Review of Knowledge Representation Techniques for Intelligent ...

2016 3rd International Conference on “Computing for Sustainable Global Development”, 16th - 18th March, 2016 Bharati Vidyapeeth's Institute of Computer Applications and Management (BVICAM), New Delhi (INDIA)

5HYLHZRI.QRZOHGJH5HSUHVHQWDWLRQ7HFKQLTXHV IRU,QWHOOLJHQW7XWRULQJ6\VWHP Neha Mendjoge D. J. Sanghvi Mumbai, India Email id: [email protected]

Abhijit R. Joshi

Meera Narvekar

D. J. Sanghvi Mumbai, India Email id: [email protected]

D. J. Sanghvi Mumbai, India Email id: [email protected]

$EVWUDFW±$FRPSXWHUVRIWZDUHWKDWVLPXODWHVDKXPDQWXWRU¶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±,76GRPDLQPRGHONQRZOHGJHXVHU5%&%0

roles such as a source of expert knowledge, a standard for evaluating the performance of student or for detecting errors, etc [1]. Domain Model

Student Model

Tutor Model

User Interface

NOMENCLATURE User

ITS: Intelligent Tutoring System RB: Rule Based ACT-R: Adaptive Control of Thought-Rational MT: Model Tracing Tutors CBM: Constraint Based Model

Fig. 1: Architecture of ITS

I. INTRODUCTION An intelligent tutoring system is computer software designed to simulate a human tutor’s behavior and guidance. The purpose of Intelligent Tutoring System (ITS) is to enable the users, i.e., learners to gain knowledge and develop their skills in a particular domain. Figure 1 shows the traditional ITS model. It contains four components: the domain model, the student model, the teaching model, and a learning environment or user interface. The domain model represents the subject matter expertise and provides the ITS with knowledge of the topic being taught. A problem presented to the user may be solved in multiple ways. The domain model can also be called as the cognitive model or the expert model. There can be various ways and methods to solve a particular problem. These different methods are included in the domain model. This model consists of the various concepts, rules, and problem-solving strategies of the domain to be learned. The domain model can fulfill a variety of

c 978-9-3805-4421-2/16/$31.00 2016 IEEE

The student model is considered as the core component of an ITS. The student model is like a window to the domain model. It provides a unique view of the various concepts incorporated in the domain model. It contains information related to a specific user. The student model provides input to the tutor model [4]. The tutor model, alternatively termed as teacher model or instructor model accepts input from the domain model and student model and makes appropriate choices about the tutoring strategies and actions. The user can request for guidance or hint at any step in the problem solving process. For example, if the user is identified as a beginner, then the tutor model may decide to show a step-by-step procedure to solve a particular problem before the problem is presented to the user. The user interface model, which is also a component of ITS, controls the interaction between the user and the system. It translates information between the system's internal representation and an interface language that is understandable to the user [2].

2508

2016 3rd International Conference on “Computing for Sustainable Global Development”, 16th - 18th March, 2016 To enable learning in a meaningful and effective manner, ITS uses a variety of computing technologies. ITS enables the users to practice their skills by solving tasks in a highly interactive learning environment. To provide such capabilities, these systems must be able to represent the knowledge effectively. It must also have mechanisms by which these representations can be used by the system to solve the various problems in the domain. These knowledge representation techniques viz, Model tracing, constraint based technique, dialogue based technique are reviewed in section II. The analysis of these techniques is presented in section III. The paper ends with conclusion. II. KNOWLEDGE REPRESENTATION ITS differs from the usual computer aided system because they easily adapt to the user’s individual requirements. This is achieved by using the student model. The basis for student model is a rich domain model. The domain model consists of the knowledge of a particular domain, i.e., it represents the subject to be taught. The domain model is common to all the users of the system, while the student model differs for each user. The student model and the domain model are used together to give feedback to the users [4]. To provide appropriate and useful feedback to the user, the domain model must be built effectively and efficiently. There are three traditional approaches used to build the domain model. Let us see these approaches one by one. $5XOH%DVHG0RGHOV Rule based models also called as Cognitive tutor. The basis of Cognitive tutors is the Adaptive Control of ThoughtRational (ACT-R) theory of cognition [5][7]. According to this theory, there are two long-term memory viz., declarative and procedural. This theory says that the human learning goes through various phases. The first phase involves learning declarative knowledge i.e. the factual knowledge such as theorems in a mathematical domain [7]. This declarative knowledge is the overall knowledge of a particular domain. This declarative knowledge is then used by the user to solve problems within a domain. This problem specific knowledge is called as the procedural knowledge. The procedural knowledge is goal-oriented. Production rules are used to represent the procedural knowledge [7]. The Rule Based (RB) system is based on a set of rules wherein the user follows the step by step problem solving process. Cognitive tutors provide response to each step that the users take to solve the problem. Due to this reason, the cognitive tutors are said to generate immediate feedback. When the user step does not match any of the valid rules or it does not match the buggy rule, an error is detected. The rule-based models are built from cognitive task analysis. which consists of producing problem spaces or task models. These problem spaces or task models are built by observing the expert and novice users. Task models represent a set of production rules in which each rule represents an action

corresponding to a task [3]. When a user tries to solve a given task, the user’s reasoning ability is analyzed based on the rules applied by the user, i.e., the user solution is compared step-bystep to the solution given by the expert. This process is termed as Model Tracing (MT). The Model Tracing (MT) tutors follows the rule based model. During teaching, the MT tutors1) suggest to the user, which should be the next step to be taken. 2) provide demonstrations. 3) evaluate the understanding of the user in terms of the skills that the user has applied. 4) infers user goals. These kinds of tutors are basically used, if the learning process is of utmost importance rather than simply checking, whether the user has obtained the correct answer or not. There are many tutors built using this model. For example, the Andes Tutor for Physics [10]. Ill-defined domains are a major limitation of model-tracing tutor. Ill-defined domain means that, for some domains, clear strategies for finding solutions are not available. Thus it can be difficult to define an explicit task model. Moreover, for complex domains, a large number of rules are used to represent the knowledge. So to determine the solution paths and designing a set of rules for a task would be very timeconsuming. Let’s say we want to find the candidate key for the relation given in Figure 2. An attribute is said to be the candidate key iff the closure of the attribute gives all the attributes of the relation. STUFF(H,I,J,K,L,M,N,O) FD1:H,IĺJ,K,L FD2:JĺM FD3:KĺN FD4:LĺO

Fig.2:Relation with functional dependencies

The production rules in cognitive tutor can be given as :

1. IF goal is to find the candidate key and closure of H and I gives all the attributes of the relation THEN the candidate key is H and I.

2016 International Conference on Computing for Sustainable Global Development (INDIACom)

2509

Buggy rule: 1. If the goal is to find the candidate key and closure of an attribute gives all the attributes of the given relation then set the attribute as the candidate key.

For the relation given in Figure 2, the candidate key is a combination of H and I. Thus, if the user enters the answer as ‘HI’, then production rule 1 is satisfied. Hence, the answer is correct. However, a relation can have multiple candidate keys. If the answer entered by the user does not satisfy any of the production rules, then the buggy rule is checked. The buggy rule is satisfied, if the answer entered by the user determines all the attributes of the relation. If none of the production rules or buggy rule is satisfied, then the answer is incorrect. %&RQVWUDLQW%DVHGPRGHOV As seen, Rule-based models capture the knowledge, which is required to generate the step-by-step solutions, while Constraint-Based Models (CBM) express the requirements, which all solutions should satisfy. This means that the rule based model analyses the path through which the solution is obtained while constraint based models analyse only the obtained result. Constraint based models are not concerned with the path chosen by the user. Ohlsson’s theory of learning from performance errors gives rise to CBM [6]. This theory suggests that even when the users have been taught the correct way of preforming a task, they still make some mistakes. The reason for making mistakes is that the declarative knowledge that the user has learnt has not been converted to the procedural knowledge. Due to this, the number of decisions that are taken by the user while performing the procedure is sufficiently large. This leads to mistakes. However, by practicing the task and catching mistakes can help the user in modifying the procedure. This helps the user to incorporate the appropriate rule that has been violated. The process of learning from errors consists of two phases as described by Ohlsson. These two phases are error recognition and error correction. Error recognition means identifying an error. The ITS recognizes an error by using the declarative knowledge. After recognizing an error, it must be corrected. An ITS can perform the role of the mentor to inform the user of the various mistakes, if he/she does not possess the declarative knowledge to identify it. A student often requires the help of a teacher to overcome a problem in his/her own understanding about the particular concept. This can be achieved in CBM by a series of carefully designed sequence of feedback messages. Thus in this way, the CBM reflects the action of a human teacher who helps the student to overcome problems in his/her knowledge. In CBM, a solution is specified by a set of constraints instead of providing an explicit task model. Each constraint consists of a relevance condition and a satisfaction condition. The domain model is thus given as:

2510

IF is true THEN will also be true. It means that, if the user solution falls into the state defined by the relevance condition, it must also be in the state defined by the satisfaction condition [7]. SQL-Tutor is an example of an ITS that uses CBM wherein the domain model consists of over 500 constraints [11]. When the user violates any constraint during solving a task, the CBM Tutor diagnoses that an error has been made and provides help to the user regarding the violated constraint [7]. When a user violates a constraint, it simply means that the user does not know the concept and needs help. As an example of constraints, consider again the example to find the candidate key as shown in Figure 2. The constraints are given as: Cr: closure contains all the attributes of the relation Cs: the attribute is the candidate key Thus in Figure 2, for the functional dependency1 (FD 1), the constraints are given as: Cr: closure of H and I contains all the attributes of the relation Cs: H and I are the candidate key To verify whether the user entered solution is correct or not, consider the above relevance condition, i.e., Cr. If the user enters the answer ‘HI’ as the candidate key for the relation given in Figure 2, then the system checks the relevance condition Cr. If the relevance condition is satisfied, then the satisfaction condition, i.e., Cs is also satisfied. &'LDORJXHEDVHG0RGHOV In Dialogue based tutors, systems hold conversational dialogue with its users. There is an agent that acts like a teacher and converses with the user. In addition to displaying graphics, text and animation, this agent also has synthesized speech, facial expressions and gestures. The dialogue based tutoring systems present challenging problems and questions to the user. The user then types in answers in English, and there is a lengthy multiturn dialogue between the user and agent till complete solutions or answers evolve [9]. Users type their responses in English in addition to the conventional point and click. These tutors help the users to actively develop their knowledge through conversations. The multi-turn dialogue between the agent and user to find the candidate key for the relation given in Figure 2 is represented as:


2016 3rd International Conference on “Computing for Sustainable Global Development”, 16th - 18th March, 2016

Agent: Enter the closure of HI. User: H I J K L Agent: There exists transitive dependency between FD2 and closure entered. Modify the closure. User: H I J K L M. Agent: There exists transitive dependency between FD3 and closure entered. Modify the closure. User: H I J K L M N. Agent: There exists transitive dependency between FD4 and closure entered. Modify the closure. User: H I J K L M N. Agent: Is there any other FD to be processed? User: No. Agent: Does the closure contain all the attributes of the relation? User: Yes. Agent: So the candidate key is HI. Looking at these domain models and how they help user in solving a problem/task given by ITS, now let us analyze these models based on various parameters in the next section. III. ANALYSIS OF DOMAIN MODEL The rule based model also called as cognitive tutor or model tracing model, is a process-centric approach. In this model, the path taken by the user to reach to the solution is evaluated. Constraint based model is a product-centric approach, in which the entire focus is on the state of the solution that the user arrived at irrespective of the steps taken by the user to reach the solution. The rule based tutors represent the domain in the form of production rules, while constraint based tutors represent the domain in the form of constraints, which are declarative knowledge of the subject to be learnt [9]. Model tracing tutors have been criticized for being too rigid because they expect the users to follow a fixed set of approaches to solve problems [8]. Model tracing systems validate the user solution by comparing it with the teacher solution. If alternate solutions exist to a problem, then all the solutions need to be stored in the system. If the user provides a solution, which is not stored in the system, then model tracing tutors term it as incorrect. For example, the use of top-down approach for writing functions is the requirement of the Lisp tutor [10. On the other hand, constraint based tutors do not impose any particular strategy for solving a task. They only represent the current state of the solution. Typical CBM systems validate the user solution by comparing it with the teacher solution. However, if any alternate solutions exist, then they are validated by checking the necessary constraints. Immediate feedback is given in case of cognitive tutors while, in constraint based tutors feedback is given on demand. Cognitive Tutors can offer hints in terms of the next step to be performed in a path as they trace the path to the solution. Hints

are provided to the user whenever the user demands or after the user crossed the error threshold. This error threshold is set by the teacher which indicated the maximum number of mistakes that the user can make. Feedback is provided by constraint-based tutors for the missing elements of the solution. In cognitive tutors, if some production rules or buggy rules are not included then it means that the knowledge base in incomplete. If the knowledge base is incomplete, then the system will term a particular step performed by the user as incomplete if the user performed step does not match any production rule or buggy rule. In CBM however, if none of the specified constraints are violated, then the solution is said to be correct. The teaching strategy followed by cognitive tutors is: evaluate every step and provide immediate feedback, if there is a problem [7]. Due to this strategy, every step that the user makes is tracked and evaluated with the correct solution by the cognitive tutors. As CBM itself does not solve the problem, it does not track and evaluate each and every step taken by the user. CBM can evaluate the solution at any time. Partial solutions may be evaluated, provided the system recognizes that the solution is not complete, leaving tests for completeness until the user declares they are done. When the user submits his/her final solution, it is checked for completeness as well as correctness [7]. Dialogue based tutors make the entire teaching-learning experience as a virtual classroom session. Dialogue based model of developing the domain model is used to bridge the gap between human tutors and computers. These tutors require that the user input, i.e., speech must be recognized by the speech recognizer. The syntax and semantics of the speech also need to be checked. The textual matter of the system must be converted into speech for the automated agent. These types of conversions from speech to text and text to speech are not required for Rule based and Constraint based models. Dialogue based tutors could be used for providing guidance to blind individuals. IV. CONCLUSION In this paper, we have reviewed the various approaches used for developing the domain model for an ITS. This paper also analyses the three traditional approaches, which are used for developing the domain model of Intelligent Tutoring Systems. We conclude that each of these approaches have their strengths and weaknesses. Model tracing is an excellent choice for domains where problem solving strategies are well-defined and comprehensive feedback is desirable. On the other hand, CBM offers a alternative when problem solving strategies are not well-defined. The development of dialogue based systems is an attempt to reduce the gap between human and computer tutors. REFERENCES [1]. Nkambou, R., Bourdeau, J., Mizoguchi, R. (Eds) (2010). Advances in Intelligent Tutoring Systems. Springer-Verlag. Berlin Heidelberg.J. Clerk Maxwell, A Treatise on Electricity and Magnetism, 3rd ed., vol. 2. Oxford: Clarendon, 1892, pp.68–73.


2511

[2]. Nwana, H.S. (1990). “Intelligent Tutoring Systems: An Overview”. Artificial Intelligence Review. 4, 251-277. [3]. Intelligent Tutoring Systems: An Overview Gigliola Paviotti, Pier Giuseppe Rossi, Denés Zarka. [4]. Martin, B., “Intelligent Tutoring Systems: The Practical Implementation Of Constraint-Based Modelling” Ph.D thesis. University of Canterbury, 2002. [5]. Anderson, J. R., Lebiere, C.: The Atomic Components of Thought. Mahwah, NJ: Erlbaum (1998). [6]. Ohlsson, S.: Learning from Performance Errors. Psychological Review 103 (1996) 241–262. [7]. Antonija Mitrovic, Kenneth R. Koedinger, and Brent Martin, A Comparative Analysis of Cognitive Tutoring and Constraint-Based Modeling P. Brusilovsky et al. (Eds.): UM 2003, LNAI 2702, pp. 313– 322, 2003. © Springer-Verlag Berlin Heidelberg 2003. [8]. VanLehn, K. et al.: Fading and Deepening: the Next Steps for Andes and other Model-Tracing Tutors. In: Proc. ITS’2000, LNCS Vol. 1839, Springer-Verlag, (2000) 474–483. [9]. Graesser, Arthur C; VanLehn, Kurt; Rose, Carolyn P; Jordan, Pamela W; Harter, Derek.,Intelligent turoring systems with conversational dialogue, AI Magazine 22.4 (Winter 2001): 39. [10].VanLehn, K. et al.: Fading and Deepening: the Next Steps for Andes and other Model-Tracing Tutors. In: Proc. ITS’2000, LNCS Vol. 1839, Springer-Verlag, (2000) 474–483. [11].Antonija Mitrovic , An Intelligent SQL Tutor on the Web, International Journal of Artificial Intelligence in Education 13 (2003) 171Ð195 IOS Press.

2512
