Selected with respect to the
student's prior knowledge
Presentation Materials Fixed sequence of
Presentation Materials
Dynamically selected accor-ding to
teaching strategies and / or the
student's preferences
Way of Following the Course Pre-defined sequence accor-ding to
the hypothetical prog-ress of an
average student
Automatic course re-palnning at run
time to match indivi-dual
differences in knowledge and
teaching strategies
Teacher's Role Excluded (unless she herself is the
Author of the course)
Assigns the teaching goal and can
manage the teaching strategy by
editing the set of teaching rules
3. Architecture and Functioning
The system implements a combination of the DCG architecture (Vassileva, 1992) which dynamically generates
a content plan of the course with a given goal and the GTE architecture (Van Marcke, 1991) which, by means of a set
instructional task hierarchies and methods, decides how to carry out the plan in an optimal for the student way
according to a set of teaching rules. The main feature of the proposed architecture is the separation between the domain
knowledge structure from the presentation materials. The idea of the DCG is to use a classical AI mechanism for
planning in an AND/OR-graph representation of the domain knowledge for automatic generation of a content plan of a
course (see Figure 1).
3
Figure 1: Content Planning in the Domain Structure.
The architecture of the system is shown in Figure 2. It will be discussed in the next sections.
Figure 2: Architecture of the System
A
B
C
D
Course Generator Domain Database
Authoring Module
Student Model
Teaching
Rules
Editor Set of Teaching
Rules
Instruct. Tasks
& Methods
Domain
Structure
PLANNER
EXECUTOR Teaching
Materials
Knowledge
History
Pers.Traits
Editor for Instruct.
Tasks & Methods
Domain Structure
Editor
Teaching Materials
Editor
Course
Pedagogical Component
4
3.1 The Domain Data Base
The subject knowledge is contained in the Domain Database Component. It contains two parts:
The Domain Structure contains the concept/topic structure of the subject knowledge that is going to be
taught. It is represented as an AND/OR graph. The nodes represent the elements of knowledge (concepts, topics,
rules etc.). If two nodes A and B are connected with a third one, C, with an "AND" link, this means that both nodes
A and B have to be developed when following this link from C. Otherwise, they would be considered as alternatives,
i.e. there is a choice of nodes to be developed, either A or B. The links represent the possible relationships between
them. These relationships can have various semantics. For example, if nodes A and B are connected with node C
with an AND-relationship of type "aggregation", this means that C contains sub-components A and B. If they are
connected with an OR-relationship of type "generalization", this means that C is a general concept with possible
instances A or B. There are many other possible semantic relationships, for example, causal relationship, temporal,
analogy, simple prerequisite etc.
The simplest way to define a Domain Structure it to use only one possible semantic of links, for example, to
link domain concepts / topics with prerequisite links encoding pedagogical following. In this way one obtains a
curriculum-like structure which can be used to guide the sequencing of content. This approach has been originally
proposed in (Peachey & McCalla, 1986) and can be seen appearing in literature under different names: content model
(Van Marcke, 1991), pedagogical structure of the domain (Vassileva, 1990), (Mitrovic, 1996), or pedagogical
content knowledge (Leinhardt, 1988), (Calderhead, 1991).
AND/OR graphs have been selected as a representation formalism, since their expressive power is very high
- it is equivalent to decomposable production rule systems (Nilsson, 1980) or to STRIPS-like operators. In addition,
AND/OR-graphs can be visualized on the screen, which has some psychological advantages for authoring and
teaching.
It is possible to organize the concepts / topics in a domain into a set of smaller, possibly interrelated
AND/OR-graphs, representing relatively independent sub-areas of the domain, or different "Views". We call such
sub-graphs "Aspects".
Every node and every link from the Domain Structure is associated to a set of TMs which represent (teach,
explain, exercise and test) this concept. The Domain Structure is used for creating a plan of the course-contents (a
sub-graph of Domain Structure) to achieve a given teaching goal (a concept). This plan is called a "Content Plan"
and the process - "Content Planning". During the course execution TMs are selected by different instructional tasks
to teach the concepts / topics to the student.
The
Teaching Materials (TMs) contains presentation- and testing-units that carry out the communication
with the student, i.e. they are in fact what the student sees on the screen. Each TM is focused on a given topic,
concept or relationship. The TMs are classified according to their pedagogical function. For example, an
introduction, a motivating problem, an explanation, help, exercise, or test. In this sense TMs are equivalent to the
"instructional primitives" in GTE (Van Marcke, 1991). TMs that carry out a dialogue with the student -- for example,
exercises and tests are represented with a set of smaller units providing a pre-stored correct answer to the exercise /
test, a hint or help, explanation, eventually intermediate stages of solving the problem etc. TMs of type "test" have in
addition two associated probabilities denoting to what extent the student's correct / incorrect answer means that the
student knows /doesn't know the concept(s) which they are supposed to test. The TMs are also classified with respect
to the media they use, i.e. textual, graphical image, animation, video etc.
5
3.2 The Student Model
The Student Model contains three parts: a model of student knowledge (the concepts / topics and relations
that have been taught), a history (the instructional tasks / methods and the TMs that have been used) and a model of
the student's personal traits and preferences.
The model of Student Knowledge is an overlay over the Domain Structure in the Domain Data Base,
containing probabilistic evaluations of the beliefs that the student a given concept. Updating of the knowledge
probabilities is carried out dynamically as a result of the student's correct/erroneous answer to a test-TMs. This is not
an original student modeling technique and we will not discuss here in detail its mechanism. More details can be
found in (Diessel et. al., 1993); the original technique is described in (Villano, 1992).
The History contains a list of all instructional tasks, methods and subtasks that have been used for every
concept during following the plan. It also stores statistics about the success of different instructional task-
decomposition methods (see section 3.3.) and statistics of the success of the various media-types of TMs used (text,
sound, graphics, animation, video).
The model of the student's Personal Traits and Preferences contains two lists of variables with their
values. The first one contains psychological features like confidence, motivation, concentration and the second one
contains the student's preferences to different types of media. The values can take three discrete values (Low,
Medium, High) and are assigned by the student at the beginning of the teaching session.
3.3 Pedagogical Component
This component contains two main parts, a set of Instructional Tasks and Methods and a Set of Teaching
Rules, each of which has a generic kernel and can be expanded with subject specific knowledge (see Figure 2). In
addition, the Pedagogical Component contains a Teaching Rules Editor which allows the teacher to modify and
assign new teaching rules.
Instructional Tasks and Methods
This part of the Pedagogical Component contains a representation of instructional tasks and their
decomposition into sub-tasks by means of different instructional methods, similarly to the GTE (Van Marcke, 1991).
Like the Domain Structure, the instructional task decompositions can be represented with AND/OR graphs, however,
here the nodes represent tasks and the links -- task-decomposition methods. The AND-links represent links to sub-
tasks of a certain task according to a certain task-decomposition method. The OR-links correspond to alternative
task-decomposition methods. For example, Figure 3 represents the generic task "Give exercise" which can be
decomposed to a sequence of the following sub-tasks: "Make exercise", "Verify", "Remedy" (adapted from Van
Marcke, 1991). The sub-task "Remedy" can be decomposed in different ways according to different methods (OR-
types of links shown in Figure 3 with different types of lines). The purpose of the instructional task-hierarchies is to
allow planning of the sequence of TMs focused on one given topic / concept from the content plan, i.e. to plan in a
pedagogically meaningful way the presentation of a certain topic / concept. Therefore, this plan will be called
"Presentation plan" and the process - "Presentation planning". The tasks and methods can be generic, but as
suggested in (Van Marcke, 1991), the deeper the sub-tasks are in the task-decomposition hierarchy, the more subject-
6
dependent they become. A special editor is provided in the Authoring Module so that the pre-defined set of generic
instructional tasks and methods in the system could be extended with subject-specific ones.
Give Exercise
Make Exercise Verify Remedy
Present Question
exercise Solution
Check Inform
Response Student
Give
Correct
Solution
Explain
Hint
Elaborate
on sub-
problems
Retry
Figure 3: An Example of a Task Hierarchy
The Set of Teaching Rules
The Teaching Rules manage the selection of content and presentation plans. Most of them are generic. The
Teaching Rules can be classified into the following cathegories:
Discourse rules
These rules manage the plan selection when planning or re-planning at the content -(Domain Structure) -
level. They establish criteria of how to select the plan when several alternative plans are possible (for example,
according to what type of semantic links to plan, when to allow switching to follow a different type of link etc.) and
how to follow the plan. One discourse rule, inspired by (Flammer, 1975), states that the appropriate way of following
the plan in case that the student is intelligent, is deductive (top-down) with respect to abstraction links (from general
to specific) or with respect to aggregation links (from whole to parts). For not well doing and not confident students
inductive (bottom-up) following the plan is better. If re-planning on the content level is needed, the teaching rules
select whether it will be local re-planning or a global change of the plan (see section 3.4 and Figure 6).
Strategy-selection rules
These rules define how to select the teaching strategy before starting the execution of the plan. The teaching
strategy defines the general principles of teaching, for example, who has the initiative in deciding what to do next -
the student or the system. We distinguish between two main types of teaching strategies: structured and unstructured.
A structured strategy means that the initiative is in the hands of the system: it selects which concept will be taught
next and how (i.e. with which instructional task). An unstructured strategy leaves the choice of a next concept to the
student. Presenting the Domain Structure on the screen and highlighting of the "ready to be learned concepts /
topics", i.e. those whose prerequisites in the plan are considered as known by the student, as in (Beaumont &
Brusilovsky, 1995), can help her navigate in the Domain Structure. The student can also choose an instructional task
and method for the current concept from the graphical representation of the task hierarchies. One strategy-selection
7
rule, for example, states that if the student is motivated, an unstructured strategy would be appropriate. If she is
unsure and not confident, the structured methods should be preferred (Siegler, 1988).
Both the discourse and strategy-selection rules take into account data from the student model (student's
knowledge and personal traits) as well as external factors like time.
• Method-selection rules
There are usually three to eight (Van Marcke, 1991) alternative task-decomposition methods for each
instructional task. The method-selection rules take into account the student's History and his / her Personal Traits
and Preferences from the Student Model in order to decide what main instructional task will be selected for the
current concept and which task-decomposition method will be chosen. This is done at the stage of presentation
planning. The method-selection rules solve the problem in GTE with the definition of relative applicability
conditions of the task-decomposition methods, i.e. how to select among alternative possible methods for a given task.
In the pedagogical literature (Einsiedler, 1976, pp. 290-293) one can find four methods for teaching a
concept which decompose the main task "teach" into sub-tasks. The "hierarchical" method teaches by a sequence of
the sub-tasks "introduce", "explain" and "give example", "give exercises", and finally "give a test". The "advanced
organizer" method performs the same task-decomposition with an additional first sub-task which presents explicitly
to the student the current teaching goal and the plan of sub-tasks which are going to be executed, and an objective
stating what is expected from achieving the current goal, i.e. what is the importance of learning the current concept
for the global goal of the course. The "basic concept" method's first sub-task is to present a problem (exercise)
whose solution requires knowledge of the goal concept. The student is not expected to solve the problem, since he /
she lacks knowledge related exactly to the concept which is going to be taught. However, the attempt to solve the
problem prepares him / her to understand the need for the new concept. Then the new concept is introduced,
followed by examples and explanations of how this type of problems are solved, then exercises, and finally a test.
The "discovery" method involves presenting a motivation problem, analyzing the problem and letting the student
solve the problem alone, hoping that s/he will discover the concept in the problem solving.
Following Einsiedler (1976) we defined a set of rules asserting that:
_ the "basic concept" method has to be preferred for successful, but not highly motivated and concentrated
students,
_ the "advanced organizer" method -- for successful, but not very concentrated and confident students,
_ the "hierarchical" method -- for concentrated students, with not very good success so far,
_ the "discovery" method -- for confident, concentrated students.
Teaching Material Selection Rules
When the current instructional sub-task is selected and decomposed down to primitive sub-tasks, the
teaching material selection rules decide how to select a TM on an appropriate type of media (i.e. text, graphics,
animation or video etc.). They take into account the model of the student's preferences in order to select among
numerous TMs for the selected primitive sub-task, those which use a type of media, preferred by the student.
Teaching Rules Editor
The Teaching Rules Editor (Figure 4) allows the teacher to define her own teaching strategy-, method- and TM-
selection rules. This is done by assigning conditions for the application of the rule (variables from the student model) and
effects (decisions of choices).
8
Figure 4: The Screen of the Teaching Rules Editor
It is clear that the Set of Teaching Rules plays an important role for the functioning of the DCG. However, the
Teaching Rule Editor itself doesn't solve the problem of creating the rules. How do we get such rules? Three approaches
are possible:
theory-based: to compile them from existing didactic theories - the current solution (Bohnert, 1995). The
disadvantage of this approach is that most didactic theories are too general and do not formulate precise rules. The
designer / author / teacher needs to interpret the general directions described in the theories to obtain some concrete rules
that can guide action in a specific situation. This interpretation is always subjective and could be criticized.
person-based: to interview teachers or to ask them to implement the rules directly themselves. This, however,
requires that the teachers are able to articulate the factors influencing their decisions which is not often the case. This
method could also lead to invalid rules, since there is a lot of evidence that people reflect on their decisions in a different
way than they actually take them (because they are aiming at a "post-mortem" logical justification of their actions).
empirically-based: to analyze protocols of real individual teaching sessions, to identify cases, and apply
machine learning techniques to generate decision trees and rules. This approach would probably give more reliable
results. However, it is more difficult to realize, since it requires a lot of empirical data. We have developed a machine
9
learning tool for generation of decision trees and rules from descriptions of cases (Horstmann, 1995) and we intend to
apply it for generating rules from protocols of teaching sessions once such data is available.
3.4 The Course Generator
The Course Generator is the component that creates the course, carries out the interaction with the student
and maintains the Student Model. The Course Generator contains the following components:
The Course Planner
The course planner is an AND/OR graph planning program which can be invoked with two purposes:
• to generate a content plan (the concepts / topics to be presented in the course) according to the teaching
goal assigned by the teacher;
• to create a presentation plan (a task-sequence) for teaching the current goal-concept.
The teacher calls the Planner for a particular student and assigns a teaching goal for the course, a given set
of aspects to be covered by the plan, link types with respect to which to plan and maximal deepness in traversing the
graph. If there are discourse rules which assign these parameters for a certain teaching goal (e.g. as in our prototype
for the goals "acquaintance", "installation", "maintenance" and "diagnosis"), the task of the Teacher is only to assign
a teaching goal-concept. The Planner is activated to create a course plan. The planning algorithm is a modification
of the AO* graph search algorithm (Nilsson, 1980). The optimization function h can be selected so as to achieve
different criteria for optimality (i.e. for plan-selection), e.g. the shortest plan, a plan avoiding a certain concept, a plan
with a certain topology-type etc. The selection of h is managed by the discourse rules. The solution graph is an AND-
graph which starts from concepts / topics known by the student (with high knowledge probabilities in the Student
Model) and leads to the goal concept / topic. The plan imposes only a partial ordering on the solution steps. The final
ordering of sub-goals is done at run time by domain-specific discourse rules and according to the selected teaching
strategy.
The Executor
The Executor receives the plan from the Planner. A main teaching strategy (structured or unstructured) for
teaching is selected, by checking the strategy-selection rules. If an unstructured strategy has been selected, the
student has to choose the next concept which s/he wants to be taught from a graphical representation of the plan and
then to select an instructional method which will be used. If a structured strategy is selected, the executor consults the
discourse rules again and chooses the current concept or link to be taught, then consults the method-selection rules,
and selects an instructional method. Then it invokes the Planner to create a plan of the instructional sub-tasks which
are needed to implement the chosen method. Finally, the TM-selection rules are consulted to select an appropriate
TM (see Figure 5).
The selected TM is presented according to the teaching sub-task, then the next sub-task from the task plan is
executed etc., until a testing TM is executed which checks whether the concept is learned. The Model of the Student's
Knowledge is updated according to the TM's conditional probabilities. With both main strategies, the unstructured and
the structured one, it may happen that the student is not able to acquire some concept within the time provided for it. A
sign for this is the insufficient knowledge probability of the concept in the student model ("sufficient" is a probability
10
threshold defined by the Author). In this case the executor invokes the Planner to find a new concept plan, bypassing the
difficult concept. During teaching there is a button available to the student which allows him / her to change the values of
his /her Personal Traits and Preferences in the Student Model. At every selection of instructional method or teaching
material for the next concept, the updated state of the Student Model is taken into account.
non-
structured
Set of Teaching Rules
Discourse Rules
Strategy Selection
Rule(s)
Method - Selection
Rules
TM -Selection Rules
Course Generator Planner
Executor
Select domain, type of link, deepness
Call Planner --> Concept Plan
Select main strategy of course
Select instructional method
Call Planner ---> Task Plan
Select TM
Execute TM
Update Student Model
Other Teaching
Rules
structured
Select current concept from plan
Select main instructional task for it
Figure 5: Course Generation and Execution
Two principle types of re-planning exist, a local plan repair and a global re-planning. Local plan repair means
that only the part of the plan related to the current goal will be changed (see Figure 6). In this way the system tries to find
an alternative way to teach a difficult concept without changing the overall plan. A global re-planning means finding an
alternative plan for the main teaching goal. The discourse rules define which type of re-planning will be chosen.
Local plan repair Global Re-
planning
Concept Structure
teaching goal
problematic
concept
Initial Plan
Figure 6: Two Ways of Re-planning
11
3.5 The Authoring Module
The Authoring Module consists of a TMs-Editor, a Domain Structure Editor and an Editor for Instructional
Tasks and Methods.
The TMs-Editor is a tool that allows "wrapping", i.e. presenting in the way that the system can use TMs
created by any authoring tool for producing multimedia materials. Ready made CAL materials, courses, videos and
graphics can be reused. A unique name is given to every TM and it is associated with one concept or link from the
Domain Structure. In order to be included in the Database, it is needed to classify the TM according to its
pedagogical type (what instructional task it implements) and media and to assign a time allotment for it.
TMs which interact with the student have to be included in the database as a set of "particles": this is a list of
pointers respectively to the "body" of the TM (the question, problem etc.), to the correct answer, to a hint, to an
explanation, a decomposition of the solution into steps. All these are individually accessible by the instructional sub-
tasks. If not all above-mentioned "particles" are present, the corresponding TMs can still be used with the main task,
but some of the task-decomposition methods will not be applicable. For example, let's assume that in the Data Base
for a certain concept A there are two TMs with the pedagogical characteristic "exercise": the first one is represented
as a list of pointers to the following particles:
Ex1 --> (*problem_statement, *correct_answer, *hint),
Ex2 --> (*problem_statement, *correct_answer, *hint, *recorded_steps_to_solution,
*explanation_of_solution).
If Ex1 is selected for implementing the task "Exercise" for concept A, only three from the five alternative
task decomposition methods of the sub-task "Remedy" are available (see Figure 7). The two other methods would be
applicable with Ex2 because it has the needed particles to implement all of their sub-tasks.
Remedy
Give
Correct
Solution
Retry
M3
M1
M2
Give a hint
Remedy
Give
Correct
Solution Retry
M3 M1
M2
Give a hint
Explain
M4
Elaborate on
sub-problems
M5
Task-decomposition methods
applicable with Ex1 Task-decomposition methods
applicable with Ex2
Figure 7: Applicability of Different Task-Decomposition Methods
12
So, in order to enable the system to execute instructional sub-tasks which give a feedback to the student's
errors, the Author has to create for every exercise an associated remedial material. This could be a hint, an
explanation, a step-wise solution, a leading question etc. They will be presented to the student by the instructional
tasks in the prescribed order by the task-plan.
At least one test-atom should be created for every node and link, so that the system can judge from the
student's success or failure whether he / she knows the associated concept. To provide means for the system to
evaluate the degree of knowledge for each of the concepts / topics addressed by a given test-atom, the Author has to
define a likelihood vector containing the probabilities of the student knowing each of the involved concepts / topics,
if he / she answers correctly to the test-atom. Even though there is no guarantee that the probabilities given are
adequate, we suppose that it is not hard for the Author to give approximate estimations of the probabilities, for
example: "If the student answers correctly to test atom A, in 85% of the cases this means that she knows concept X
and in 90 % - that she knows concept Y".
The Domain Structure Editor is a graphical editor which allows developing, extending and modifying the
Domain Structure. It supports creating, deleting and switching between aspects; for a selected aspect it allows to
insert, delete and move, name and re-name nodes on the screen; to insert, delete and connect links; to represent the
different semantics of the links with different colors; to view the existing teaching materials in the data-base and to
associate them with the nodes and links from the Domain Structure.
The Editor for Instructional Tasks and Methods is similar to the Domain Structure Editor. It allows creating,
deleting, and modifying of instructional task-structures. Alternative task-decomposition methods are represented by
linking the task-nodes with arcs which have different colors, thickness and pattern. Every leaf-node (not decomposable)
sub-task is provided with a list of the appropriate pedagogical types of TMs which can be presented. The majority of task
hierarchies and decomposition methods are generic and defined in advance. The editor allows the author to define
subject-specific instructional tasks.
4. Implementation of the System in an Engineering Domain
In engineering domains it is important to distinguish between the basic physical concepts (e.g. power, force,
mass, velocity, electricity, voltage, current etc.) and concepts corresponding to certain devices, parts of devices,
functions etc. It is also important to distinguish between different types of relations, corresponding to their different
semantic (e.g. aggregation, abstraction, analogy, causal, temporal and spatial). A plan of a course is a sub-graph with
given properties, e.g. starting with a goal - node, terminating with certain leaf- or known by the student nodes,
following links with a certain semantic and meeting certain optimality criteria. Different teaching goals can be
achieved by taking different views of the same knowledge. In order to generate a meaningful course, the Planner
needs knowledge about the semantic of the links that have to be followed, so it has to be assigned explicitly by the
Teacher. This, however, would be difficult, since the Teacher will need to know not only what concepts are there, but
also all different types of links between them and how they can be traversed for a specific teaching goal. A way of
organizing the Domain Structure is needed to separate the different semantic views over the subject. This is where a
decomposition of the Domain Structure into smaller relatively independent sub-graphs or "aspects" (see section 3.1.)
is particularly useful.
13
4.1 "Aspects" within the Domain Structure
It is impossible to describe a complex engineering system without considering a structured representation
allowing different aspects and levels in the description, (Hewett & Hayes-Roth, 1994). We will use further the word
"aspect" as a synonym for a "view-point", since it seems to us a more general notion. Usually technical systems can
be well described in three aspects: functional, geometrical and structural. In other domains different aspects may be
more relevant, for example, costs, behavior etc. A relatively self-contained description of the domain according to
any of these aspects should be possible.
The links among the concepts in one aspect may have different semantics (aggregation, abstraction or
causal). For example, an aggregation type of links may be used to describe a component hierarchy in the structural
aspect, or generalization links - to describe functional hierarchies, or causal links - to show operation flows in the
functional aspect. There are also cross-aspect links, which have a subject-specific semantic. For example, a link with
the semantic "is performed by"<-->"performs" can connect a given function-node with the corresponding
structure node(s) that implement(s) the function. As an example, the Domain Structure with two aspects - structural
and functional - of a toaster is shown in Figure 8.
Figure 8: The screen of the Domain Structure Editor
14
4.2 Discourse Rules for Course Planning
It is often enough to consider one isolated aspect for generating a course. For example, if the teacher wants a
course on how to assemble a toaster, she should consider the "Structure" aspect and assign the concept "Toaster" as a
teaching goal. The Planner generates a set of possible plans for the course in this aspect, according to the main type
of link in this aspect with the semantic "aggregation" ("a part of"<-->"contains"). The plan includes all
components needed to build a toaster. The Discourse Rules determine which one out of all possible plans should be
selected for execution and how it should be followed (bottom up or top-down).
Results from the field of automatic text generation indicate that in technical domains, devices are most often
described referring to two different aspects: structural and functional. Paris (1993) analyzed encyclopaedia texts
describing technical devices. She found out that the explanations that were targeted at novices (junior
encyclopaedias) usually describe the way of functioning - a "process trace" of the device and include only the
necessary minimum of information about the structure components carrying out these functions. On the contrary,
texts intended for experts (manuals for technicians) focus mainly on the structure and component details - they use
the "constituency schema" (McKeown, 1985). For an individual user whose level of knowledge is somewhere
between these two extremes, the description of a device should be composed as a combination of the process trace
and the constituency schema.
Our course planner is able to generate plans which cover more than one aspect because it can plan with
respect to different types of links, including cross-dimensional, at the same time. However, in order to keep the
description focused, it is needed to select one aspect as a main "back-bone" of the course. Depending on the changes
in the student's knowledge and in the environment restrictions the executor should be able to switch to the other
aspect according to certain rules. Paris (1993) shows a way to use an explicit model of the user's knowledge of the
domain concepts (which is similar to our model of student's knowledge) for generating individually tailored
explanations.
Our student model, similarly to the one proposed by Paris, contains information about the student's
knowledge about basic concepts and about other concepts. The decision how to choose an overall schema of the
course, in analogy with Paris (1993), is taken by using the following discourse rule:
A process trace (i.e. the functional aspect) is selected as a main aspect for planning only in case that all of
the following conditions are present:
1) there exists a functional aspect and a corresponding function to the structural concept (the teaching goal),
and
2) the student has no knowledge about the goal concept, and
3) no knowledge about any super-ordinate concept with respect to a "generalization"- type of link,
4) the plan in the functional aspect does not involve unknown basic concepts, and
5) the student doesn't know most of the functionally important structural components (related with cross-
aspect links to the most interconnected nodes in the function aspect).
If any of these conditions are not fulfilled, a constituency schema (the structural aspect) is chosen as the
main one. The rule for choosing an overall schema is encoded in the set of Discourse Rules. It is an example of a
domain-dependent rule, in contrast with the generic discourse rules examples of which were given in section 3.3.
Once a overall schema for the course is selected, a plan is generated for teaching the goal concept in the selected
aspect. This can be the structural aspect (in case that the constituency schema is chosen) or the functional aspect (in
case that the process trace is chosen).
15
Combining two aspects can be done at a specific decision point in the algorithm of the Executor (see Figure
5). This is the point when a new concept is introduced and needs to be described, i.e. at the point "Select current
concept from the plan". At this point the system has to decide whether to provide structural or functional information.
The decision whether to switch to the other aspect is taken by the same discourse rule and one additional discourse
rule that checks whether there is still the needed time to perform the switch. During the execution of the initial plan it
may become clear that the student model has changed, the selected plan is no longer appropriate and has to be
changed locally or even for the whole course. In case of re-planning the course, the Discourse Rules will select a
different plan, taking into account the new state of the Student Model.
The described discourse rule is activated for the teaching goal "initial acquaintance". There are other
discourse rules for other teaching goals, like "installation", "maintenance", " diagnosis".
5. Evaluation
The platform chosen for the implementation is IBM PC 486 in a MS-Windows environment. The system is
implemented in C++ and OpenScript. ToolBook © Asymetrix is used as an authoring tool for creating the TMs. It allows
a very easy creation of TMs with advanced graphics and animation and permits linking photos, videos, and sound-
tracks. At this stage a prototype of the system has been implemented and tested for teaching about electric toasters.
Even for such a simple device the Domain Structure is quite complicated. It was structured according to three aspects
— structure, geometry and functions. In the functions-aspect 12 functions (nodes) are connected with time-relations of
3 types: "before", "after" and "in parallel". In the structure-aspect there are 18 nodes organized in a
hierarchy connected with links of the type "is a part of". There are 5 cross-aspect links between the structure-
and functional-aspects.
The main part of time (one week) an Author spent for "paper and pencil" development of the Domain
Structure. Editing the Instructional Tasks and Methods took one afternoon (there were only 12 rules) and the Domain
Structure - two days. More time - three days - was spent in editing TMs with ToolBook. Once the data base with the
TMs is ready, the time for automatic generation of a course-plan when a specific teaching goal is assigned is less than a
minute. The Teacher found reasonable 14 different teaching goals divided in 3 groups: initial acquaintance, montage,
maintenance, diagnosis and repair. The length of the generated plans (in terms of nodes and links to be presented)
varied in wide interval, depending on the position of the teaching goal in the Domain Structure and the initial
knowledge of the student. The time-duration of a course varied between 10 and 30 minutes, depending on the length of
the plan and the duration of the selected tasks with the selected TMs.
In order to evaluate the effort spent for creating an hour of instruction, the time spent for authoring has to be
divided by the sum of the durations of all possible courses that can be generated by the system (with all possible
teaching goals). If we take 8 hours as an average duration of a working day, 6 hours - for one afternoon and 20 minutes
as an average duration of a course, we obtain an approximate ratio of 86 hours of authoring for 5 hours of instruction,
i.e. the ratio is 17.2 to 1. This is a quite favorable result in comparison with other authoring approaches for IST and
even for traditional CAL courseware, since the lowest average time of design and authoring for one hour of intelligent
instruction quoted by different authors is around 100 hours. If the Domain Structure allows the generation of numerous
alternative courses for different goals, the extra-efforts for design and editing the Domain Structure are justified. We
believe that authoring with our system is far more effective than authoring in the traditional sense. However, we will be
able to claim that the system is more effective only after experimenting in a more complicated technical domain.
16
Recently, the DCG has been re-implemented in a more modest version on the WWW (Vassileva & Deters,
1997), and has been applied as an authoring tool for creating a WWW-based course of lectures on Computer-Based
Learning, given by the author of this paper at the Federal Armed Forces University. The authoring effort was
approximately the same - 18 hours of authoring for one hour of instruction. The results of the initial tests showed that
the DCG on WWW will not make a “breakthrough” in the paradigm of university teaching, but it can be easily and
usefully integrated in the existing organization. The following main applications were outlined:
Lecture Support, Distance and Continuous Education. The courses generated with the DCG can be used as
additional learning materials (as an interactive script) supporting lectures given regularly or occasionally at the
University. The specifics of our university is that the all students are officers and are obliged to serve in the Army
five years after graduating. Interactive courses on the WWW accompanying the lectures offered at the university
would provide a “umbilical cord” between our students and their Alma Mater. It this way they can deepen and
refresh their knowledge permanently.
Re-use and Sharing of Domains. The distributed architecture of the DCG allows for authors to collaborate and
cooperate in editing domain structures and relating TMs to the concepts / topics. It also allows a reuse of TMs and
domain structures. Libraries of often used concepts / topics and corresponding URLs can be developed. An
electronic domain represented in the DCG, linked to actual documents on the WWW, such as ”hot” scientific
papers, or just textbook explanations can be shared by lecturers teaching the same subject at different universities,
where everyone can make extensions and modifications according to his/her personal view.
Learners as Authors. Modern learning theories point out the positive effects of letting the learner create his/ her own
understanding and knowledge structures feeling: motivation because of feeling ”ownership” of the problem,
development of searching – and meta-cognitive – skills. For this reason often students are left to plan a lesson
themselves, for example, by organizing lectures as seminars. This enables them to create an own view of the
domain, and an ability to search for new information. The DCG as an authoring tool can be used for carrying out
this type of projects. For example, a student or a team can be assigned the task of authoring a certain theme (sub-
domain). The student/ team has to review the literature, to discover important concepts / topics and relationships and
create a domain structure, to create or find related materials on the WWW and to link them to the concepts / topics
in the domain structure. The structure and materials will be discussed and criticized by the lecturer and the class, and
the domain produced in this way could be used later by the DCG for automatic generation of instructional courses
on this theme.
We haven't been able to evaluate the learning effect of the system empirically with enough students and to
compare their results with a control group. However, our first experiences show very positive attitude and we expect
good results especially for introductory courses in basic disciplines read at the University, like mathematics, where the
students are coming from school with very different background knowledge and adaptive composition of the course for
every individual student would have obvious advantages.
6. Conclusions
Dynamic Courseware Generation provides an alternative to the traditional approach for authoring in CAL. Its
main advantages are:
flexibility in the goals of courses. By means of a multi-aspect organization of the subject concepts / topics
and the possibility to define and use different types of semantic links among them the system can decide how
to plan a course for any given goal in an optimal way according to various discourse rules.
17
individualization of instruction. Teaching of complex technical systems requires to take into account that
the students have different level of knowledge, motivation, confidence. By continuous monitoring of the
student's behavior and requesting student's feedback, the system maintains a model of student knowledge,
personal traits and preferences and is able to find alternative course-plans, instructional methods, and TMs
which are dynamically tailored to the student's benefit.
possibility to assign and change the teaching rules of the system. A human teacher must have a clear
metaphor of the mechanism of the system's functioning. She can think of the system as a teaching agent. This
agent can be "instructed" by means of modifying or creating new teaching rules.
possibility for easy authoring, re-use of already developed courseware and of using technical documentation
as a basis for developing Teaching Materials. This is obtained by separation of the Domain Structure from the
actual TMs which allows them to be updated and extended directly without changing the structure of the
domain.
The course-authoring is shared between the Author (designer of the data-base for a particular domain) and
the Teacher. That makes teachers actively involved in the creation of a course without too much efforts and special
knowledge of authoring. The actual authoring process is shifted to a higher level: to represent explicitly the structure
of the concepts / topics and instructional tasks. This is a non-trivial task. However, we believe this is a justified effort
in order to obtain a system able to teach in a variety of ways for a variety of goals.
References
I. BEAUMONT, P.BRUSILOVSKY (1995) Adaptive Educational Multimedia: from Ideas to Real Systems, In
Proceedings of ED-MEDIA'95, Graz, AACE: Charlottesville, VA.
A. BOHNERT (1995) Analyse und Entwicklung pädagogischer Lehrstrategien für ein intelligentes tutorielles System,
Magisterarbeit, Philosophischen Fakultät der Rheinischen Friedrich-Wilhelms-Universität zu Bonn.
P. BRUSILOVSKY (1992) A Framework for Intelligent Knowledge Sequencing and Task Sequencing, In
Proceedings ITS'92, Lecture Notes in Computer Science No. No 608, Springer: Berlin-Heidelberg, pp.499-506.
CALDERHEAD J. (1991) Representations of Teachers' Knowledge, in Teaching Knowledge and Intelligent
Tutoring, P.Goodyear (Ed.), Ablex: Norwood, N.J., pp.269-278.
T. DIESSEL, A. LEHMANN, J. VASSILEVA (1994) Individualized Course Generation: A Marriage Between CAL
and ICAL. Computers and Education, Vol. 22, No.1/2, pp.57-64.
W. EINSIEDLER (1976) Lehrstrategies und Lernerfolg: eine Untersuchung zur lehrziel- und schülerorientierten
Unterrichtsforshcung. Weilheim: Beltz.
M. ELSOM-COOK, C. O'MALLEY (1989) Bridging the gap between CAL and ITS. CITE Report 67, I.E.T., The
Open University, Great Britain, 1989.
A. FLAMMER (1975) Wechselwirkungen zwischen Schülermerkmalen und Unterrichtsmethoden (WSU). In.
Schwarzer R., Steingaden K. (Hrsg.): Adaptiver Unterricht: Zur Wechselwirkung von Schülermerkmalen und
Unterrichtsmethoden. Munchen: Kösel, pp. 27-41.
R. HEWETT, B. HAYES-ROTH (1994) Physic Systems Modeling for Integrated Reasoning, Modularity and Reuse,
International Journal of Expert Systems, vol.7, no.4, pp.359-386.
HORSTMANN, E. (1995) Entwicklung und Implementierung eines Moduls zur Generierung von pädagogischen
Entscheidungsregeln, Diplomarbeit ID 10/95, Fakultät für Informatik, Universität der Bundeswehr München.
18
J. LARKIN, R. CHABAY (Eds.) (1992) Computer Assisted Instruction and Intelligent Tutoring Systems: Shared
Goals and Complementary Approaches, Introduction chapter, pp.1-10, Lawrence Erlbaum Assoc: Hillsdale, NJ.
G. LEINHARDT (1988) Situated Knowledge and Expertise in Teaching. In J. Calderhead (Ed.) Teachers'
Professional Training, London:Falmer, pp. 146-168.
J.-L. LEONHARDT, M. BAKER, C. BESSIÈRE (1991) Towards Integration of Multimedia and Artificial
Intelligence in Courseware Production SHIVA: A Step in European Context. In Proceedings of CALICSE'91,
Lausanne, pp. 35-42.
M. LINARD, R. ZEILIGER (1995) Designing Navigational Support for Educational Software, in Blumentahal B.,
Gornostaev, J. and Unger C. (Eds.) Proceedings EWHCI'95, Moscow, Springer Lecture Notes in Computer Science
No 1015, pp. 63-78.
A. MITROVIC, S. DJORDJEVIC and L. STOIMENOV (1996) INSTRUCT: Modeling Students by Asking
Questions, User Modeling and User Adapted Interaction, 6 (4), pp.273-302.
K. MCKEOWN (1985) Text Generation: Using Discourse Strategies and Focus Constraints to Generate Natural
Language Text, Cambridge University Press: Cambridge.
T. MURRAY (1992) Tools for Teacher Participation in ITS Design, Proceedings ITS'92, Lecture Notes in Computer
Science No 608, Springer: Berlin-Heidelberg, pp.593-600.
N. NILSSON (1980) Principles of Artificial Intelligence, Tioga Press: Palo Alto, CA.
C. PARIS (1993) User Modelling in Text Generation, Communication in AI Series, Pinter: London.
D. PEACHEY, G. MCCALLA (1986) Using Planning Techniques in Intelligent Tutoring Systems, International
Journal of Man-Machine Studies, 24, pp.77-98.
R. SIEGLER (1988) How Content Knowledge, Strategies and Individual Differences Interact to Produce Strategy
Choices. In Interaction among Aptitudes, Strategies and Knowledge in Cognitive Performance. (Schneider &
Weinert , Eds.) Springer: New York, pp. 74-88.
K. VAN MARCKE (1991) A Generic Task Model for Instruction, in Proceedings of NATO Advanced Research
Workshop on Instructional Design Models for Computer Based Learning Environments, Twente, July, 1-4, 1991.
J. VASSILEVA (1990) An Architecture and Methodology for Creating a Domain-Independent Plan-Based
Intelligent Tutoring System. Education and Training Technology International, 27 , 4, pp. 386-397.
J. VASSILEVA (1992) Dynamic Courseware Generation within an ITS-shell Architecture. Proceedings ICCAL'92,
Lecture Notes in Computer Science No 602, pp. 581-591, Springer: Berlin-Heidelberg.
J. VASSILEVA (1995) Reactive Instructional Planning To Support Interacting Teaching Strategies, Proceedings of
AI-ED 95, World Conference on AI and Education, pp.334-342, AACE: Charlottesville, VA.
J. VASSILEVA and R. DETERS (1997) Dynamic Courseware Generation on the WWW. Proceedings of PEG'97,
Sozopol, Bulgaria, 30.05.-1.06.1997, pp. 111-117.
M. VILLANO (1992) Probabilistic Student Models: a Bayesian Belief Networks and Knowledge Space Theory,
Proceedings of ITS-92, Lecture Notes in Computer Sciences No 608, pp. 491-498, Springer: Berlin-Heidelberg.
B. WOOLF (1987) Theoretical Frontiers in Building a Machine Tutor. In Artificial Intelligence and Instruction:
Applications and Methods (G. Kearsley, Ed.), pp.229-267, Addison-Wesley: Reading.
... Personalization rules derive, in part, from teaching practices. For example, one personalized learning system (Vassileva, 1997) generated courses and customized them dynamically to better fit learning progress and learners' preferences based on the teaching rules proposed by Van Marcke (1992). The teaching rules defined how the learning system selected content, teaching strategies, instructional tasks, and materials. ...
Thesis
Current online learning approaches are sometimes criticized for a “one- size- fits -all” approach, low levels of interactivity, and insufficient feedback, which may result in low levels of learning satisfaction and high dropout rates. To mitigate these shortcomings, this study proposed a set of rules to design personalized motivational feedback based on students’ personal achievement goals. The researcher expected this specially designed personalized feedback to be able to improve student motivation and learning outcomes. To examine the effectiveness of such feedback, an explanatory mixed-methods study was implemented, which included two consecutive phases. The first phase was a quasi-experimental study. A 2018 online master’s degree program course offered by a large R-1 University in the U.S. served as the study context. Twenty-eight students were selected as the test group where personalized motivational feedback based on the proposed rules was delivered along with regular instructor feedback. Another forty students were selected as the control group who only received regular instructor feedback. Students’ motivation and perceived satisfaction were measured by using pre and post surveys. Students’ learning performance was measured by using the collected assignment scores after the semester ended. The second phase was a set of post interviews, in which 13 students from the two groups were asked about their perceptions of the impact of the feedback they received and how they used feedback in their learning process during the study. In the first study phase, ANCOVA F test results indicated the post-test scores of learner motivation and perceived satisfaction in the test group were significantly higher than those of the control group. The mean value of the cumulative assignment scores in the test group was somewhat higher than that of the control group, but this difference was not statistically significant based on the results of Wilcoxon Two-Sample test and ANCOVA F test. In the second study phase, the post-interviews showed that students in the test group expressed more consistently and strongly that they had an overall positive perception of the feedback received in the course. The participants from the test group further explained the underlying mechanism of this personalized motivational feedback was that it affected students’ learning positively by helping them set and regulate learning goals, activate self-regulation mechanisms, and adjust their learning behaviors. Based on the results and the features of the study design, the researcher concluded that the personalized feedback designed by following the set of rules proposed in this study has the potential to improve learner motivation in the online learning context. While its effect on learning outcomes was not significant, the researcher speculated that learning outcomes might have been affected by more complex factors, such as ceiling effects and predominant class structures. The researcher suggested online instructors and instructional designers consider students’ achievement goals when conducting learner analysis and creating learner profiles. She also suggested developers of next-generation LMSs include achievement goals in the learner model and include such rules in a personalization mechanism. One primary limitation of this study was that a ceiling effect on learning performance emerged leading to insufficient variation for the researcher to detect a statistically significant difference in learning performance. Therefore, the researcher suggests future researchers in this area replicate this approach by using automated feedback delivery tools and consider employing personalized feedback in different types of classes and using specific instructional approaches, such as problem-based learning and competency-based learning. Future research should also consider achievement goal’s mediating factors, such as students’ self-regulation skills, in learner analysis.
... It has been established that an AI powered expert module is more efficient than a non AI powered module. With the expert module the tutor can easily redirect his teaching techniques according to data of the student that is stored in the student module (McNally, Shute, & Psotka, 1996;Vassileva, 1989). It is also designed to facilitate teachers in making teaching decisions. ...
Article
Full-text available
Web Learning is a powerful, world-class Virtual Learning Environment VLE that allows, information users like Students Faculty and Research Scholars to engage in meaningful online flexible delivery of courses and assessments to on-site or distance learners, while meeting their specific needs and learning styles. this concept represented multifaceted ideas. Transactional distance consisted of four dimensions—instructor-learner, learner-learner, learner-content, and learner-interface transactional distance. The results inform researchers and practitioners of Web-based instruction concerning the factors of transactional distance that need to be taken into account This article describes a year-long application of critical information literacy theory for social-science-related library workshops. Each of these workshops had a customized section that included working with special collections and university archives. The students who participated ranged from incoming freshmen to seniors at Stanford University. The curriculum development method for these workshops.
Chapter
Ubiquitous internet access over the past decade has provided opportunities to learn languages online. However, most of the online language learning studies focus on higher education perspectives, while the related learner contribution on the commercial platform has yet to be explored. This study investigates Chinese adult learners' perception of learning English on the online course platform through a qualitative research approach. The semi-structured interview questions comprise five parts: learners' background, learning situation (frequency and time spent), assessment of instructors and teaching materials, preference of delivery model (with traditional classroom compared), and their attainment from the online learning platform. These findings can help platform operators better understand the perception of Chinese amateur adult learners and fill the literature gap to provide educators with insights and references on current research agendas.
Article
Full-text available
This paper investigates how humans adapt next learning activity selection (in particular the knowledge it assumes and the knowledge it teaches) to learner personality and competence to inspire an adaptive learning activity selection algorithm. First, the paper describes the investigation to produce validated materials for the main study, namely the creation and validation of learner competence statements. Next, through an empirical study, we investigate the impact on learning activity selection of learners' emotional stability and competence. Participants considered a fictional learner with a certain competence, emotional stability, recent and prior learning activities engaged in, and selected the next learning activity in terms of the knowledge it used and the knowledge it taught. Three algorithms were created to adapt the selection of learning activities' knowledge complexity to learners' personality and competence. Finally, we evaluated the algorithms through a study with teachers, resulting in an algorithm that selects learning activities with varying assumed and taught knowledge adapted to learner characteristics.
Chapter
Full-text available
An ideal face-to-face tutor learner interaction aims to offer learning to the learner in a manner that best suits an individual learner's learning level and learning style. This ability of differentiated instruction has been built in Seis-Tutor Intelligent Tutoring system, developed to offer subject matter knowledge of ‘Seismic Data Interpretation,' a field of geo-physics. The detailed architecture of learner-centric curriculum sequencing module, built to this effect, with its components, sub-components, their interconnected functioning, to generate exclusive learning path, have been described. An algorithm for learner-centric curriculum sequencing, a mathematical model and proposed implementation using a case study has been elaborated.
Article
Face to face human tutoring in classroom environments amply facilitates human tutor-learner interactions wherein the tutor gets opportunity to exercise his cognitive intelligence to understand learner's pre-knowledge level, learning pattern, specific learning difficulties, and be able to offer course content well-aligned to the learner's requirements and tutor in a manner that best suits the learner. Reaching this level in an intelligent tutoring system is a challenge even today given the advanced developments in the field. This article focuses on ITS, mimicking a human tutor in terms of providing a curriculum sequence exclusive for the learner. Unsuitable courseware disorients the learner and thus degrades the overall performance. A bug model approach has been used for curriculum design and its re-alignment as per requirements and is demonstrated through a prototype tutoring recommender system, SeisTutor, developed for this purpose. The experimental results indicate an enhanced learning gain through a curriculum recommender approach of SeisTutor as opposed to its absence.
Article
An ideal face-to-face tutor learner interaction aims to offer learning to the learner in a manner that best suits an individual learner's learning level and learning style. This ability of differentiated instruction has been built in Seis-Tutor Intelligent Tutoring system, developed to offer subject matter knowledge of 'Seismic Data Interpretation,' a field of geo-physics. The detailed architecture of learner-centric curriculum sequencing module, built to this effect, with its components, sub-components, their interconnected functioning, to generate exclusive learning path, have been described. An algorithm for learner-centric curriculum sequencing, a mathematical model and proposed implementation using a case study has been elaborated.
Article
Full-text available
This paper presents a further development of our architecture for dynamic courseware generation which allows explicit representation of teaching expertise. The instructional course is generated automatically for a given teaching goal and can be dynamically changed according to specified teaching rules to suit better the student's individual progress and preferences. The architecture of the system is based on explicit representation of the structure of the domain concepts and relationships (what has to be taught) and the instructional tasks and methods (how to teach). The separate representation of the actual teaching materials allows better flexibility and individualisation, as well as easier updating and re-use of ready CAL materials. In this way our approach provides an alternative to traditional CAL-authoring. An implementation of this approach in a simple engineering domain has been done and an attempt for evaluation of the advantages for authoring is presented.
Chapter
Full-text available
In the past 20 years, views of the interaction between knowledge and strategies have shifted radically. In the 1960s and early 1970s, the role of generally useful strategies such as rehearsal and organization were emphasized. Little if any attention was paid to knowledge of the content on which the strategies were performed, for example, the numbers or words that children were trying to remember. Individual differences were conceptualized at a quite general, traitlike level that did not often make contact with either strategies or content knowledge.
Article
Preface Introduction 2. Discourse structure 3. Focusing in discourse 4. TEXT system implementation 5. Discourse history 6. Related generation research 7. Summary and conclusions Appendices Bibliography Index.
Article
This chapter treats instruction as a generic knowledge-based task. In this view, the domain knowledge about the instructional process will be separated from the knowledge about the domain to be taught. The chapter presents an architecture for capturing this instructional knowledge, based on generic instructional tasks, generic instructional methods, and generic instructional primitives. The instructional process is generated by instantiating those entities in a given context or teaching situation. The instructional tasks embody the tasks to be realized by an instructor in an instructional process. They are defined in terms of a description of the contexts in which they can be instantiated, and a set of alternative methods that can be used for their realization. An instructional method is a knowledge-based representation of a method or procedure that can be used for executing a task. It contains conditions to determine when a method is to be preferred, or whether a method has succeeded in realizing a task. Finally, the instructional primitives are the primitive objects that are manipulated by the methods during the instructional process.
Article
An architecture of a “slightly-intelligent” Computer Assisted Learning (CAL) system is proposed, able to carry out a rather rigid, goal-driven, presentation-oriented style of teaching. However, it is able to generate individualized courses of instruction for students with differing background knowledge and with differing purposes. Already developed CAL courseware is used as a “raw material” for every individual course: thus the architecture builds upon an existing database of learning materials that will be used as a source of atomic course-pieces. The architecture includes domain-knowledge representation at three levels: a level of the teaching goals, a concept-level and a presentation level, containing the database of learning materials. An individualized course is generated dynamically out of the learning materials following the links in the concept-structure to meet the teaching goals assigned by the teacher. Individualization is obtained at three levels: in the teaching goals, in the selection of the concepts to be presented and in the selection of learning materials for a given concept. Three different types of student models are used for this purpose.We believe that this architecture is practical in cases when there is much available learning material and students with different backgrounds who need various and frequently changing teaching goals.