輔助超媒體科學學習環境中自我調適學習之適性化鷹架系統

Z

Z Z Z Z

㋲

Z

A Novel Adaptive Scaffolding Scheme for Self-Regulated Science

Learning in Hypermedia-based Learning Environments

A Novel Adaptive Scaffolding Scheme for Self-Regulated Science

Learning in Hypermedia-based Learning Environments

Student

Huan-Yu Lin

Advisor

Shian-Shyong Tseng

A Dissertation

Submitted to Department of Computer Science College of Computer Science

National Chiao Tung University in partial Fulfillment of the Requirements

for the Degree of Doctor of Philosophy

in

Computer Science

November 2012

I

A Novel Adaptive Scaffolding Scheme for Self-Regulated Science Learning in

Hypermedia-based Learning Environments

student

Huan-Yu Lin

Advisors

Dr. Shian-Shyong Tseng

Department of Computer Sciecne

National Chiao Tung University

ABSTRACT

Science education aims to build learners' scientific knowledge structure and varied

process skills. The scientific learners who have various prior knowledge and learning

styles usually need various learning processes to master the concepts or skills. This

learning requirement can be fulfilled by Hypermedia-based Learning Environments,

where the free learning environments can provide a non-linear learning process for

various learning needs. In the non-linear learning process, learners can freely select

appropriate learning paths to achieve the learning goal and the diversified presentation

can demonstrate varied process skills. However, the large number of learning choices

provided by this kind of flexible learning environment usually make learning more

difficult if learners lack self-regulated learning (SRL) abilities to decide their learning

processes and strategies. Thus, scaffoldings, which suggest or guide learners when

learners cannot self-regulate their learning, are usually used to help low-SRL-ability

learners. According to previous researches, adaptive scaffoldings, which dynamically

provide learners assistance according to learners' status, can improve learning

performance and facilitate SRL behaviors better than fixed ones, but providing

adaptive scaffoldings would cause heavy loads on teachers. Although some of existing

Intelligent Tutoring System (ITS) approaches can provide adaptive scaffoldings,

applying these approaches in the non-linear learning processes is still difficult. This is

because the diverse portfolios and prior knowledge generated by various processes

cause the teaching strategies more complex than ones for linear learning processes.

Thus, In this dissertation, three subproblems about representing non-linear learning

plans, adapting learning content to diverse learners' requirements, and diagnosing

learners' status by heterogeneous portfolios are defined. For solving these

subproblems, a novel adaptive scaffolding scheme is proposed, where a generalized

finite state machine, a multi-granularity learning content model, and an

ontology-based knowledge structure are designed to solve the three subproblems,

respectively. The evaluation results and the applying cases are also provided in this

dissertation.

trash

Table of Content

Abstract (In Chinese) ...I Abstract (In English) ...II Acknowledgement ...III Table of Content ...IV List of Figures ...VI List of Figures ...VIII List of Abbreviations ...IX List of Symbols ...X

Chapter 1 Introduction ...1

Chapter 2 Preliminaries ...8

2.1 Scientific Inquiry and Scientific Process Skill ...8

2.2 Self-Regulated Learning and Its Scaffolding Survey ...8

2.3 Intelligent Tutoring System Approach Survey ...11

2.3.1 Knowledge-based Learning Planning Support Mechanism ...14

2.3.2 Knowledge-based Strategy Control Support Mechanism ...15

2.3.3 Knowledge-based Learning Diagnosis Support Mechanism ...18

Chapter 3 Novel Adaptive Scaffolding Scheme ...21

3.1 Learning Process Representation Subproblem ...24

3.2 Personalized Content Adaptation Subproblem ...25

3.3 Process Skill Diagnosis Subproblem ...26

3.4 Interoperability of Scaffolding Providers in Adaptive Scaffolding Scheme ...28

Chapter 4 Generalized Finite State Machine ...31

4.1 Definition of Generalized Finite State Machine ...31

4.2 Rule Class Generation ...32

4.3 Extended Model for Non-Linear Learning Process ...34

4.3.1 Adaptive Learning Planning ...35

4.4 Experiment and Experimental Result ...36

Chapter 5 Multi-Granularity Content Model ...40

5.1 Definition of Multi-Granularity Content Model ...40

5.2 Content Version Management Scheme ...43

5.2.1 Content Version Clustering ...47

5.2.3 Content Version Cluster Decision Tree Maintenance Process ...52

5.3 Content Adaptation Process ...53

5.3.1 Satisfaction Measure on the Quality of Media ...54

5.3.2 Satisfaction Score of the Media Parameter ...55

5.3.3 Content Version Reuse Decision ...58

5.3.4 Content Adaptation and Synthesis ...61

5.4 Experimental Result ...64

5.4.1 Result of Actual Experiments ...64

5.4.2 Results of Simulated Experiments ...67

Chapter 6 Heterogeneous Knowledge Diagnosis Model ...75

6.1 Definition of Heterogeneous Knowledge Diagnosis Model ...76

6.2 Domain Ontology in Ability-Centered Level ...76

6.3 Learning Activity Frame ...78

6.3.1 Key Operation Action Pattern ...78

6.3.2 Embedded Assessment Function ...80

6.3.3 Assessment Portfolio ...83

6.4 Diagnosis Proccess ...84

6.4.1 Diagnostic Rules ...86

6.4.2 Diagnostic Report Generation ...90

6.5 Experimental Results ...92

6.5.1 Experimental Plan and Execution ...92

6.5.2 Analysis of Learners' Scores with Prior Knowledge Measures ...93

6.5.3 Assessment Accuracy of the OPASS System through Domain Experts ...98

6.6 Analysis of Learners' Feedback ...99

Chapter 7 Conclusion ...100

Appendix 1. Object-Oriented Learning Activity System ...103

Appendix 2. Personalized Learning Content Adaptation Mechanism ...109

Appendix 3. Online Portfolio Assessment and Diagnosis Scheme ...116

List of Figures

Figure 3.1: Subproblems and related SRL phases ... 21

Figure 3.2: Novel Adaptive Scaffolding Scheme Architecture ... 24

Figure 3.3: Knowledge interoperability between the three scaffolding providers ... 29

Figure 4.1: Designing and Executing Phase of Object Oriented Leanring Activity System .... 35

Figure 4.2: Flowchart of Rule based adaptive learning method ... 35

Figure 5.1: Content Version Management Sheme ... 44

Figure 5.2: Case Decision Tree Construction ... 45

Figure 5.3: CADT based on HP in Table 5.1 ... 52

Figure 5.4: Flowchart of the CADT maintenance process ... 53

Figure 5.5: Experiment results of the automatic dynamic bandwidth detection scheme ... 65

Figure 5.6: DT of different requests and transmission data size based on various bandwidth settings ... 66

Figure 5.7: Comparison among the inadaptation, static adaptation, and PLCAM approaches 67 Figure 5.8: Comparison of (a) the difference of query time; and (b) the difference of satisfaction between the PLCAM without and with the CADT based on different bandwidths and requested MOs ... 69

Figure 5.9: Comparison of (a) the delivery time; (b) the query time; and (c) the satisfaction score between the PLCAM without and with CADT based on 500 KB bandwidth and different requested MOs ... 70

Figure 5.10: Comparison of (a) the average delivery time; (b) the average query time; and (c) the average satisfaction score between the PLCAM without and with the CADT based on different bandwidths and requested MOs ... 70

Figure 5.11: Comparison of (a) the delivery time; (b) the query time; and (c) the satisfaction score between the PLCAM without and with the CADT on random bandwidths [50 KB, 500 KB], random maximum DT [1,8], random requested MOs [1,9], and eight HP data points in Table 5 ... 71

Figure 5.12: Most suitable threshold of rebuilding CADT based on the different amount of nodes in the Block Version Base ... 72

Figure 5.13: Resultant transcoding time of the PLCAM with auto-adjustment scheme ... 73 Figure 5.14: Comparison of (a) query time; and (b) the satisfaction score of the PLCAM

points in Table 5... 74 Figure 6.1: Example of a Partial CM of the Biology Transpiration Experiment... 77 Figure 6.2: Example of a Partial Scientific Process Skill Map for the Scientific Inquiry

Experiment ... 78 Figure 6.3: Flowchart of the OAPDP ... 86 Figure 6.4: Example of the Personalized Diagnostic Report Generated by DRGalgo in OPASS

... 92 Figure 6.5: (a) Learners Practicing the OPASS, (b) Taking the Examination, and (c) Reading

the Diagnostic Report Regarding the Scientific Inquiry Experiment in the Physics

Domain ... 93 Figure 6.6: Statistical Results of Teachers’ Evaluations for Diagnostic Report Accuracies .... 99 Figure 6.7: Statistical Results of the Questionnaire Concerning the Learners’ Satisfaction .... 99 Figure x1.1: Flowchart of designing a learning activity in OOLA system... 104 Figure x1.2: A partial concept hierarchy of freezing in water cycle... 105 Figure x1.3: The partial misconception hierarchy of freezing in water cycle ... 105 Figure x1.4: The teaching strategy of the scaffolding instruction with misconception diagnosis

... 106 Figure x1.5: A part of flowchart of learning activity “The evaporation, condensation and boil

of water” ... 107 Figure x1.6: A part of flowchart of misconceptions diagnosis and remedial instructions ... 107 Figure x1.7: Screenshot of the authoring tool of OOLA system ... 108 Figure x2.1: Operational flow for a user to retrieve the learning content by the PLCAM

system ... 111 Figure x2.2: (a) Adapted content version of LPa; (b) delivered the adapted content version of

LPa for LPb due to the higher similarity; and (c) delivered the content version created by

LPb in advance for LPc ... 112

Figure x2.3: Screenshots of the learning content adaptation process performed by the PLCAM system ... 114 Figure x2.4: Monitoring interface screen of the LCAMS Web server in the PLCAM system

... 115 Figure x3.1: Architecture of the Prototypical OPASS ... 116 Figure x3.2: Assessment Activities of the Web-based Scientific Inquiry Experiment in: (b)

List of Figures

Table 4.1: The pretest-posttest of learning achievement ... 37

Table 4.2: The one-group pretest-posttest t-test ... 37

Table 4.3: The pretest-posttest of learning achievement of high grade group ... 38

Table 4.4: The one-group pretest-posttest t-test of high grade group ... 38

Table 4.5: The pretest-posttest of learning achievement of low grade group ... 38

Table 4.6: The one-group pretest-posttest t-test of low grade group ... 38

Table 5.1: Example of block-level nodes having leveli = 1 ... 43

Table 5.2: Result of applying CVClustering with the cluster parameters (K =5, Ts =0.01, Tm =1.0, Tn =1, Ti=50, Tp=1) based on data in Table 5.1 ... 49

Table 5.3: Result of mapping the numerical value in HP ... 50

Table 5.4: HP in block-level nodes with cluster label classified by attribute, "Sound Precision" ... 51

Table 5.5: HP in ∑ data used for the simulation experiments ... 68

Table 6.1: Illustration with the Description of each KOAP ... 80

Table 6.2: Example Logs of Planning Data ... 84

Table 6.3: Example Logs of Operational Data ... 84

Table 6.4: Example of Three Types in the DR Definition ... 88

Table 6.5: Example of WrongStep($S, $P) Definition Associated with Problem Description, Reason, and Suggestion Description in the OPASS ... 89

Table 6.6: Questionnaire of Learners’ Degrees of Satisfaction of the OPASS System (Five-Level Likert Scale from 1 (Strongly Disagree) to 5 (Strongly Agree)) ... 93

Table 6.7: Summary Statistics for Prior Knowledge and OPASS Measures – Grade 10, Physics domain ... 95

Table 6.8: Correlations of OPASS Scores with Prior Knowledge Measures in TIPS – Grade 10, Physics Domain ... 95

Table 6.9: Correlations of OPASS Scores with Prior Knowledge Measures in TIPS – Grade 10, Physics Domain ... 96

Table 6.10: Summary Statistics for Prior Knowledge and OPASS Measures – Grade 9, Biology Domain ... 98

Table 6.11: Correlations of OPASS Scores with Prior Knowledge Measures – Grade 9, Biology Domain ... 98

List of Abbreviations

AC Action Continuity

ADP Adaptation Decision Process

AR Assessment Rules

AS Action Sequence

CADT Content Adaptation Decision Tree

CAP Content Adaptation Process

CM Concept Map

CSP Content Scaffoldings Provider

CVClustering Content Version Clustering Algorithm

DNF Disjunction Normal Form

DR Diagnosis Rule

DRGalgo Diagnostic Report Generation Algorithm

DT Delivery Time

EO Experiment Operation

GFSM Generalized Finite State Machine

HANd Hierarchical Atomic Navigation Concept HKD Heterogeneous Knowledge Diagnosis model HLE Hypermedia-based Learning Environment

ID3 Iterative Dichotomiser 3

ISODATA Iterative Self-Organizing Data Analysis Technique algorithm

ITS Intelligent Tutoring System

IRT Item Response Theory

KA Key Action

KOAP Key Operation Action Pattern

LAMS Learning Activity Management System

LCI Learning Caution Indexes

LCS Learning Content Synthesizer

LMS Learning Management System

LOR Learning Object Repository

MGC Multi-Granularity Content model

NORM New Object oriented Rule Model

OAPDP Online Assessment Portfolio Diagnosis Process

OC Object Continuity

OOLA Object-Oriented Learning Activity System

OPASS Online Portfolio Assessment and Diagnosis Scheme PLCAM Personalized Learning Content Adaptation Model PPFO Preferred Picture Format Ordering

PSP Plan Scaffolding Provider

SCORM Sharable Content Object Reference Model

SM Skill Map

SRL Self-Regulated Learning

SSP Suggestion Scaffolding Provider

SWV Satisfaction Weight Vector

List of Symbols

Symbol Description Chapters

si A state, denoting a suggested learning activity Chapter 3 and 4

µ

∑ A set of inputs, denoting learner models Chapters 3, 4, 5, and 6

S A set of states, which represent learning activities

Chapters 3 and 4

cj Contents Chapter 3

C A set of contents Chapter 3

el A learning event Chapters 3 and 6

s0 The initial state Chapter 4

δ A transition function Chapter 4

F A set of final states Chapter 4

aij An attribute Chapter 4

N The number of attributes in

µ

fnow A fact of current state Chapter 4

fnext A fact of next state Chapter 4

fai A fact of attribute ai Chapter 4

NLA A set of states denoting lecturing activities Chapter 4

NAP A set of states denoting education applications Chapter 4

NEA A set of states denoting examination activities Chapter 4

F A set of possible feature sets Chapter 5

N A set of nodes of all granularities in cases Chapter 5

ni A node of a content version Chapter 5

Fi A set of features Chapter 5

childi A set of children nodes Chapter 5

leveli The level of granularity Chapter 5

contenti The original content of the version ni Chapter 5

SF A satisfaction function Chapter 5

ℜ The degree of the adaptation quality Chapter 5

CPi A set of concept properties Chapter 5

HPi A set of hardware properties Chapter 5

LPi A set of learner properties Chapter 5

MPi The media parameters Chapter 5

typei The media type Chapter 5

sizei The size of this media version Chapter 5

WV A Weight Vector Chapter 5

wk A weight in a weight vector Chapter 5

E The Entropy Chapter 5

I The information gain Chapter 5

MPset A set of media parameters Chapter 5

Texpected The average expected time of deivering each

requested media-level nodes

Chapter 5

Tused The actual deliver time Chapter 5

MPcandi The candidate MP list Chapter 5

Ttranscoding The estimated transcoding time Chapter 5

TMDT The maximum available delivery time Chapter 5

Tdeliver The estimated deliver time of the ni Chapter 5

ci A concept Chapter 6

Ontology A set of ontology Chapter 6

F A set of frames Chapter 6

fi A frame of learning activity Chapter 6

Ei A set of all learning events Chapter 6

Vi A set of all slot values Chapter 6

CRi The learning status crystalization rule set Chapter 6 P The set of predicates of learning status Chapter 6

DR The diagnostic rule set Chapter 6

CM A concept map Chapter 6

C A set of concepts Chapter 6

ci A concept Chapter 6

R A set of relations Chapter 6

cri A concept relation Chapter 6

SM A skill map Chapter 6

S A set of skills Chapter 6

s1 A skill Chapter 6

sr1 A skill relation Chapter 6

EO All actions that a learner can operate Chapter 6

a1 An action Chapter 6

Ari An assessment rule Chapter 6

Csi A condition setting Chapter 6

Stepi The name of an experiment step in the scientific

inquiry assessment experiment

Chapter 6 Problemi A checking predicate function of Stepi Chapter 6

Chapter 1

Introduction

In science education, learners are asked to have deep understanding and master scientific inquiry skills in science domain [33]. Thus, many kinds of scientific inquiry assessment [34, 101, 102] and learning activities [77] are used to assist learners in understanding the complex knowledge structure and varied process skills. In addition to traditional face-to-face learning activities, Hypermedia-based Learning Environments (HLE) are also used widely to support science education to enhance learning efficacy and balance teachers’ loading [58, 71]. The hypermedia-based learning environments are suitable to the science learning [51] because the free learning environments can provide non-linear learning processes, which can facilitate learners to construct knowledge structures on the basis of their own prior knowledge, and the diversified presentation are suitable to demonstrate varied process skills. However, without any support, most of the learners cannot obtain high learning performances due to the lack of self-regulated learning (SRL) abilities [36, 79], including planning goals, controlling strategies, monitoring performance, and reflecting on status [72]. SRL scaffoldings, which suggest or guide learners to regulate their learning when learners lack abilities to do the SRL well, are widely used to help learners learn in HLE [7].

SRL Scaffoldings are usually categorized into fixed scaffoldings, which are the same documents or suggestions for all learners, and adaptive scaffoldings, which can provide suggestions according to learners' learning status [5]. Azevedo and his colleague [6] apply fixed scaffoldings, adaptive scaffoldings, and no scaffoldings in learners' learning process to evaluate how various scaffoldings can affect learners' understanding of topics and SRL behaviors. In this research, the fixed scaffolding was a list of questions to remind learners to self-regulate their learning, and the adaptive scaffolding was suggestions provided by teachers

according to learners' status when learners cannot regulate their learning well. The experimental results of this research showed that adaptive scaffoldings based on ongoing learning diagnosis can significantly improve learners’ self-regulation and learning performances. However, in the real learning situation, it is costly for teachers to take care of all learners to provide adaptive scaffoldings. Thus, this dissertation aims to propose a Novel Adaptive Scaffolding Scheme, which can adopt teachers' expertise, to provide adaptive suggestions to help learners regulate their learning in the hypermedia-based learning environment. The main difficulty toward this goal is how to adopt teachers' expertise by using Information Technology (IT) to provide appropriate scaffoldings.

The mechanisms used in the IT domain to adopt teachers' knowledge in a system, called Intelligent Tutoring System (ITS), can be categorized into the conventional-program-based approach and the knowledge-based approach. The former uses hard-coded algorithm to simulate teachers’ teaching strategies. However, as an ITS is used wider and wider, more teaching strategies are required to be applied to satisfy various learners' needs. The teaching strategies embedded in the algorithms are difficult to be maintained and acquired, so the cost of refining and maintaining algorithm would grow rapidly. The latter separates the domain expertise, represented by knowledge models, from the inference logics, so the teachers' knowledge can be refined easily without changing program codes. The knowledge model, explicitly representing teachers’ teaching strategies, can also facilitate to maintain and acquire teachers' knowledge. However, the variety of learners' portfolios and requirements in the non-linear learning processes of HLE makes the adopted teaching strategies complex, so the major challenge to apply knowledge-based approaches is how to design a suitable knowledge model to satisfy the teaching requirements of the non-linear learning processes provided by the free HLE. This dissertation defines and solves three subproblems caused by providing

adaptive scaffoldings in the non-linear learning processes, where three knowledge models are proposed in the Novel Adaptive Scaffolding Scheme to solve the three subproblems, respectively. The subprobems, the related ITS mechanisms, and the proposed ideas are listed as follows.

Learning Process Representation Subproblem

The learning process in HLE is non-linear, which usually makes learners difficult to plan their learning goals, because these learners cannot choose the most suitable learning paths to their learning status among the large amount of choices. Thus, in order to suggest learners appropriate learning paths, teachers' typical learning paths for various kinds of learners and rules of selecting learning paths need to be adopted in the adaptive scaffolding scheme.

Existing ITS mechanisms which provide adaptive navigation support [2, 14, 18, 20, 44, 78, 91] can facilitate teachers to generate typical learning plans and suggest learners with the appropriate learning paths. However, for the non-linear learning processes, the typical learning plans need to be complex with many candidate learning paths, and this kind of mechanisms still lack the knowledge model which can fulfill both expressive power and understandability. A Learning Process Representation Subproblem occurs where a knowledge model needs to be designed to satisfy both adequate expressive power to represent teachers' learning path selection knowledge and good understandability for teachers to provide their expertise.

In order to solve the Learning Process Representation Subproblem, the proposed scheme generalize a finite state machine, which is an easy-to-understand model to represent conditional processes, to represent the learning processes. The new model proposed in this

dissertation is named Generalized Finite State Machine, where the states denote learning activities and the extended transition functions express the complex teaching strategies of learning path selection.

Personalized Content Adaptation Subproblem

In the non-linear learning processes in the free HLE, learners have diversified learning progress and even various learning environments and meida. Choosing appropriate content to satisfy learning requirements is a key task of controlling learning for a learner. However, the diverse requirements need to be satisfied by providing huge number of versions for each content, and it is difficult for learners to find appropriate versions by themselves.

Learning recommender mechanisms in ITS [29, 46, 54, 65, 66] can recommend existing learning materials to learners according to learning styles, prior knowledge, and environments, but the huge number of content is needed for the learners' diversified needs. Content Adaptation mechanisms [11, 15, 30, 53, 57, 75, 97, 99] can dymanically generate new content for learners' requirements by fragmenting and recombining original content. However, managing large number of content fragments to efficiently provide learners appropriate content is still difficult, which is called a Personalized Content Adaptation Subproblem.

In order to solve the Personalized Content Adaptation Subproblem, the original content is decomposed into blocks with the inner media and text. Various versions of content can be generated for various requirements by transcoding some media. A Multi-Granularity

Content model is proposed to manage these versions of content, where a version is

represented as three granularities: page level, block level, and media level. When a request is received, the system can retrieve the most suitable version of the page firstly to find the more

or less suitable content version and adapt it to the detailed requirements by replacing the blocks from other versions and transcoding the media if the request time is acceptable. Thus, the content adaptation process from coarse-grained to fine-grained can efficiently retrieve good-enough content and effectively refine it to the suitable content.

Process Skill Diagnosis Subproblem

In addition to the traditional lectures and examinations, an HLE can provide varied learning media and activities, such as virtual laboratories, to enhance the effectiveness of science learning. These activities can generate various learner portfolios to record learners' behavior and performances for monitoring and reflecting on their learning status. However, the various learning paths and diverse learner portfolios in the non-linear learning processes make moitoring and reflecting difficult because learners are difficult to refer to peers' performances and progress.

Existing learning diagnosis mechanisms [17, 38, 45, 56, 62] in ITS domain can effectively diagnose learners' performance by traditional assessment results, but the ideas of these mechanisms are difficult to be applied to diagnose scientific process skills and learning behaviors for the science learning because heterogenerous learner portfolios generated by the scientific learning activities cannot be analyzed by the existing appraoches. Thus, a Process Skill Diagnosis Subproblem is defined as how to manage and organize the heterogenerous learner portfolios and provide learning diagnosis by applying the teaching expertise to these diversified portfolios to assist learners in monitoring and reflecting on learning status.

For the Process Skill Diagnosis Subproblem, because the heterogeneous learner portfolios are difficult to be analyzed and used in learning diagnosis, a Heterogeneous

Knowledge Diagnosis model is proposed where the learners' behaviors in science learning

activities can be identified as correct or incorrect actions by using the proposed key operation action patterns. These heterogeneous learning events are organized into an ontology-based knowledge structure, which is a tree of concepts and process skills with the relations connecting among the nodes. With the relations of concepts and process skills, it is easy to find the causal relationships between events and learning performances. Thus, the high-level learning diagnosis knowledge is easy to be applied to the organized portfolios.

Generally speaking, this dissertation proposes a novel adaptive scaffolding scheme to assist learners in self-regulating their learning in the scientific learning domain. This knowledge-based scheme applies teachers' educational knowledge to diagnose learners' learning status and provide scaffoldings to fit individual learners' needs. In order to evaluate the effectiveness of the proposed novel adaptive scaffolding scheme, three sub-systems based upon the three proposed models, including a Generalized Finite State Machine, a Multi-Granularity Content model, and a Heterogeneous Knowledge Diagnosis model, were constructed and the corresponding experiments were conducted. The results show that the adaptive scaffoldings for planning learning based on Generalized Finite State Machine can significantly improve low-grade learners' learning performances. For supporting learning content selection and adaptation, the proposed Multi-Granularity Content model is more efficient than the previous approaches to provide more appropriate content. For monitoring and reflecting on learners learning status, the scaffoldings based on the Heterogeneous Knowledge Diagnosis model can also improve learners' motivation to understand learning problems.

scaffolding systems and the issues raised for non-linear scientific learning processes are introduced in Chapter 2. In order to overcome these issues, a Novel Adaptive Scaffolding scheme is proposed and described in Chapter 3, where three subproblems about non-linear scientific learning processes and the corresponding models used to solve the problems are introduced. In Chapters 4, 5, and 6, the three models and their evaluation are detailedly described, respectively. Afterward, a conclusion and the references used in this dissertation are provided in Chapters 7 and 8, respectively. Finally, the cases of applying the proposed models to real learning situations are given in appendices.

Chapter 2

Preliminaries

The free learning processes and varied presentation can satisfy more requirements of learners to learn scientific concepts and scientific inquiry capabilities, but the learners' performance might become low because the learners cannot make decisions among the large number of choices in the flexible learning environment without enough ability of SRL. Thus, scaffoldings are necessary to help these learners regulate their learning. In this chapter, the scientific inquiry and process skills are described firstly, and SRL models and the existing scaffolding approaches are also introduced.

2.1 Scientific Inquiry and Scientific Process Skill

Today, Scientific Inquiry-based learning receives widespread attention. The purpose of such learning is to promote students’ knowledge and understanding of scientific ideas as well as how scientists study the natural world [19]. If students possess scientific inquiry skills, they are capable of conducting an investigation, collecting evidence from a variety of sources, developing an explanation from the data, and communicating and defending their conclusions [35]. Scientific inquiry can be considered as a set of process skills that consists of questioning, hypothesis-making, experimenting, recording, analyzing, and concluding, which can be regarded as "hands-on" learning [19, 52]. The knowledge and capabilities of scientific inquiry are multidimensional [19, 34, 92] and can be divided into three types: (1) Substantive Knowledge, e.g., scientific concepts, facts, and processes; (2) Procedural Knowledge, e.g., procedural aspects of conducting a scientific inquiry; and (3) Problem Solving and Integrative Abilities, e.g., the ability to solve problems, pose solutions, conceptualize results, and reach conclusions [50].

2.2 Self-Regulated Learning and Its Scaffolding Survey

defined self-regulation as the planning, monitoring, controlling, and reflecting phases. The behaviors of cognition in each phase are described as follows:

The Planning Phase: A learner activates prior knowledge and plans learning goals and processes.

The Monitoring Phase: A learner monitors what he/she has learned and evaluates the learning performance toward the goals.

The Controlling Phase: A learner controls and adjust learning strategies and materials to achieve the learning goals.

The Reflecting Phase: A learner reflects on and refines the learning strategies and processes to continuously improve the learning effectiveness.

Besides, other researchers also define various models of SRL to describe a learner's cognitive process. Winne and Hadwin [94] posited that learning happened in four phases: task definition, the goal setting and planning, the studying tactics, and adaptation to metacognition. Zimmerman [103] defined that the SRL includes three main phases: the forethought phase, including task analysis and self-motivation beliefs, the performance phase, including self-control and self-observation, and the self-reflection phase, including self-judgement and self-reaction. Although the definitions of all researchers' models are different, the described learning actions in all models are similar and can be mapped to Pintrich's model in general. For example, the task definition, the goal setting and planning phases in the Winne and Hadwin's model can be regarded as the planning phase in Pintrich's model; the studying tactics can be mapped to the control phase; and the adaptation to metacognition phase can be considered as the monitoring and reflecting phases in Pintrich's model.

define what scaffoldings should be provided by teachers while learners cannot successfully complete the tasks of their self-regulation:

In the planning phase, learners plan their learning activities, so teachers need to suggest learners who cannot plan their own learning with typical learning plans.

In the monitoring phase, learners evaluate their learning performances, so teachers need to assist learners in evaluating their knowledge and skills.

In the controlling phase, learners control learning strategy and the presentation of materials, so teachers have to assist learners in determining which content and presentation is appropriate and adapt the content presentation for learners’ needs.

In the reflecting phase, learners should reflect on learning status and find out themselves’ learning barriers, so a learning diagnosis is usually needed to determine how to remedy the learners’ learning barriers.

Azevedo [6] categorized scaffoldings into fixed and adaptive scaffoldings. Several fixed

scaffolding systems were proposed in recent years to help learners be aware of each phase of

self-regulated learning.

Abrami [1] proposed an E-portfolio system to assist learners in planning their learning, where learners could create learning works, set learning goals, upload learning results, and share these plans and results to teachers, peers, and parents.

In order to teach learners to plan their learning and problem solving activities, Ge [28] provided learners prompts of five problem solving steps in a problem-based learning activity.
Shih [80] developed a platform to facilitate learners to plan and monitor their learning schedules, where learners could customize their own learning schedules based on teacher-provided schedule templates and monitor learning time, attempts, and progress.

define relations bwtween these segments to plan the navigation path. The tool can also regenerate the hypertext by integrating the segments according to the navigation plan to facilitate to read.

Siadaty and her colleaegues [82] use competence ontology to model a company's necessary skills and provide workers suggestions of learning goal planning and tools to monitor workers' learning status.

According to previous evaluation, the fixed scaffoldings can make learners be aware of planning and monitoring their learning, but these scaffoldings lack personalized support to address learners’ indivudual learning needs [6]. Adaptive scaffoldings were provided to help learners overcome their barriers of SRL according to learners' status. In the research [6], adaptive scaffoldings provided by teachers can offer learners better learning effectiveness, but the wide use of this kind of adaptive scaffoldings in the real learning environments would cause heavy loads on teachers. ITS mechanisms, aiming to use IT mechanisms to guide learners to learn and overcome their learning barriers, could be solutions to widely provide adaptive scaffoldings without much increase teachers' loads.

ITS approaches can be categorized into conventional-problem-based approachs and knowledge-based approaches. The former develop intelligent logics by hard-coded programs, and the latter aims to seperate the teaching knowledge from system control logics. The following sections introduce these two kinds of intelligent tutoring system approaches.

2.3 Intelligent Tutoring System Approach Survey

Many existing intelligent tutoring systems were designed by conventional programs, where the teachers’ teaching strategies are simulated by using artificial intelligent algorithms.

Planning Support

Some existing learning systems, providing adaptive navigation support to guide learners learning, can assist learners in planning their learning.

Iglesias [49], Su [84] and their colleagues provided learners adaptive learning sequences according to learners’ learning performances and prior knowledge by using reinforcement learning and planning algorithm, respectively.

Hsiao and her colleagues [42] proposed parameterized questions, having attributes such as difficulty and concepts, and applied adaptive navigation support to select questions for learners to enhance learning effect and motivation.

Context-dependent parameters [16] and learning styles, such as Field Independent/Field Dependent, visual/verbal, abstract/concrete, etc. [73], were also used to compute the adaptive learning sequences.

Hwang [48] proposed an adaptive game-based learning system where the learning styles, global/sequential, are used to determine the game sequence.

Flores [24] grouped learners by using high/low prior knowledge and high/low motivation, and provide adaptive tutorials by using the groups.

Shih and her colleagues [81] used online-test to diagnose learners' abilities of concepts and gave adaptive remedial instrucction according to the concept abilities.

Despotović-Zrakić [21] clustered learners by learning styles, such as active/reflexive, sensitive/intuitive, visual/verbal, and sequential/global, and provided adaptive course to each cluster.

Huang and his colleagues [43] used sequential pattern mining to find recommended concept-learning path. In order to provide adaptive presentation, the user-voting approach and Item Response Theory (IRT) were used to determine the learners' ability levels and learning objects' difficulty levels.

Monitoring and Reflecting Support

For helping learners monitor and reflect on their learning processes in the science education, there were plenty of customized virtual laboratories [23, 41, 52, 93, 100] that constructed environments of specific experiments for scientific inquiry to assess learners’ integrated abilities including scientific knowledge and process skills, but constructing a hard-coding virtual experimental environment for each specific experiment was costly and time consuming.

The learning diagnosis mechanisms were defined to assist learners in reflecting on their learning status and finding their learning barriers.

Liu and Yu [64] proposed an Aberrant Learning Detection approach, which finds learners who have low learning performances due to non-cognitive factors by using Learning Caution Indexes (LCI) to detect the difference between the real learning performance and the estimated performance from Item Response Theory (IRT).

Moridis and his colleague [69] constructed an affect recognition system by formula-based method and Artificial Neural Network (ANN) method to predict learners’ mood in online self-assessment and give affective feedbacks after or before assessment.

Gonzalez and his colleagues [31] proposed a math problem diagnosis system, where mistakes in Math solutions are matched by predefined mistake patterns and provide corresponding remedial action suggestions.

Wu [95] proposed an intelligent tutee system to encourage learners learning by teaching in a concept mapping activity. The adaptive prompts are used to elicit learners' reflection on cognition and meta-cognition.

decision processes, which are constructed using previous learners’ solution.

These conventional-program-based intelligent tutoring system approaches could provide learners effective assistance in parts of self-regulated learning processes. However, the teaching environments and subjects are continuously changing, but the teaching knowledge embedded in the systems is difficult to be acquired and maintained. Thus, the requirements of knowledge-based approaches appear to provide learners learning systems having higher maintainability. The following subsections introduce the existing knowledge-based ITS mechanisms, which can support learners to plan, control, monitor and reflect on their learning.

2.3.1 Knowledge-based Learning Planning Support Mechanism

In order to provide learners adaptive learning paths to help plan their learning, some editable adaptive navigation support mechanisms and specifications were proposed.

SCORM Sequencing and Navigation [2] is one of the most popular adaptive learning activity specifications, where teachers can represent their learning strategies as the sequencing rules to control the learners’ learning paths among the learning materials.

Sakurai and his colleagues [78] proposed a dynamic storyboarding to manage didactic knowledge, representing learning sequence templates, and assist learners in planning university subjects. The results showed the system was beneficial for learning.

Clemente, Ramírez and Antonio [18] proposed a rule-based learning diagnosis, which can find appropriate learning materials according to learning objectives, learners’ abilities, and materials’ topics.

Although teachers can represent their teaching strategies as rules by using these approaches, for teachers to take care of the detailed inference of adaptive learning rules is still difficult. Thus, graph-based models were proposed to enhance the understandability of

learning processes by visualizing the designed processes.

LAMS [20] is a user-friendly learning activity planning system, where teachers can design a collaborative learning process for the whole class, but the linear learning process designed by LAMS cannot represent adaptive learning strategies for personalized learning.

The dynamic fuzzy Petri Nets (DFPN) [14, 44, 91] was used to represent the behavior of tutoring agent, where the learning activity contains a main learning sequence. After a post test, the remedial learning contents can be shown if the score of test is lower than the threshold.
Inference diagrams [12, 61] were also used to describe the courseware diagram and support the evaluation of learners’ learning performance. The learning sequence of each learner can be various with different score range after an examination. Similar to the researches of DFPN, adaptive navigation support is only based on single test score and cannot express the learners’ complete learning statuses.

These models can provide adaptive navigation support according to the single test score, but the lack of expressive power make it still difficult to express teaching strategies for complex learning portfolios. Thus, how to facilitate teachers to intuitively design the adaptive learning plan to support learners in planning their learning processes is a critical issue.

2.3.2 Knowledge-based Strategy Control Support Mechanism

For supporting learning strategy control, most of existing approaches focus on content selection and adaptation. Learners having various styles, prior knowledge, learning paths, and learning devices require personalized content presentation to satisfy their learning needs. When learners control their learning, selecting an appropriate learning content is necessary to ensure learning effectiveness. Thus, some learning content recommenders were proposed to assist learners in choosing existing learning content in the repository.

learners’ learning styles in LMS for teachers to understand their learners. Learning styles include attempts of learning, preference of content types and learning types, and navigation styles.

Mampadi and his colleagues [65] proposed an adaptive navigation support system to recommend learning materials according to the learners’ learning styles, which emphasize on Pask’s Holist-Serialist dimension.

Manouselis and his colleagues [66] proposed a collaborative filtering recommender for learning resource, where teachers used multi-attributes ratings to parameterize resources and shared with others. Ghauth and Abdullah [29] also proposed a learning material recommender by incorporating keyword-based content-based filtering and average good-learner ratings.
Klasnja-Milicevic, Vesin, Ivanovic and Budimac [54] proposed a recommender to recommend learning materials by clustering learners according to learning styles and finding habits and interests using frequent sequences mining.

For the ubiquitous learning, Hwang and Chang [46] proposed a mobile learning approach which provides location-based formative assessment to encourage learners to observe the real environment and find the answers.

However, because of various learners' styles, prior knowledge, and learning devices, the number of learners' requirement combination can be large. The mechanisms mentioned above can only select existing content for learners, so the huge number of content versions, which should be prepared, cause the content version management and large search space problem. To cope with this problem, content adaptation mechanisms were proposed to adapt single content to satisfy wide range of requirements.

static adaptation and dynamic adaptation. The former usually generates multiple variants for each content component, attaching a layout description for the presentation of component-based Web content [40, 68, 89]. These static adaptation approaches can reduce download time, but they require preprocessing tasks and greater storage allocation. Another limitation is that they do not take into account the user's preference and the situation of the wireless network.

Many dynamic adaptation approaches, including content structure analysis and context-based adaptations, have been proposed to resolve these issues [25].

A Hierarchical Atomic Navigation Concept (HANd) was proposed by González-Castaño and his colleagues [30] to navigate on small-scale devices, using the content structure analysis approach. In the HANd approach, an automatically generated navigator page is used to indicate some or all elements embedded in a World Wide Web (WWW) page. To generate the navigator page, a Web page is analyzed and fragmented into several separate “clipped” versions with the degrees of importance. According to the ability of the browsing device, the navigator page can determine a threshold of importance degree to control the amount of elements delivered to users.

Based on a similar concept, many fragmentation and summarization processes have been proposed to organize a Web page into a thumbnail representation that indexes detailed information [15], breaks each Web page into several text units [11], and detects the important parts [99] or the interesting fragments in dynamic Web pages [75], thus reducing delivery latency.

However, not all Web pages are suitable for text summarization because summarized statements, as lossy information, may mislead users. To help improve understanding, the

semantically coherent perceivable units of the Web content can be extracted and presented together on a mobile device according to their semantic relationships [53, 57, 97].

However, most of the aforementioned content adaptation approaches need to manage large number of content fragments, and face a combination explosion problem when the number of requirements become large. Thus, how to manage the content versions to facilitate personalized content adaptation still needs to be solved.

2.3.3 Knowledge-based Learning Diagnosis Support Mechanism

Learners in the non-linear science learning process require to monitor and reflect on their learning status of scientific concepts and process skills during the learning activities, including lectures, traditional examinations, and process skill learning activities, such as virtual labortory assessment. Existing learning diagnostic mechanisms can evaluate learners' learning status and provide remedial learning suggestions according to their test results.
Lin and her colleagues [62] used item-concept relations and learners’ correctness of items to calculate the learners’ learning performances of concepts.

Furthermore, Hwang [45] and Heh [38] proposed mechanisms which can determine remedial learning paths by referring learners’ learning performances and the knowledge structure representing as a concept ontology. Hwang [47] further proposed a group decision approach which can integrate multiple experts' knowledge structures by using the rules-based approach. The integrated knowledge structure could be also used to diagnose learners' weak concepts and suggest remedial learning paths.

Afterward, Chu, Hwang and Huang [17] proposed an Enhanced Concept Effect Relationship to represent concepts and their difficulty levels to improve the effectiveness of learning diagnosis.

Although these approaches can effectively provide diagnosis for traditional examination results, the lack of considering learners' learning behaviors in other kinds of learning activities causes the limitation of finding learners' barriers, especially for process skills learning.

Kosba and his colleagues [56] proposed a mechanism, which can automatically generate adaptive feedback for teachers and learners according to the collected learning performances and learning behaviors by modeling teachers' high-level diagnostic knowledge using a rule-based approach.

Mitrovic [67] proposed a constraint-based intelligent tutoring system, where the key action patterns are modeled as rules associated with the positive feedbacks. The system could give positive feedbacks when learners' actions are correct and uncertain.

However, the considered learning behaviors belong to traditional learning situations. The approache is still difficult to be applied to diagnose learners' scientific process skill learning.

In addition to the traditional assessments, in order to facilitate learners to evaluate their process skills, some editable virtual lab systems were also proposed for teachers to design scientific inquiry experimental tests.

Higgins and his colleagues [39] proposed an authoring tool for teachers to construct diagram-based free-response assessment in electronics like logic circuit design. While teachers design the question, they can also set the marking file which is used to input their system for scoring of learners’ answer. The teacher can use this authoring tool to create different diagram-based assessment in electronics.

Yaron and his colleagues [98] proposed an authoring tool to provide teacher the flexibility of adding new chemicals and chemical equations. Learners can operate predefined

devices to conduct a chemical experiment, doing actions like mix chemicals or heat device to observe the reaction.

These mechanisms can provide virtual environments for learners to train their process skills, but the learning diangostic mechanisms were still lacked for evaluating learners' learning behaviors in the virtual experiments.

Current learning diagnostic mechanism can deal with the traditional learning situations well, and the high-level diagnostic knoweldge can also be modeled by using rule-based approaches in previous researches. However, for scientific learning, heterogeneous learning behaviors from process skill training and scientific concept learning are still difficult to be analyzed and applied by teachers' high-level diagnostic knowledge.

Chapter 3

Novel Adaptive Scaffolding Scheme

According to the analysis in the Chapter 2, although the previous studies have proven that the existing ITS approaches can facilitate teachers and learners to teach and learn effectively in specific domain and learning situations, scaffolding learners for each SRL phases among the non-linear scientific learning processes still causes some probems, as shown in Figure 3.1:

Figure 3.1: Subproblems and related SRL phases

Learning Process Representation Subproblem: Because of the lack of the appropriate

learning process representation, teachers are difficult to design the adaptive learning plan to support learners in their planning phase.

Personalized Content Adaptation Subproblem: Due to the lack of content versions

management approaches, efficiently providing personalized content adaptation to support learners in their controlling phase is difficult.

Process Skill Diagnosis Subproblem: Because of the heterogeneous learning behaviors

from process skill training and scientific concept learning, analyzing learners' learning portfolios and applying teachers' high-level diagnostic knowledge to give learning diagnosis to support learners in their monitoring and reflecting phases is difficult.

To cope with these problems, a Novel Adaptive Scaffolding Scheme, including the Generalized Finite State Machine, the Multi-Granularity Content model, and the Heterogeneous Knowledge Diagnosis, is proposed. As shown in Figure 3.2, the Novel Adaptive Scaffolding Scheme includes three scaffolding providers to support learners to self-regulate their learning:

Plan Scaffolding Provider (PSP): A PSP, based on Generalized Finite State Machines

to solve the Learning Process Representation Subproblem, can provide suggested learning plans to facilitate learners to plan their learning.

Content Scaffoldings Provider (CSP): When learners aim to control their learning

materials, a CSP, based on a Multi-Granularity Content Model to solve the Personalized Content Adaptation Subproblem, can manage content versions and adapt learning content to the learners' requirements.

Suggestion Scaffolding Provider (SSP): During the learning process, learners need to

monitor and reflect on their learning. A SSP, based on a Heterogeneous Knowledge Diagnosis Model to solve the Process Skill Diagnosis Subproblem, can give learning diagnosis and remedial suggestion to help the learners understand their own learning status.

All learners' learning portfolios, including records of reading content, test results, and portfolios of process skill training are stored into a Learner Portfolio Database. Learning resources, such as learning content and test items, are stored in the Learning Content

Repository and Test Item Repository, respectively. Besides, learning applications, such as

virtual laboratories, are stored into a Learning Application Repository. These learning portfolios and resources are refered and fired by the three scaffolding providers to provide learning scaffolding services.

In the beginning of learning, PSP can suggest next learning activities to help a learner plan learning processes according to the learning portfolios when planning learning processes. Afterward, in the controlling phase, the learner aims to learn with a learning content, CSP can provide an appropriate learning content according to the learner's portfolio and the planned learning activity. After learning and testing, SSP can support the learner to monitor and reflect on their learning status according to the learning portfolios and provide diagnostic report for the learner to plan the next round of learning. Thus, in the proposed scheme, the interoperability of all scaffolding providers is concerned to scaffold the learner's whole learning process.

Figure 3.2: Novel Adaptive Scaffolding Scheme Architecture

3.1 Learning Process Representation Subproblem

In a non-linear learning process, three kinds of necessary elements should be represented in the learning activity plan: learning paths, learning activities, and the learning-path-selecting strategies. Various learning paths should be designed for various kinds of learners. Among the learning paths, learning activities, such as examinations, lectures, or projects, should be determined. Besides, in the branches of the learning process, the learning-path-selecting strategies based on teachers’ teaching knowledge should be designed to guide learners to select appropriate learning processes. However, designing a model for teachers to design processes and learning-path-selecting strategies flexibly and easily is difficult.

Thus, the Learning Process Representation Subproblem is defined as how to design a

learning plan model, such that

the model can represent the learning paths and learning activities,

executed to conduct learning planning, and

the model is easy to understand and be used to design processes for teachers?

In order to intuitively represent the non-linear learning process to solve this Learning Process Representation Subproblem, a generalized finite state machine is used to model the knowledge of learning activity plans. The Generalized Finite State Machine (GFSM) is designed by generalizing a traditional finite state machine, where a compound input is used to represent multiple attributes of a learner’s status and the rules of the Disjunction Normal Form (DNF) are used in the new transition function to express the teachers’ learning-path-selecting strategies.

3.2 Personalized Content Adaptation Subproblem

In order to fulfill individual learners’ learning styles and learning status, the presentation of a learning material should be various. Existing content adaptation mechanisms can adapt the text or multimedia items to various presentation needs, but how to manage and reuse the adapted presentation versions to efficiently provide learning content is still an important issue. The existing learning content recommender systems consider a learning material as a static item. If a huge number of presentation versions are adapted for various requirements, the recommender should manage all the versions and have large search space for recommending a material for a new learner. If all the materials are stored as the detailed chunks of all adapted versions, the recommendation is still inefficient due to the combination explosion of these chunks for forming a complete learning material.

Thus, the Personalized Content Adaptation Subproblem is defined as how to control

the granularity of the stored content presentation versions, such that

the mechanism can retrieve or adapt suitable learning materials to the learners'

requirements?

According to the preliminary studies, the learning materials stored in a single content granularity cause the inefficient problem in providing content having adaptive presentation to a new learner. Thus, a Multi-Granularity Content model (MGC) is proposed to represent and store the learning materials as multiple granularities. For a new requirement, the coarse-grained learning material fulfilling the most learning needs is retrieved as the main body of the provided content. Afterward, the fine-grained parts of the retrieved material, which are less appropriate for the learner, are replaced by other fine-grained parts to enhance the quality of the adapted content. The content adaptation mechanism adapts learning materials from coarse-grained to fine-grained can prevent the combination explosion problems and the combination costs of detailed chunks.

3.3 Process Skill Diagnosis Subproblem

Learners need to monitor and reflect on their learning status in the monitoring and reflecting phases of self-regulation, so learning diagnosis mechanisms were proposed to support learners to evaluate their own learning. In the linear learning process, all learners’ learning portfolios are homogeneous, so assessing learners’ learning performance and status is easy by ranking or scoring. However, in the non-linear learning process, all learners’ learning processes are various, so how to assess the heterogeneous learning portfolio to provide the learning diagnosis is more difficult than the homogeneous ones. Teachers’ high-level diagnosis knowledge can be generally applied for various learning processes, but the existing approaches only focus on evaluating learners’ results of traditional tests. Without considering learners’ detailed learning behaviors in learning activities, such as operations in a scientific

inquiry experiment, the learning diagnosis cannot precisely capture learners’ learning status and process skills, especially in the science education.

Thus, the Process Skill Diagnosis Subproblem is defined as how to provde learning

diagnosis for scientific learning, such that

the heterogenerous learner portfolios generated in the scientific learning can be

managed and organized,

the teaching expertise can be applied to these diversified portfolios, and
the diagnosis can find learners' weekness of concepts and process skills?

To cope with the Process Skill Diagnosis Subproblem, a middle-level knowledge representation is needed to extract the learners’ learning status from heterogenerous learning events and provide structural learner models for learning diagnosis. Thus, a Heterogeneous Knowledge Diagnosis model (HKD) is proposed where an Ability-Centered Level is defined to connect high-level diagnosis knowledge and low-level learning events. In the Ability-Centered Level, all the learning behaviors and test results are structured for further diagnosis. In the Ability-Centered Level, the background knowledge, including concepts or process skills, is represented as the ontology, where concepts and skills are represented as nodes and the prerequisite relations and extended relations are represented as the relations between nodes. All learning behaviors and test results are extracted and represented as predicates of learning status. For example, after learners get a score 0.8 of a concept c1 in a

test, a predicate is recorded as Score(c1, 0.8), and after reading a lecture about c1 during the

inadequate reading time, the learning behavior is also be recorded as LearningTime(c1,

inadequate). Besides, assume a learner does a wrong operations about the measurement skill

the structured learning status from the heterogeneous learning events, the frame-based knowledge representation is used to model all the learning activities. For example, the frame of a reading activity records the lecture’s expected reading time and its associated concepts. For an experiment-based test, the frame records all necessary and wrong operation patterns and their associated skills and concepts. The embedded rules are defined to transform a learner’s learning events to the predicates in the Ability-Centered Level according to the slots of the frames. Besides, the high-level learning diagnosis knowledege can be represented by using rule-based representation, which can infer learning status and learning barriers from the predicate of learning status and the relations in the ontology of the Ability-Centered Level.

3.4 Interoperability of Scaffolding Providers in Adaptive Scaffolding

Scheme

The three scaffolding providers can be interoperable to provide the complete adaptive scaffoldings for learners. As shown in Figure 3.3, in the planning phase of self-regulated learning, the Plan Scaffolding Provider based on Generalized Finite State Machines can provide the suggested learning activity si ∈ S (Step 2) according to the learner’s learner model

µ

PSP: ∑ × S  S, where ∑ is a set of learner models and S is a set of learning activities.

If the learner takes a reading activity, the Content Scaffolding Provider based on a Multi-Granularity Content Model can adapt the content cj ∈ C (Step 4) according to the

requirements of learning activity si and the learner model

µ

learning in the control phase.

CSP: ∑ × S  C, where C denotes a set of contents.

The learner can read the content cj or take a test in the suggested learning activity si and

reading leanring activity and Learning represents the learning behavior in other learning activities:

Reading: S × C  E, where E denotes a set of learning events. Learning: S  E

Finally, in the monitoring and the reflecting phases, the Suggestion Scaffolding Provider based on the Heterogeneous Knowledge Diagnosis Model can infer new learner model according to these learning events el (Step 5).

SSP: E × ∑  ∑

The three scaffolding providers in the scheme mentioned above can provide adaptive scaffoldings to help learners in their SRL processes. The information of learners' status can be shared among these scaffoldings providers to produce ongoing diagnosis and suggestions. In the following chapters, the three knowledge models used in the three scaffolding providers are introduced precisely.