DRAFT VERSION Bayesian Network to Manage Learner Model in Context-Aware Adaptive System in Mobile Learning Viet Anh NGUYEN, Van Cong PHAM University of Engineering and Technology, Vietnam National University, Hanoi vietanh@vnu.edu.vn Abstract Ubiquitous learning is of interest to many researchers and developers to build adaptive course for learners at anytime and in anywhere To create personalized learning content suitable for each learner, one challenge is to manage and evaluate the learning model, known as the learner's profile Our previous study represented a model of CAMLES [1] system which is a personalized context – aware adaptive system in mobile learning to support students to learn English as a foreign language in order to prepare for the TOEFL test as a case study in Vietnam This paper represents how to apply Bayesian Network in order to manage learner model which is a key factor to determine the learning content adaptation for the learner’s demands and knowledge of individual learners Uncertainty factors used to determine the level of understanding of learners for each concept in the content model Keywords: Bayesian Network, context-awareness, adaptive system, CAMLES Introduction The development of mobile devices, especially mobile phones is booming now Especially there is an upward trend in mobile applications in the areas of modern life such as communication, entertainment, banking and education In particular, academic knowledge on the mobile phone is an important direction of research in the field of e-learning One of the benefits of mobile learning (m-learning) is the ability to provide and access learning materials at any time in anywhere CAMLES system, our model focuses on the identification of learning content through the learning model based on context and awareness of learners Using this system, users can easily browse through the course content to adapt as they wish The learner model is based on the context parameters and learner’s demands The level of knowledge of learners is represented by a set of discrete values corresponding to the results of the tests Disadvantages of the model represented by discrete values not fully represent several learner models, especially when they depend on learner’s knowledge; the level of knowledge is constantly changing during learners participating in the courses Therefore in the next version of CAMLES, we used Bayesian Network to represent learner model in order to overcome those limitations Recently, Bayesian Network model has been widely used for measuring learner knowledge instead of using vector or concrete value There are several systems applying this model to manage their learner model SQL Tutor [2] presents domain knowledge in terms of many constraints, which are factors for Bayesian making multiple predications about students HYDRIVE [3] models student’s skill at troubleshooting an aircraft hydraulics system Bayesian is used to update the student model, using as evidential student factors In CATs system [4], Bayesian Network used to select new questions for adaptive tests; it was constructed as nodes that measure student’s knowledge and gathers evidence with two kinds of links: aggregation relationships among knowledge variables, and relationships among knowledge and evidential variables MEDIA system [5] managing attitude model, knowledge model are two sub-models of student models Bayesian is constructed for aggregation and prerequisite relationships The rest of this paper is structured as follows Firstly, we will review related researches on context-aware learning depending on location In the next section, the factors used in the context of our model to adapt the course content for each learner as well as CAMLES model will be introduced In the fourth section, we focus on how to apply Bayesian Network to manage learner model of the CAMLES system as well as system prototype implementation Finally, discussions and conclusions are summarized Literature Review Our literature review presents recent context–aware m-learning applications for language learning Especially, those support students to learn foreign languages These applications can be classified into two categories: context-aware locationindependent learning and context-aware location-dependent learning Now, we focus on several typical applications: • TenseITS [6] teaches English language to foreign students through meeting their demands Learner model is designed based on four context factors: location, interruption/distraction, concentration and available time Appropriate learning materials for different learners are selected based on the information represented in learner model • CAMCLL [7], context-aware location-independent learning, teaches Chinese to the students whose language levels are not enough to make conversations in Chinese by supporting appropriate sentences to different learners based on contexts The CAMCLL context includes time, location, activities and learner levels Adaptive engine of it is based on ontology and rule-based matching • LOCH [8], context-aware location-dependent learning, supports students to learn Japanese while involving in real time situations By monitoring the positions of the learners, teachers can establish the communication with the students and guide them The context factor in LOCH system is location 3 Context Aware Mobile Learning Architecture In order to select personalized mobile learning materials based on the context as well as learner’s preferences, we propose architecture with their layers described in Figure It includes several layers: Context-awareness detection layer, Database layer, and Adaptation layer The following brief describe the system components Context Data Context Detection Layer Learner model Context Factors Detection Content model Database Layer Adaptive Rules Adaptive Layer Adaptive engine Fig Context Aware Mobile Learning Architecture 3.1 Context-awareness detection layer The function of the context-awareness detection layer is to identify the context factors such as location, time interval, manner of learning and learner’s knowledge that have impact on selecting adapted learning materials for different learners The core of this layer includes main functions: i) Detecting location, ii) Collecting time interval request, iii) Collecting the learner’s preferences, iv) Testing for learner’s knowledge evaluation 3.2 Database layer Database layer consists of context data, content data, learner’s profile and test Context information includes two categories: learner’s request and learner knowledge The first category is the information obtained from the learner’s request such as location, interval of time to learn and concentration These factors require the learners to fill in before they participate in the course In this model, we define location as a place where learners use mobile devices to take part in the course Each location is described by a corresponding discrete value such as (Bus Terminal: 1; Restaurant: 2; Outing: 3; Campus: 4; and Home: 5) This represents the factors that have impact on learning activities such as concentration level, the frequency interruption as well as available time to learn The lower value indicates that the location affecting context factors is higher, whereas the higher value indicates that impact is lower Interval of time is available time that the learner will spend learning Similarly, we use discrete values to identify the level of concentration The learner can choose one of parameters before participating in the course The concentration parameter is designed to determine the learners’ requirements about concentration on learning while student uses mobile devices to browse the course Three concentration levels are low, medium and high The second category is learner’s knowledge that is assumed to be a context factor because of knowledge level variation In our model, learner knowledge is evaluated in several ways The first one is by several test questions at the first time they participate in the course The second way occurs when learners finish one topic; the system requires them to take several questions in order to test their knowledge on this topic The third way is that learner can change their knowledge levels as desired We use concept-based domain to describe the course content It is represent a graph with node denotes concept, and vertex denotes relationship among each concepts The difference in the use of concepts in the studies is the determination of the unit of measurement concepts Depending on the content areas, different applications and design point of view, measure different concepts such as knowledge units [9], the rules [10], and constraint [2] In our model, we propose concept as definitions below: Definition 1: The concept is a basic unit to present a specific content In the content model, the concept is understood as the smallest unit of course content, in other words it would not exist as a Ci concept which is a part of Cj concept To determine the relationship between these concepts, we propose the prerequisite concept definition as follows Definition 2: Prerequisite concepts: Ci concept is called the concept's prerequisite of Cj concepts when to understand the Cj concepts necessary to understand the concept Ci (Denotes is: Ci Cj) Defining prerequisite describe relations between concepts in the model, we only consider the prerequisite relationships between concepts rather than considering the relationship component is used in some models [11, 12], because in our model composition concept is considered the smallest unit To show information about the level of knowledge of learners, we build a model based on the overlay model We use the probability values to quantify the level of understanding for the learner's concept 3.3 Adaptation layer Adaptation layer includes some functions designed to adapt learning materials for each learner Based on the results of test as well as learner’s background, Learner’s knowledge evaluating component used to identify how learner’s knowledge level is The heart of this layer, learning resource selection component, is used to select appropriate adaptive learning content for each learners according to their learner modeling We designed several rules to choose learning resources from content model as traveling of tree nodes The child node describes detailed information about parent node Learning material is adapted to different learners in two ways The first way is that when learner selects one topic from the suggested list, the content belonging to this topic is adapted based on learner model of different learners The second way occurs when the learners finish a test, the system recommends one or more topics that students need to learn The Rules we used to select learning resources in this model is if – then rules [1] Defended on learner model, the adaptive rules include three elements such as height of tree, number of topic and number of test question The height of tree informs that how information detail is The number of topic denotes the number of child nodes or sub topics of determine topics Having several sub topics, the number of topics will decide how many topics are supplied to different learners Similarly, the number of test questions denotes how many test questions will be required to take after different learners browsing the definite topics How to apply Bayesian Network to manage learner model 4.1 Probability values represent the knowledge level of learners In the model, with each concept we use two state variables to quantify the level of knowledge of learners for the reasons as follows: i) using the overlay model needs the variable to store value that indicated level of knowledge to the learner's concept; ii) Assess the level of knowledge of learners needed for quantify concepts: The notion that knowledge of learners; respectively level learners not understand that concept In the model, we represent each level through a state For each concept C, the two state variables are used to represent their understanding of learners to this It is: • Not_acquired: represents that knowledge level of learner does not acquire the concept • Acquired: represents that knowledge level of learner acquires the concept For each concept C, p (C = not_acquired), p (C = acquired) denotes the probability value representing that the state may be not acquired or acquired the concept C It has: p(C = not_acquired) + p(C = acquired) = Bayesian Network applications [13] quantify the level of understanding for the learner's concept From the general formula for the probability distribution, we determined the quantitative formula of the level of knowledge for Cn concept through the following propositions: The C1, C2, , Cn-1 concepts are the prerequisite concepts of Cn concept Meanwhile the value of quantitative level of knowledge of learners to the Cn concept is determined by the following formula: P(Cn| Cn-1,…,C2,C1) = P(Cn| Pa(Cn)) with Pa(Cn) ⊆ { Cn-1,…,C2,C1} (1) 4.2 How to evaluate learner’s knowledge This section presents the research results, new contributions in building adaptive mechanism to adapt courses to meet learner’s demands to select the concept suggests that learners need to learn during the course participants To quantify the level of knowledge, we build a Bayesian network model, with the inference mechanism to quantify the level of understanding of each concept of learners The adaptive rules [14] are applied on the basis of quantitative values to select the concepts that learners need to learn After constructing Bayesian Network, we carry out reasoning to quantify the level of understanding concepts of learner The goal of this step is to quantify the knowledge level of learners for each concept in each stage of learning for the course, as a basis for adaptive content selection to suit each learner We use two strategies of quantitative reasoning: i) Diagnostic reasoning: This mechanism is used in cases where learners not understand a C concept, to determine the value of the probability of the learner’s understand prerequisite concepts of C concept ii) Predictive reasoning: This mechanism is used in cases to determine the quantitative value of the probability of understanding level of the C concept when the probability that the value of the degree of understanding level of prerequisite concepts is known Rules based to select concepts that learners need to learn This step, we shall select concepts to guide the learners need to learn as well as point out the concepts that can be ignored Learners are allowed to ignore concepts to learn if the learner has to understand the concepts The quantitative value of the probability level of understanding of the learners has been identified in the previous step, which is a basis to determine the learners who understand the concept The problem is how much the learners are deemed to have understood the concept with the probability value In the study by Millan [4], Wei [15] considered if the learner understands the concept of the probability, values are from 0.7 to 1, if he/she does not understand the concept, the probability values are from to 0.3, and unspecified when the value is about 0.3 to 0.7 Our opinion, the choice of threshold in this model is not good because with the identification of such threshold, the concepts are considered equal However, the concept of the course content has different level of difficulty Therefore, assessment of their understanding of learners needs to consider the level of concepts System Prototype Implementation We implemented CAMLES prototype based on J2ME technology In order to use CAMLES, the learners need to download and install applications alone in their mobile phones At this stage, we develop content model which consists of five main topics: Adjectives and Adverbs, Pronouns, Questions, Noun Phrases and Commands Those are considered parent topics for the entire contents of the system Under each topic, there will be corresponding child topic, for example, the child of Adjectives and Adverbs topic are Adjectives, Adverbs The learner inputs context parameters via mobile interface The topic content was adapted to him Finishing this topic, the system suggests some question tests to evaluate learner’s knowledge about topic and shows the test results as well as recommend in the next screen Fig Learner inputs context parameters and adaptive content showed (two left pictures) as well as Test questions for evaluating learner’s knowledge (two right pictures) To examine our experimentation, we designed a questionnaire includes six questions to survey 35 students who used CAMLES system with their mobile phone which supports GPRS or 3G to connect to Internet In order to evaluate our system, students check to one of from to values that was the lowest and was the highest We classify student into three categories: group one includes students who never taken the TOEFL test before, group two contain students who have never taken TOEFL test and get below 450 score (paper test), and group three are students have get above 500 score According to Question and Question 2, the students who are satisfied with system would like to use the system again Question was used to survey learners who choose the context which was/is true to them or not As you see, in Group result, students who never take the TOEFL test before are interested in our system However, Average score of Question is 3.0, which shows that they often choose the context which is not true as they in For example, they can choose Restaurant location while they in class Discussions Our target users are graduate students who intend to take TOEFL test However, this approach can be applied to general learners to study English as a foreign language Our model, context-aware location-dependent learning, adapts learning content according to context as well as learner’s knowledge background To find interests in our system, we compare it with early systems In TenseITS [6] learner’s knowledge parameter only calculate at current stage, so if the learner, from second time, backs to the system with the same context factors such as inputted previously, the adaptive contents are similar In our model, learner’s knowledge background is stored and is evaluated after the students finish the topic The results are basic for calculating learner model value for next time learners use system The CAMLL [7] is also based on learner level to adapt suitable sentences; however, how the learner level updates learning progress has not been specified Conclusions This paper has introduced new version of CAMLES, a context aware mobile learning for supporting Vietnamese students to learn English language to prepare for TOEFL test as well as describe how to apply Bayesian Network to manage overlay learner model It adapts learning materials according to the learner’s knowledge as well as their location, their available time, their concentration At this stage, our learner model is still not distinct for all context cases Therefore, there are several different contexts which have the same value in learner’s model In the future work, we will consider refining the content model as well as adaptive engine in order to match the learner’s requests One notable problem is how to fragment content to display in accordance with the size of the mobile phone, which also needed to be considered Acknowledgements This work is partly supported by a research grant of National Foundation for Science and Technology Development (NAFOSTED) References NGUYEN, V A., PHAM, V C., & HO, S D.: A Context-Aware Mobile Learning adaptive System for Supporting Foreigner Learning English Proceedings of IEEE-RIVF International Conference on Computing and Telecommunication Technologies (2010) Mitrovic, A.: An Intelligent SQL Tutor on the Web International Journal of Artificial Intelligence in Education, 13, 173-197 (2003) Mislevy, R., & Gitomer, D H.: The Role of Probability-Based Inference in an Intelligent Tutoring System User Modeling and User-Adapted Interaction , 5, 253-282 (1996) Millan, E., & Perlrez-De-La-Cruz, J L.: A Bayesian Diagnostic Algorithm for Student Modeling and Its Evaluation User Modeling and User Adapted Interaction , 12, 281-330 (2002) Carmona, C., Millan, E., Perez-de-la-Cruz, J L., Trella, M., & Conejo, R.: Introducing Prerequisite Relations in a Multi-layered Bayesian Student Model Proceedings of International Conference on User Model (pp 347-356) LNAI 3538 (2005) Cui, Y., & Bull, S.: Context and Learner Modeling for the Mobile Foreign Language Learner Science Direct System , 33 (2005) Al Mekhlafi, K., Hu, X., & Zheng, Z.: An Approach to Context-aware Mobile Chinese Language Learning for Foreign Students Proceedings of Eighth International Conference on Mobile Business (2009) Otaga, R J., A.Saito, H N., Yin, N A., Yano, C., Oishi, Y., & Ueda, Y.: LOCH: Supporting Informal Language Learning Outside the Classroom with Handhelds Proceedings of the 2005 IEEE International Workshop on Wireless and Mobile Technologies (2005) Oberlander, J., O'Donell, M., Mellish, C., & Knott, A.: Conversation in the Museum: Experiments in Dynamic Hypermedia with the Intelligent Labeling Explorer The New Review of Multimedia and Hypermedia, 4, 11-32 (1998) 10.Ritter, S PAT Online: A Model-Tracing Tutor on the World Wide Web Proceedings of 8th World Conference of Artificial Intelligence in Education, (pp 11-17) (1997) 11.Bra, P D., & Ruiter, J P.: AHA! Adaptive Hypermedia for All Proceedings of World Conference of the WWW and Internet AACE, (pp 262-268) (2001) 12.Trella, M., Carmona, C., & Conejo, R.: MEDEA: an Open Service-Based Learning Platform for Developing Intelligent Educational System for the Web Proceedings of Workshop on Adaptive Systems for Web-based Education at 12th International Conference and Artificial Intelligence in Education (2005) 13.Pearl, J.: Probabilistic Reasoning in Expert Systems: Networks of Plausible Inference San Francisco: Morgan Kaufman Publishers, Inc (1998) 14.NGUYEN, V A., NGUYEN, V H., & HO, S D.: Developing Adaptive Hypermedia System Based on Learning Design Level B with Rules for Adaptive Learning Activities Journal of Natural Science, Vietnam Nation University, 25 (1), 1-12 (2009) 15.Wei, F., & Blank, G D.: Student Modeling with Atomic Bayesian Networks Proceedings of International Conference in Intelligent Tutoring System (pp 491-502) Lecture Notes in Computer Science, Springer 4053(2006) ... foreign language Our model, context- aware location-dependent learning, adapts learning content according to context as well as learner s knowledge background To find interests in our system, we compare...Recently, Bayesian Network model has been widely used for measuring learner knowledge instead of using vector or concrete value There are several systems applying this model to manage their learner model. .. after different learners browsing the definite topics How to apply Bayesian Network to manage learner model 4.1 Probability values represent the knowledge level of learners In the model, with each