Student-adaptive educational systems Haiying Deng ICS UCI Papers for today • Methods and techniques of adaptive hypermedia (Brusilovsky, P) • MetaDoc: An Adaptive Hypertext Reading System (Boyle, C. and Encarnacion) • Using Bayesian Networks to Manage Uncertainty in Student Modeling (Conati, C. et al ) Methods and Techniques of AH Peter Brusilovsky HCII, School of CS Carnegie Mellon University Outline • Overviews of AH • Methods and techniques of Content Adaptation • Methods and techniques of Adaptive navigation support Definition of AH • All hypertext and hypermedia systems which reflect some features of the user in the user model and apply this model to adapt various visible aspects of the system to the user. • Adaptation techniques refers to methods of providing adaptation in existing AH systems. • Adaptation methods are defined as generalizations of existing adaptation techniques. Adapting to what – Knowledge: overlay model or stereotype model – User’s goal: similar to the overlay model hierarchy (a tree) of tasks – Background and experience – preference Methods of content adaptation • • • • • Additional explanations Prerequisite explanations Comparative explanations Explanation variants Sorting (the fragments of info by the relevance) Techniques of Content Adaptation(1) • Lower level: conditional text – all possible info is divided into several chunks of texts, which is associated with a condition on the level of the user – the info chunk presented only when the condition is true – ITEM/IP, Lisp-Critic, C-book Techniques of Content Adaptation(2) • Higher level: stretchtext – replace the activated hotword extending the text of the current page. – Collapse the non-relevant stretchtext extension, uncollapse the relevant ones. – Collapsed and uncollapsed hotwords can be transferred with each other – KN-AHS Techniques of Content Adaptation(3) • page variants techniques: two or more variants of the same page with different presentations of the same content for different user according to the user stereotype – ORIMUHS, WING-MIT, Anatom-Tutor, C-book. • Fragment variants: variants of explanations for each concept -- Anatom-Tutor • Combination of the two above: Anatom-Tutor Techniques of Content Adaptation(4) • Frame-based technique: info about a concept in form of a frame, frames forms a slot, slots forms a scheme. Slots or schema chosen by some rules. – Hypadapter and EPIAIM – PUSH: a combination of stretchtext and frame-based technique, which has its own entity type of info, similar to frame-based model and a interface similar to MetaDoc stretchtext interface. Methods of adaptive navigation support(1) • Global guidance: – give suggestion at each step of browsing about the next link: WebWatcher – Adaptively sort all the links from the given node according to the global goal: Adaptive HyperMan and HYPERFLEX Methods of adaptive navigation support(2) • Local guidance: – Similar to the global guidance, but different in terms of the local goal, based on the preferences, knowledge and background Methods of adaptive navigation support(3) Local orientation support: to help the user in local orientation - providing additional info about the current node - Limiting the navigation opportunities and let user concentrate on the most relevant links Methods of adaptive navigation support(4) • Global orientation support – Help understand the overall structure of the hyperspace and the user’s absolute position. – Instead of visual landmarks and global maps directly, provide more support by applying hiding and annotation technology. – Providing different annotation based on the knowledge level. Methods of adaptive navigation support(5) • Managing personalized views: – Protect users from the complexity of the overall hyperspace by organizing personalized goal-oriented views, each of which is a list of links to all relevant hyper documents – BASAR Techniques of adaptive navigation support(1) • HYPERFLEX: provides with global and local guidance by displaying an ordered list of related nodes. • Adaptive HyperMan: inputs including user background, search goal interest of current node, etc, outputs an ordered set of relevant doc. • Hypadapter: use a set of rules to calculate the relevance of links for each slot. Techniques of adaptive navigation support(2) • HyperTutor and SYPROS: use rules to decide the visible concepts and nodes based on the concept types, the types of links to other concepts and the current state of user’s knowledge. • Hynecosum: supports both goal-based and experience-based methods of hiding using hierarchies of tasks. Techniques of adaptive navigation support(3) • ISIS-Tutor, ITEM/PG and ELM-ART: support several methods of local and global orientation support based on annotation and hiding, links to the concepts with different educational states are annotated differently using different colors. • HYPERCASE: only known example of map adaptation: supports local and global orientation by adapting the local and global maps Summary • Identified seven adaptation technologies for AH: – – – – – – – adaptive text presentation Adaptive multimedia presentation Direct guidance Adaptive sorting Hiding Annotation of links Map adaptation MetaDoc: An Adaptive Hypertext Reading System • Craig Boyle •Antonio O Encarnacion Overview Simple online text documentation: fixed organization. Hypertext: present through link selection Adaptive Hypertext: actively participate the reading. Adaptivity • Extends the conventional flexibility of the hypertext from the network level to the node level. • MetaDoc: Stretchtext Example: Stretchtext User model • Adapts to the reader, instead of a document • Contains a representation of the reader’s knowledge. • Participates in the reading process. Related work • “Stretchtext”: (Nelson, 1971)change the depth of the information in a node. • Stretching: replace the whole node , similar to GOTO links • Replacement-buttons • DynaText: limited form of stretchtext. MetaDoc to other doc forms • User Modeling: active document • Stretchtext: three dimensional reading and writing • Hypertext: non-sequential reading and writing • Online Documentation: hierarchical retrieval • Printed Text: linear reading and writing Interactive Agent • Store the knowledge about the reader • Used to vary the level of detail in the doc. User level and levels of information • Users and Stereotype: novices, beginners, intermediates or experts based on the knowledge of Unix/AIX and general computer concepts. • Concept levels: the same as above. • MetaDoc varies the amount of explanation or detail info to present the correct level of info based on the internal stereotype info of a concept and the reader’s knowledge level. MetaDoc document • Choose different versions of a single node manually or automatically • Selectively adjust parts of the node instead of adjusting the whole node Writing Stretchtext • • • • Smooth transition Familiar landmarks for different levels Common node identifiers Be ordered Stretchtext in MetaDoc • Vary the info in terms of either explanation or amount of detail • Choose the embedded and appended stretchtext: less confusing • Selected by mouse operations which is context-sensitive and recursive Default presenting rules • Explanation of concepts associated with higher levels are automatically provided for lower level users. • Explanation of concepts associated with lower levels unnecessary for higher level users are suppressed. • Higher level details not necessary for understanding a concept are suppressed for lower level users • Details of equal or lower level concepts are automatically displayed for higher level users. Architecture of MetaDoc(1) • 3D Document component: determines the final form of the node presented to the user and receives commands from the user, composed of the Document Presentation Manager and the Base Document Architecture of MetaDoc(2) • Intelligent Agent: dynamically keeps track of the user knowledge level, automatically matching the presented info depth to the user level, composed of a user model and the inference engine • Domain Concepts: bridge the gap between the above two User Modeling • Explicit modeling: give user the option of explicitly changing the user model within the session • Implicit modeling: stretchtext operation: request for more or less explanation command for less or more detail Evaluation MetaDoc • Evaluated with respect of comprehension and location of specific info. • Compared three systems: MetaDoc, hypertext-only and stretchtext versions. MetaDoc evaluation Reading comp. time (sec.) Reading comp. correct Mean search time (sec.) Hypertext Stretchtext MetaDoc Expert Novice Expert Novice Expert Novice 1780 1930 1250 1780 810 1420 5 3 6.5 7 7 7 755 725 645 530 555 575 Significant results of MetaDoc evaluation Discussion of results • Users of AH doc spent less time answering the comprehension questions correctly • Users of adaptive documents spent less time answering search and navigation questions • MetaDoc had greater impact on novice users than experts. Conclusion • MetaDoc provides an environment in which the user read a hypertext document that will adapt to his/her needs. • Can Help improve readers’ performance. Using Bayesian Networks to Manage Uncertainty in Student Modeling • CRISTINA CONATI • ABIGAIL GERTNER and KURT VANLEHN Andes system’s main contribution • Provides a comprehensive solution to the assignment of credit problem for both knowledge tracing and plan recognition • supports prediction of student actions during problem solving, Problem solving interface • Provides two kinds of help: – Error help – Procedural help Example studying interface • Under SE-Coach which gets the students to self-explain examples – Step correctness: by Rule Browser – Step utility: by Plan Browser Andes’ approach to student modeling Issues in real world(1) • • • • • 1. Context specificity 2. Guessing 3. Mutually exclusive strategies 4. Old evidence 5. Errors Issues in real world(2) • • • • 6. Hints 7. Reading latency 8. Self-explaining ahead 9. Self-explanation menu selections Networks of Andes • Data structure: solution graph • Knowledge-based model construction approach – For problem solving: all the correct solution – For example studying: one single solution R-try-Newton-21aw: if the problem’s goal is to find a force then set the goal to try Newton’s second Law to solve the problem R-normal-exists: If there is a goal to find all forces on a body And the body rests on a surface Then there is a Normal Force exerted on the body by the surface. Encodings • Givens: (SCALAR (KIND MASS)(BOD Y BLOCKA)(MAGNI TUDE 50)(UNITS KG)) • Problem goal: (GOAL-PROBLEM (IS FINDNORMAL-FORCE)(APPLIED-TO BLOCK-A)(APP LIEDBY TABLE)(TIME 1 2)) • Sub-goals Structure of the networks(1) • The domain-general part: represents student long-term knowledge Structure of the networks(2) • Task-specific part: Modeling for Problem Solving • Errors of Omission and Errors of Commission • Updating the student model after a hint • Using the network to generate help Modeling for Example Studying(1) • P(RA= T | all parents = T) =1 - a – address the issue of self-explaining ahead – represents a student’s tendency to selfexplain an inference as soon as she has the knowledge to do so Modeling for Example Studying(2) • The student’s reading time: Low, ok, long – The longer to view an example item, the higher prob. to self-explain it – P(RA=T | Rule=T, All preconditions=T, Read Є {LOW,OK}) =1 - a Modeling for Example Studying(3) • Use of the self-explanation Menus: the higher the number of wrong attempts, the higher the P(SE=T | Context-rule = F), which implements that in this situation it is more likely to achieve the correct action through random selection in the tools rather than reasoning Modeling for Example Studying(4) • Use the student model to support selfexplanation: if the model contains the certain proposition nodes with prob. Lower that the threshold for self-explanation, prompt the students to explain further or read the lines more carefully Evaluation of Andes’ • Machine learning style evaluation: 65% • Evaluation with real students: 1/3 of a letter to 1 letter grade better than the control group • Evaluation of the student model for example studying: Discussion • Empirical evaluations of the resulting coaches indicated that students learned more with them than with conventional instruction. • How did Andes achieve the success: accurately represent the probabilistic dependencies in the task domain.