The Design and Implementation of an Active Peer Agent Providing ...

The Design and Implementation of an Active Peer Agent Providing Personalized User Interface Kwangsu Cho1 , Sung-il Kim2, , and Sung-Hyun Yun3 1

3

Learning Research and Development Center, University of Pittsburgh, USA [email protected] 2 Dept. of Education, Korea University, Seoul, Korea [email protected] Div. of Information and Communication Engineering, Cheonan University, Korea [email protected]

Abstract. This study describes the design and implementation of a naive computer peer agent called KORI-2. Unlike prevalent intelligent tutoring systems that implement a machine tutor replacing a human expert, teachable agent systems try to implement a machine peer tutee. Human student tutors are motivated to learn domain knowledge effectively while tutoring a machine tutee. With KORI-2, a human student who plays a tutor role teaches KORI-2 by constructing concept maps. In this tutoring process, KORI-2 actively initiates and guides tutoring interactions by raising (re-)questions or refuting tutor’s explanations. In addition, the KORI-2 system implements interface to contextualize the navigational situations of human tutors.

1

Introduction

Although the importance of motivation to learning in education contexts is evident, most learning systems do not focus on this issue [1]. Rather, they bypass the issue by including external motivators such as rewards and games, assuming that those components will motivate users who would otherwise grow bored. Although learners enjoy games and rewards, they often overlook learning content, which leads to poor learning. Therefore, it seems important to integrate motivation learning activities. To serve the goal, we have developed KORI-2, an intelligent peer tutee agent that is taught by a human student tutor. Previous research supports the potentials of our approach, showing that similarities between peers such as status and knowledge motivate learners to be deeply engaged in peer tutoring activities [2]. It was found that novice students may learn better from peers than from adults or experts [3]. Although both tutor and tutee can benefit from the participation of peer tutoring [4], tutors learn better than tutees [5] by responding or elaborating their knowledge to tutees’ questions [14].

Corresponding Author.

R. Khosla et al. (Eds.): KES 2005, LNAI 3681, pp. 62–68, 2005. c Springer-Verlag Berlin Heidelberg 2005

The Design and Implementation of an Active Peer Agent

63

There have been many attempts to develop tutee agents, peer agents, or teachable agents. One of the well-known systems is Betty [6]. The system is a multi-agent system consisting of a machine tutee agent called and an expert tutor agent as major components. Betty interacts with a human tutor by responding to questions Forbes’ qualitative reasoning algorithm. However, it seems questionable what aspects of Betty define the system a teachable agent. For example, while a student who takes the tutor role tutors Betty by constructing a concept map, Betty is simply assume to learn the content on the map. However, Betty does not have a machine learning algorithm unlike a general expectation. Instead, it simply accepts content from student teaching. More importantly, Betty is a passive learner in that it reacts to questions raised by humans. It does not have actively engaging learning behaviors such as initiation, question, requestion, and rebuttal like human peer tutees. We try to advance the research and development of teachable agents by designing KORI-2 to have more human peer tutee-like behaviors. The goal of this study is to describe the design and development of KORI-2 that interacts with fifth to seventh graders in order to learn geology concepts. As an updated version of KORI [13], KORI-2 is active in that KORI-2 may initiate and shape interactions and also that KORI-2 is personalized according to the level of student tutors. In this paper, we describe a major improvement of the system focusing on the activeness and personalized interface of the system. The other major components can be referred to the previous study [13].

2

The Internal Structure

The brain of KORI-2 consists of inference engines and a knowledge base. The classes of concept maps and inference engines are coded to represent and compute the knowledge of KORI-2. The system also has two independent user interface modules: a knowledge sharing module and a resource module. In the knowledge-sharing module, human tutors teach knowledge to KORI-2 through concept maps with the help of an action-language translator component, which translates between actions on the concept map and structured written language. In the test module, KORI-2’s learning is evaluated with a built-in expert module. The resource module consists of web site links or documents related to the learning content. Figure 1 shows the user interface for human tutors. Human tutors teach KORI-2 which makes its own knowledge and updates it through the feedback mechanism. Concept maps are used to structure and organize knowledge into objects and relations. Concept maps take a form of diagram consisting of nodes and arcs which play a role of external representations. Concept maps have been used in many disciplines such as education, policy studies and organizational studies to visualize knowledge structures and arguments. In educational psychology, Novak developed a system of concept maps [7, 8]. Toumin [9] developed scientific arguments based on concept maps. In AI, Quillian developed concept maps called semantic networks [10]. Along with the development of formal semantics, networks have been widely used for knowledge representations [11].

64

Kwangsu Cho, Sung-il Kim, and Sung-Hyun Yun

Fig. 1. Knowledge Acquisition and Representation Interface between Students and the TA.

In the concept map of KORI-2, a node takes the form of a rectangle which represents an instance or a concept. An arc stands for the causality between two nodes. Labels on arcs provide relational specifications between two nodes. KORI2 represents inheritance between objects as in semantic networks [12]. Domain knowledge is represented in two layers. On a bottom layer, domain ontology is represented in an object-oriented scheme. The top layer is domain task-specific representations using a semantic network. As a form of the semantic network, we use a concept map consisting of nodes and arcs that play a role of external representations. As in Figure 2, for example, the object sedimentary rock has relation to the object deposit. To represent relation between these two objects, an arc is used with the label ’be weathered’. Users can represent their knowledge into concept maps using this method. Individual nodes and arcs have a threshold value, which indicates the importance of the concept or learning difficulty. The threshold value ranges from 0 to 1, and its default value is 1. Concepts deemed important or known to be difficult to understand will be assigned a smaller threshold value to prompt further questioning from the tutee. The value interacts with a tutee activation algorithm, determining whether KORI-2 initiates questions.

3

Active Tutee Behaviors

Since we have developed a computer tutee agent to motivate human tutors to more deeply engage with tutoring, it is important to know what kinds of tutee behaviors prompt or discourage desirable tutor behaviors. To build guidelines to model effective peer tutee behaviors, we observed six pairs of 5th graders interact in peer tutoring situations. Qualitative protocol analyses revealed the following design guidelines of peer tutee behaviors:


65

1) A student tutor engages in tutoring when a student tutee takes a role of a less capable peer relative to the student tutor. 2) A student tutor engages in deeper levels of learning when a student tutee asks questions about the tutor’s explanations. 3) A student tutor appears to lose motivation when a student tutee is unresponsive. 4) A student tutee accepts wrong explanations from a student tutor rather than correct wrong explanations. Following these guidelines, KORI-2 can operate in a naive mode where the agent assumes the role of a less capable peer. The tutee can ask the human tutor questions such as “How sedimentary rock is formed?” or “Are you sure?”. It asks these questions based on both threshold values assigned to the concepts in the domain map and tutee activation values. If the concept the student is currently focused on has a concept threshold value smaller than a tutee activation value, then the system is prompted to question and keeps questions until the activation value is smaller than the threshold value.

Fig. 2. Representation of Objects and their Relations.

As in Figure 2, we use rock cycles as an example of learning contents. The object sedimentary rock has an arc to the object deposit. To represent the relation between these two objects, the arc is labeled with ’be weathered’. Students can represent their knowledge into concept map by using this method. The inference engine uses forward and backward to infer about new relations between objects. Its reasoning result is reflected in the concept map window. After the tutor and tutee finish the concept map, the system activates the inference engine with predetermined rules and tries to find missing or incorrect relations between the sedimentary rock object and the soil object. The dotted arrow with ’be weathered’ label shows this reasoning result.

4

Personalized User Interface

“Personalization is a toolbox of technologies and application features used in the design of an end-user experience. Features classified as ’personalization’ are wideranging, from simple display of the end-user’s name on a web page, to complex

66


catalog navigation and product customization based on deep model of user’s needs and behaviors.” [15]. As an initial stage of implementing personalized user interface, we took the approach of personalizing navigational context [16]. The KORI-2 system is designed to allow human tutors to select their learning options before tutoring KORI-2. Mainly, it would be effective to vary the form of concept maps with which human tutors teach KORI-2: A null map, a partial map and an incorrect map. The partial map shows parts of a full solution. The null map interface presents a blank space without any node or arc. The incorrect map includes partially wrong solutions. Table 1. Measurement Results of Students’ Interest and Understanding to Learn

Interest Interest Interest Interest Understanding Understanding Understanding Understanding

grade competence null map partial map incorrect map 4th low 2.10 3.30 2.85 4th high 3.38 3.53 3.30 5th low 3.11 2.92 2.94 5th high 4.05 3.53 3.81 4th low 2.50 3.23 3.05 4th high 3.69 4.32 3.43 5th low 3.54 3.25 3.29 5th high 4.36 3.46 3.65

In table 1, to support the personalized user interface, we collected empirical data for the level of competence of 4th and 5th grade elementary students. The participants were instructed how to use the concept map. Then, the participants used concept map in forms of a null map, a partial map and an incorrect map. The participants were asked to indicate which type of concept maps they wanted to work on. Results reveal that the participants with high learning motivation and competence tended to prefer the null map or the incorrect map to the partial map. By contrast, the users with low learning motivation and competence prefer the partial map to the other types. A more interesting finding is that users’ interest on the types of concept maps tend to have a high correlation with their understandings, r = .87. Therefore, the results support the personalized user interface approach. The following packages are used to create the interface. KORI-2 is implemented on JAVA platform. For GUI user interface, we use JAVA Swing and Jgraph components. conceptmap - knowledge representation and acquisition using a concept map editor - to manage concept maps in the main window with objects and relations ask answer - to manage Q&A interface using inference modules main - frame structure and main function To simulate the knowledge of KORI-2, the following classes are used. ConceptMap class - manage objects, relations, link paths between objects InferenceEngine class - inference rules, reasoning process, inherit methods from ConceptMap class


67

The reasoning is processed based on concept map knowledge structure. In InferenceEngine class, the following three methods are used to implement reasoning process. get reasoning() - acquisition of new knowledge through reasoning process generate reasoning() - make reasoning result to human tutors’ question ask question() - the agent asks a question to a human tutor in case that there exist several paths to go to the next state during reasoning processes. The ConceptMap class is used to structure and organize the tutee’s knowledge. It gets knowledge from both human tutors and reasoning. During the reasoning process, the following search methods are used to get right results from concept map representations. SearchPath() - the method to access and get knowledge from concept maps buildLinkPath() - set link path between objects addNextPath() - set next object and its relationship

5

Conclusion

This study describes an approach to an intelligent peer agent called KORI-2 that actively interacts with human student tutors and provides the personalized user interface to enhance learner’s motivation and cognition. Since research and development of teachable agents are still in an embryonic stage, we face various technical and educational challenges. However, the system offers various assets that traditional intelligent tutoring systems and teachable agent systems do not provide. In addition, the design and development of intelligent peer tutee systems may inform learning theories as to the necessity of understanding under what conditions tutors learn better and what kinds of tutee behaviors are effective for human tutor learning.

Acknowledgments This research was supported by Brain informatics Research Program sponsored by Korea Ministry of Science and Technology.

References 1. Barth, S., “KM horror stories.,” Knowledge Management, 3, pp.36-40, 2000. 2. Cazden, C. B, “Classroom discourse. In M. C. Wittrock (ed.),” Handbook of reseach on teaching, 3rd Ed. New York: MacMillan, 1986. 3. Cho, K. & Schunn, C. D. (in press), “Scaffolded writing and rewriting in the discipline,” Computers & Education, 2005. 4. Devin-Sheehan, L., Feldman, R., & Allen, V., “Research on children tutoring children: A critical review,” Review of Educational Research, 46 (39), pp.355-385, 1976. 5. Webb, N., “Peer interaction and learning in small groups,” International Journal of Educational Research 13, pp.21-40, 1989.

68


6. Biswas, G., D. Schwartz, K. Leelawong, N. Vye, and TAG-V, “Learning by Teaching: A New Agent Paradigm for Educational Software,” Applied Artificial Intelligence, special issue on Educational Agents, vol. 19, no. 3, March 2005. 7. Novak, J. D, “Concept mapping as a tool for improving science teaching and learning,” In:D. F. Treagust; R, Duit; and B.J. Fraser(Eds), Improving Teachable and Learning in Sceince and Mathematics, London: Tearchers College Press, pp. 32-43, 1996. 8. Novak, J.D., Gowin, D.B., “Learning How To Learn,” New York, Cambridge University Press, 1984. 9. Toulmin, S., “The Uses of Argument,” Cambridge, UK, Cambridge University Press, 1958. 10. Quillian, M.R., “Semantic Memory,” In: Minsky, M., Ed.: Semantic Information Processing, Cambridge, Massachusetts, MIT Press, pp. 216-270, 1968. 11. Gaines, B.R., “An Interactive Visual Language for Term Subsumption Visual Languages,” IJCAI’91: Proceedings of the Twelfth International Joint Conference on Artificial Intelligence, San Mateo, California, Morgan Kaufmann, pp.817-823. 1991. 12. Ames, M., Charles, N., “Snitch: Augmenting Hypertext Documents With A Semantic Net,” 1993. 13. Kim, S., Yun, S.H., Yoon, M., So, Y., Kim, W., Lee, M., Choi, D. & Lee, H., “Design and Implementation of the KORI: Intelligent Teachable Agent and Its Application to Education,” Lecture Notes in Computer Science 3483, Springer-Verlag Berlin Heidelberg, pp.62-71, 2005. 14. Chi, M. T.H., Siler, S. A., Jeong, H., Yamauchi, T., Hausmann, R. G., “Learning from Human Tutoring,” Cognitive Science, 25, pp.471-533, 2001. 15. Kramer, Joseph et al. “A User-Centered Design Approach to Personalization,” CACM, Vol. 43, No. 8, pp.45-48, 2000. 16. Gustavo, R., Daniel, S., Robson, G. “Designing Personalized Web Applications,” ACM, pp.275-284, 2001.