Third International Workshop on Culturally-Aware Tutoring Systems (CATS2010) Workshop Co-Chairs: Emmanuel G. Blanchard McGill University, Canada W. Lewis Johnson Alelo Inc, USA Amy Ogan Carnegie Mellon University, USA Danièle Allard Sherbrooke University, Canada
http://www.iro.umontreal.ca/~blanchae/CATS2010
Preface This year is already the third edition of the workshop on Culturally-Aware Tutoring Systems. The first edition of CATS took place in 2008 in conjunction with ITS2008, which was followed by a second edition, held in conjunction with AIED2009. As in the previous two years, the current workshop aims to continue the investigation of the way culture can be represented within the overarching goal of imparting knowledge via intelligent tutoring systems. It also aims to stimulate discussion on the impact of human cultures on ITS systems and reflect on those emerging technologies that need to be developed to more fully integrate cultural considerations. Finally, it aims at sharing and expanding the knowledge we have about culture, while raising new research questions and opening research opportunities for the ITS community. This year, we have decided to utilize the gathering of intercultural researchers to foster greater discussion and interaction. Hence, we have organized an interactive session around the question “Cultural modelling: Is it intractable?” This session aims at discussing cultural modelling from the point of view of (a) learner understanding, and (b) ITS and pedagogical/conversational agent development. This session was inspired by the observation that the questions currently being faced about cultural modelling are similar to those that John Self addressed about the feasibility of student modelling in general in his seminal 1988 paper, Bypassing the intractable problem of student modelling. Furthermore, similarly to previous editions, CATS2010 also features several traditional paper presentations. Papers concerned with issues of representing culture as well as its impact and influence in the domain of ITS were solicited. A total of five full papers have been selected to be included in these proceedings and to be presented during the workshop. An additional team of researchers is also invited to present their on-going project during the workshop. Our interdisciplinary program committee consists of researchers from ten different countries, each with a unique view on culture. We would like to thank them for their support, along with the ITS2010 organizers for making CATS2010 possible.
June, 2010 Emmanuel G. Blanchard, W. Lewis Johnson, Amy Ogan & Danièle Allard
Program Committee WORKSHOP CO-CHAIRS Emmanuel G. Blanchard, McGill University, Canada W. Lewis Johnson, Alelo, USA Amy Ogan, Carnegie Mellon University, USA Danièle Allard, Dalhousie University, Canada PROGRAM COMMITTEE Jacqueline Bourdeau, TELUQ, Canada Elisabeth Delozanne, Paris VI University, France Benedict Du Boulay, University of Sussex, UK Birgit Endrass, Augsburg University, Germany Paul Fishwick, University of Florida, USA Isabela Gasparini, UFRGS / UDESC, Brazil Monique Grandbastien, University of Nancy, France Jihie Kim, Information Science Institute/USC, USA H. Chad Lane, Institute for Creative Technology/USC, USA Christopher Miller, SIFT, USA Riichiro Mizoguchi, Osaka University, Japan Yukiko Nakano, Seikei University, Japan Elaine Raybourn, Sandia National Laboratories / University of New Mexico, USA Matthias Rehm, Aalborg University, Denmark Katharina Reinecke, University of Zurich, Switzerland Isabelle Savard, Laval University, Canada Julie Sykes, University of New Mexico, USA Heike Winschiers-Theophilus, Polytechnic of Namibia, Namibia Robert Wray, SOAR Technology, USA Shumin Wu, IBM Silicon Valley Lab, USA
Table of Contents
Full Papers An adaptive e-learning environment based on user’s context Isabela Gasparini, Amel Bouzeghoub, José Palazzo M. de Oliviera, & Marcelo S. Pimenta
1
Using immersive simulation to develop intercultural competence W. Lewis Johnson, & Alicia Sagae
13
Creating virtual synthetic cultures for intercultural training Samuel Mascarenhas, & Ana Paiva
25
CAMPERE: Cultural adaptation methodology for pedagogical resources in e-learning Frank Mpondo Eboa, François Courtemanche, & Esma Aïmeur
37
Interactive PhrasebookTM – Conveying culture through etiquette Peggy Wu, & Christopher Miller
47
Additional Presentation (not included in these proceedings) Possibilities of describing culturally-aware characteristics in learning objects metadata Júlia Marques Carvalho da Silva, Tiago Thompsen Primo, Rosa Maria Vicari
Full Papers
An adaptive e-learning environment based on user´s context 3
Isabela Gasparini1,2,3, Amel Bouzeghoub , José Palazzo M. de Oliveira1, Marcelo S. Pimenta1 1
2
Instituto de Informática, Universidade Federal do Rio Grande do Sul, Porto Alegre, Brazil Departamento de Ciência da Computação, Universidade do Estado de Santa Catarina, Brazil 3 TELECOM & Management SudParis, France {igasparini, palazzo, mpimenta}@inf.ufrgs.br;
[email protected]
Abstract. Since e-learning systems are used by a wide variety of students with different characteristics, being adapted to user’s model and profile is an essential feature. Although there are several approaches for adaptive e-learning environments, they focus mainly on technological and/or networking aspects without taking into account other contextual aspects, such as cultural and pedagogical context. This paper presents an improvement of an adaptive elearning environment called AdaptWeb®, based on a rich context model as an extension to the traditional student modeling. Keywords: Adaptation, context-aware, learning environments, cultural-aware.
1 Introduction E-learning environments (ELE) have gained wider acceptance in the last ten years and are currently being used by teachers in courses with a wide variety of students that have different skills, background, preferences, and learning styles. Many authors now agree that one of the most desired characteristics of these environments to cope such diversity is that of being adaptive and personalized [1]. Personalization is the process of adapting a computer application to the needs of specific user (in e-learning domain – a student or teacher), and it takes advantage of knowledge acquired and registered about them [1], [2]. In fact, the use of personalization techniques aims at improving the ELE usability since a personalized system automatically customizes the user interface, content, access rights, navigation, etc, considering the user profile and each user may think that the system was designed specifically for him/her. In contrast to customization - which is a user-initiated and user-driven process, personalization is system-initiated and system-driven, thus requiring the system to monitor the user’s profile changes in order to adapt the system dynamically. The most common contents of existing student models for e-learning are: students’ interests; knowledge, background and skills; experiences; goals; behavior; interaction preferences; individual traits and learning styles. However, ELEs may be dynamically
1
adjusted not only according to the student’s model but also depending on a richer notion of context [3]. In our work, context is considered as having cultural and technological perspectives, in addition to previous individual and pedagogical contents related to student modeling. We organize this contextual information in a multidimensional space where each dimension is represented by a specific ontology, which may be handled separately or jointly by the system. A personalized ELE based on context provides the learner with exactly the material he needs, and appropriate to his knowledge level and that makes sense in a special learning situation, which together with external factor is known as a scenario in our work. Thus, for each scenario, an ELE is dynamically adjusted depending on the context information available. In this paper, we present an architecture that increments even more the actual systems' personalization capabilities making use of ontologies to model the student’s context in different scenarios. Our goal is to improve user’s interaction with the system by a more dynamic and adapted feedback to user’s needs in a particular time. Thus, we believe that the learning process can be improved as well. This paper is structured as follows. Section 2 presents the fundamentals involved in this paper. Section 3 explains some related works. Section 4 shows AdaptWeb system. Section 5 discusses about Ontology-based context modeling. Section 6 presents our Architecture for context-aware learning. Section 7 shows some examples of use and section 8 some conclusions.
2 Fundamentals The research and practice in Computer Science and Education have evolved with the introduction of the Internet and Web-based courses. However, some educational applications are usually developed without taking into account the dynamic capabilities and personalization that the Web environment can provide, such as analyze the bandwidth and provide videos or learning objects according to it at a given time, discover user´s culture background and dimensions to present interface and content adapt to him, and so on. This lack raises serious usability issues [4] such as (i) guidance problems, i.e., pages always presenting the same fixed content, and (ii) navigation problems, which are the consequence of a totally open linking schema and without focusing on a more adequate for each student. Basically, adaptation approaches automatically adapt system elements considering the user profile (i.e., the student model in the ELE domain). Student modeling is nowadays essential as a non-intrusive way of gaining information about students' characteristics, and the most common contents of student models for e-learning include student’s interests; knowledge; background and skills; experiences; goals; behavior; individual traits and learning styles, all of them having a well-known and straightforward description. In ELE, we have to consider at least three points-of-view, related to (1) learner, (2) learning activities, and (3) environment (technical and physical commuting, device, location, time) [15]. About the learner, the inclusion of multicultural aspects into ELE is vital. Cultural aspects are preferences and ways of behavior determined by the
2
person’s culture. Indeed, the cultural aspects are just the features that distinguish between the preferences of students from different regions [22]. Nevertheless, taking in account culture for personalizing ELE is a new approach. As well as other software applications, ELE are usually restricted to one personalization strategy per country. However, a predefined localized personalization cannot be assigned to all people of a nation, as some might have many cultural influences and are, therefore, culturally ambiguous [8]. This means that the ELE must be personalized in relation to a particular set of student’s cultural properties. Thus, modeling cultural profiles can be a way to improve cultural awareness in global knowledge sharing and learning processes. Cultural characteristics can be described on different levels, such as national and regional, organizational, professional, and individual characteristics.
3 Related Work Context adaptation is not a new challenge in context-aware environments or in pervasive learning environments. Recent works aim to provide the capacity for identifying the right contents, right services in the right place at the right time and in the right form based on the current situation of the learner. There is an interesting theory of learning for a mobile society [9] but our work is closely related to others like [10], [11], [12], [13], [14] and [15]. The interesting propositions of GlobalEdu [10], [12] in terms of architecture, for instance, have distributed and central alternatives with different models (learner, context and environment). Specifically about cultural aspects, Blanchard and Mizoguchi [27] described preliminary work on an upper ontology of culture. By working at the meta-level of culture, their work aims at identifying major constituents to be considered when dealing with any kind of cultural issue without having to endorse a particular culture’s representational framework. They propose using this approach to deal with many CATS (Culturally-Aware Tutoring Systems) related issues by providing objective formalism for cultural representation. Savard, Bourdeau and Paquette [28] investigate the application of cognitive informatics in the domains of education and culture. They focus particularly on cultural diversity in computer-assisted distance learning environments. Chandramouli et al. [26] introduce the notion of the CAE-L Ontology for modeling stereotype cultural artifacts in adaptive education. They use a Cultural Artifacts in Education (CAE) questionnaire to gather the information required to determine if there is a significant cultural bias within online education, specifically Adaptive Educational Hypermedia. Motz et al. [29] present an architecture used in the ongoing e-learning EduCa Project, based in a strong use of ontologies for the retrieval, management and clustered of electronic educational resources according to user’s cultural aspects, like degree of impatience, attitude, treatment, language, learning styles and activities. These cultural aspects are specified in a MultiCultural Aspects Ontology, which follows the standard Learning Object Metadata (LOM) and uses OWL (Web Ontology Language). Sieg, Mobasher and Burke [30] proposed a framework that integrates critical elements that make up the user context, namely the user’s short-term behavior, semantic knowledge from ontologies that provide explicit
3
representations of the domain of interest, and long-term user profiles revealing interests and trends. They present a novel approach for building ontological user profiles by assigning interest scores to existing concepts in a domain ontology. Our research has a different point of view of these related works because we are looking for integrating cultural, technological and pedagogical aspects as part of a context user model that makes sense in a special situation, in a given time. Despite the considerable scientific production on customization, personalization and adaptation in e-learning, surprisingly there is very little research work devoted to how to incorporate context-aware and culture-oriented concerns to such systems.
4 AdaptWeb® Environment The adaptive learning system AdaptWeb® aims to adapt content, presentation, and navigation in an educational web course. It is an open source environment fully operational, available at SourceForge, and actually being used in different universities. The educational content of AdaptWeb® is defined during the authorship’s phase and then stored in XML. The XML documents go then through a filtering process, which happens dynamically as the student interacts with the environment. So, the instructional content provided by the AdaptWeb system is dynamically adjusted according to the student’s model. This model combines the student information with a structure of concepts that result from the authorship process. The student’s model describes users in terms of theirs characteristics as knowledge, interactions preferences, background, technological resource, navigational history and cognitive learning styles. A Computer Science student has distinct background from a Mathematic student and they need distinct contents of the logic discipline. The teacher, through his experience, determines which content will be provided for each group of students (knowledge depth), thus enabling the content to be adapted for each student. User’s interaction preference set one of two navigation modes for a discipline: guided tour or free mode. Guided tour takes into account the dependencies among contents set by the author at the stage of authorship, and the student can only access a concept if its prerequisites are already known. In the free mode, the student can navigate freely. The student’s knowledge is monitored and the student’s progress is constantly updated in the user profile.
5 Ontology-based Context Modeling To be effective, learning process must be adapted not only to the student’s profile but by the learner’s context as well, creating some kind of matching between context and profile to provide for example the appropriate content, navigation, and recommendations. Learning processes have to provide extremely contextualized content that is highly coupled with context information, limiting their reuse in some other context. If the context information is represented independently from content information, the possibilities for reuse increase.
4
In a broader sense, context describes the circumstances under which something occurs as well as the interrelationships of those circumstances. Such interrelationships provide a semantic perspective that restricts and narrows the meaning of “something” [17]. A context-aware ELE is an application that adapts its behavior according to its students’ context. Context-aware applications not only use context information to react to a user’s request, but also take the initiative as a result of context reasoning activities [18]. Thus, an improvement in the user’s contextual information leads to a better understanding of users’ behavior in order to adapt (i) the content, (ii) the interface, and (iii) the assistance offered to users. We have developed a model based on upper-level ontology. In this model, a user might be involved in several overlapping contexts (such as technological and personal contexts), and consequently, his/her educational activity might be influenced by the interactions between these contexts. Overlapping contexts contribute to and influence the interactions and experiences that people have when performing certain activities [15, 19, 20]. As deeply described in [16], our model has three levels: meta-model, model (ontologies), and object (Figure 1). The meta-model level is represented by an upper ontology, the model level with several ontologies to describe the elements that populate the context and, in the lower level, we find the instantiations of the context ontologies. In other words, the ontology concepts of one level are the instantiations of its immediate superior level. Thus, the concepts of the object level are instances of the model level which is further formed by instances of the meta-model level.
Fig. 1. Three-level context model We personalize an ELE for each user based on the information stored in a student model. In our work, the typical characteristics of students are extended to include the context dimensions having personal, cultural, technological and pedagogical aspects. Personal context is widely considered in e-learning, e.g. EPIAIM, KN-AHS, ITEM/PG, ISIS-Tutor, HyperTutor, ELM-ART, [2], AHA! [21], eTeacher [23] and AdaptWeb®. This type of context is usually gathered in the student model and contains the most important or interesting facts about the user, including interests; knowledge, background and skills; experiences; goals; behavior; interaction preferences; and last but not least individual traits and learning styles. Cultural context includes cultural background of a student and may have a great impact on their ability and efficiency to learn a given set of content [26]. Cultural context is referred to different languages, values, norms, gender, social or ethnic aspects. Thus, modeling of culture profiles can be a means to improve cultural awareness in global knowledge sharing and learning processes. Cultural characteristics can be described in different levels: indeed, there are cultural characteristics related to a nation, to an organization or even to a small group. There
5
are different cultural dimensions proposed in the literature, but the most accepted were the five dimensions proposed by Hofstede [31], which are based on value orientations and shared across cultures. According to Bossard [32] there are two categories of topics that are affected in human computer interaction localization, (a) presentation of information (e.g. time, date and color format) and language (e.g. font, writing direction, etc.); and (b) dialog design (e.g. menu structure and complexity, layout, positions) and interaction design (e.g. navigation concept, interaction path, interaction speed, system structure, etc.). Despite some HCI (Human-computer interaction) works now focusing on cross-cultural aspects in HCI, the research of cultural-dependent aspects of HCI, is still embrionary [33]. Technological context is related to many different technological constraints (e.g., device processing power, display ability, network bandwidth, connectivity options, location and time). Pedagogical context consists of multifaceted knowledge due to many distinct viewpoints of pedagogical information needed to personalize e-learning. In practice, many adaptive systems take advantage of users’ knowledge of the subject being taught or the domain represented in the hyperspace, and the knowledge is frequently the only user feature being modeled [2]. Recently, various researches started using other characteristics, such as personality model OCEAN [5], cognitive [6] and learning styles (for example, from Felder's model [7], [23]. Among all the possible information gathered in the student model, we are specially interested in modeling scenarios because they change according to context. Scenarios may depend on the situation the student is now in and on external factors. Therefore, it is important to model in which context the student prefers something. The concreteness of scenarios helps students and teachers to develop a shared understanding of the proposed contextual information, and allows assimilating and representing complex idiosyncrasies of that they would otherwise misunderstand. Scenarios are extremely related to the contextual background information and describe a simple perspective for execution of a situated action [24]. The term situated action emphasizes the interrelationship between an action and its context of execution and the notion that people's behavior is contextualized, i.e., the situation is a very important factor in determining what people will do. We define a scenario as a tuple containing an entity that the student prefers in a given situation, a relevance denoting the student’s preference for that entity, a certainty representing how sure we are about the student having that preference and a date to indicate when that preference is stored: Scenario = {entity, situation, relevance, certainty, date}
Situations are the key to include temporal aspects of context in a comprehensive ontology for context modeling, since they can be related to suitable notions of time [25]. As context varies during certain time intervals, it is vital to consider it within the concept of Situation. Examples of situations could be “John was at home using his notebook to read lesson number 3 of the Human Computer Interaction course” or “A Japanese Professor, who speaks English, is adding new exercises to the course Introduction to Java using a high speed connection while she travels by train”. Therefore, we define situation as a set of contextual information in a particular period of time: Situation = {Context, initial time, final time}
6
An example of contextual information would be: “The student named John is reading lesson number 7”. This is a description relating an entity (the student John) to another entity (the lesson number 7) via a property (is reading). We represent this contextual information as (Student.john, isReading, Lesson.lesson#7). We define the context as a set of triples composed by concepts, instances and relations between them. It is important to emphasize that the concepts and instances might belong to the same ontology or different context ontologies: Context = {(Ca1.Ia1, R1, Cb1.Ib1), ..., (CaN.IaN, RN, CbN.IbN)}; (C: concept, I: instance and R: relation).
To clarify these ideas, let us consider again John example. John prefers reading visual learning material in a situation when he is at home using his notebook to read lesson number 3 of the Human Computer Interaction course. Hence, the corresponding context1 will be: Context1={ (Person.John, locatedIn, Location.home), (Person.John, uses, Device.notebook), (Person.john, reads, Lesson.lesson#3), (Lesson.lesson#3, belongsTo, Course.HCI)} Situation1={ Context1, 4:00PM, 7:00PM} Scenario1={User, Situation1, relevance.high, certainty.95%, date.05-02-2010}
Figure 2 presents the proposed model. The meta-model is an upper-level ontology describing abstract concepts like user, application, user profile, situation or date. The model depicts the different contextual dimensions. Each contextual dimension is represented by a different ontology such as a cultural ontology (with concepts like culture, social norm or language), education ontology (course, learning style, discipline, etc.), personal ontology (name, gender, birthday, etc.) or technological ontology (operating system, browser, network bandwidth, etc.). Finally, the object model will comprise instances describing the context of a particular user like a concrete name (John Smith), a course (Human Computer Interaction) or a particular language (English).
Fig. 2. Example of a scenario-oriented situation
7
6 Architecture for Context-Aware Learning Environment The AdaptWeb® architecture has been extended in order to support context awareness, being adapted to specific scenarios, as mobility, social interaction, cultural-aware and device independence. First, we have restructured the existing AdaptWeb® architecture to be Service-Oriented (SOA), having different data and context modules that are individually adapted and orchestrated by a set of high level services, allowing our environment to communicate with other ones in a flexible manner. The new architecture is based on three servers that operate together to provide and manage contextualized data according to the student's scenarios. Each server manages specific data related to the user context, being respectively responsible for the storage and adaptation of (i) information about students (personal data, preferences, objectives, knowledge background, behavior, learning styles, cultural context, etc.), (ii) environmental context (information related to the user environment, tasks, activities, time interval, devices, location), and (iii) learning object’s information (documents provided by the educational environment to its users for their learning). However, since this information is managed and stored separately and the context needed to the adaptation environment is diverse, it is orchestrated by an internal component called Context Management Service (Figure 3). This service is responsible for analyzing the context managed by the servers, generating different scenarios that can be experienced by the students in a specific period. These scenarios are used to guide the adaptation (in the Adaptation Engine), and materialized in the interface rendered to the user. The main goals of the architecture are: (i) easily reuse of educational resources, since they will be adapted to the user scenario while the stored content remains the same, (ii) integration into the existing architecture, since the new architecture is supposed to take advantage of the existing functionalities and (iii) extensibility to other educational systems, using standard technologies. The personalization is possible with the combination of contextual data related to whom and where the user is, what he/she is doing and what does he/she needs to achieve his/her educational targets.
Fig. 3. Context-Aware Learning environment architecture
8
7 Examples of context adaptation and recommendation In this section, we describe some improvements of the personalization’s capabilities of AdaptWeb® in order to provide support to this contextual modeling purpose. In order to explain the contextual adaptations developed in AdaptWeb®, we start by describing different learning situations, and then we detail how those situations trigger the corresponding contextual adaptation in AdaptWeb. We show some examples of possible contexts in an Artificial Intelligence course. For a simplification purpose, we have a few variables: student's knowledge, subject, network connection, learning style, Language, LanguageLevel and Country. In Context1, João is a student who lives in Brazil, his mother tongue is Portuguese, and he has a low level knowledge in English. He is trying to learn about the subject “Bayesian networks”, which is explained in English. He is doing exercises about that subject, but unfortunately he is not obtaining satisfactory results. In addition, he has a high network connection and according to Felder’s model [7] he is active. The user model checks his number of mistakes and identifies he needs help resolving the exercises. At the same time, the situation model detects via teacher’s agenda that a chat with the students was previously scheduled by the teacher to happen in 10 minutes. These events will start a service of notification in the Context Management Service, informing that a change of the current scenarios related with these events may change. After a new orchestration by the Context Management Service, the User Interface sends a message to the student, notifying him of this possibility to solve his doubts and shows the “chat” link in a different and highlighted color. In another scenario, Context2, Maria, a Spanish-speaking PhD student of Engineering from Argentinean, has very good skills in three different foreign languages (English, Portuguese and French). She is also learning the subject “Bayesian networks” and not having good results. She has a low network connection and her Felder's learning style is reflective. In consequence, AdaptWeb sends a message by email to her teacher advising to contact the student and changes the order of the links, putting links related to video material with low quality resolution in the end and disabling links related to video material with high quality resolution (those who are heavy and difficult to see). Furthermore, AdaptWeb detects some important links for learning material written in English and shows this in the top of the list. These contexts are formalized as following: Context1 = {(Student.João, isLearning, Subject.bayesianNetworks), (Subject.bayesianNetworks, isExplainedIn, Language.english), (Student.João, hasUserKnowledge, UserKnowledge.bad), (Student.João, hasConnection, NetworkConnection.high), (Student.João, hasStyle, LearningStyle.active), (Student.João, hasMotherTongue, Language.portuguese), (Student.João, hasLanguageSkill, Language.english), (Student.João, hasEnglishLanguageLevel, LanguageLevel.low), (Student.João, isCitizenOf, Country.Brazil)} Context2 = {(Student.María, isLearning, Subject.bayesianNetworks), (Student.María, hasUserKnowledge, UserKnowledge.bad), (Student.María, hasConnection, NetworkConnection.low), (Student.María, hasStyle, LearningStyle.reflective), (Student.María, hasMotherTongue, Language.spanish), (Student.María, hasLanguageSkill, Language.english), (Student.María, hasEnglishLanguageLevel, LanguageLevel.high), (Student.María, isCitizenOf, Country.Argentina)}
In summary, the adaptation mechanisms in AdaptWeb can be for example the following actions/recommendations:
9
Context1 “send notification to student only in Portuguese” + “show highlighted links”+ recommend learning objects and content about the same subject (same concept in the domain ontology) in Portuguese with a low level of difficulties; Context2 “order links” + “hide or disable links” + “show highlighted links” + “recommend learning objects and content about the same subject written in Spanish, English or French”.
These contexts are used by the context management service within logical rules in order to predict future recommendations. This is an example where a prototypical context (Contexti) is used to generate new recommendation: Contexti = {(Student.S, isLearning, Subject.s), (Student.S, isUsing, LearningObject.lo), (Subject.s, isExplainedIn, Language.l), (Student.S, hasUserKnowledge, UserKnowledge.bad), (Student.S, hasConnection, NetworkConnection.high), (Student.S, hasStyle, LearningStyle.active), (Student.S, hasMotherTongue, Language.l’), (Student.S, hasLanguageSkill, Language.l), (Student.S, hasEnglishLanguageLevel, LanguageLevel.low), (Student.S, isCitizenOf, Country.c)} If S is in Contexti then recommend new LO such as LO.Subject= lo.subject and LO.language= S.motherTongue and LO.level is less than lo.level.
A learning object (LO) is defined as any entity, digital or non-digital, that may be used for learning, education or training [34]. Currently, this model is being under evaluation with real students and actual courses in AdaptWeb environment.
8 Conclusions Naturally, as e-learning systems become more sophisticated, new opportunities and new challenges are emerging. One meaningful example is the need to deal with context modeling and its relation with user modeling. Context modeling extends traditional user modeling techniques, by explicitly dealing with aspects we suppose to have a significant influence on the learning process assisted by an e-learning environment. We propose the use of ontologies to model this contextual information, as a three level model to capture different levels of detail. This paper described the combination in the same approach of ontology-based and web-services-oriented aspects, to represent explicitly the rich context as an extension of traditional user-student modeling. Our ultimate aim is to increment even more the actual systems personalization capabilities making use of ontologies to model the student’s context in different scenarios, adapting the systems content, navigation and presentation. Our future work includes increasing cultural aspects as different values, norms, social or ethnic aspects in the context model, to have a more understanding of users and a better communication with learning objects provide by teacher, meaning effective learning for students in AdaptWeb environment. We plan to carry out more experimental tests, with our partners in different countries, using the same content, in various contexts and a wide variety of students. Acknowledgments. This work has been partially supported by Conselho Nacional de Desenvolvimento Científico e Tecnológico - CNPq, Brazil, and by the projects AdaptSUR 022/07 (CAPES, Brazil) - 042/07 (Secyt, Argentina), and AdContext 54707 (CAPES-COFECUB).
10
References 1. Brusilovsky, P., Peylo, C. Adaptive and intelligent Web-based educational systems. In: International Journal of Artificial Intelligence in Education 13 (2-4), Special Issue on Adaptive and Intelligent Web-based Educational Systems, 159-172 (2003). 2. Brusilovsky, P., Millan, E. User Models for Adaptive Hypermedia and Adaptive Educational Systems. The Adaptive Web, 4321, 3-53 (2007). 3. Eyharabide, V., Gasparini, I., Schiaffino, S., Pimenta, M. and Amandi, A. Personalized elearning environments: considering students’ contexts. IFIP World Conference on Computers in Education, v. 302, pp 48-57, Springer (2009). 4. Kalbach, J. Designing Web Navigation: Optimizing the User Experience, O´Reilly (2007). 5. Goldberg, L. R. The structure of phenotypic personality traits. American Psychologist, 48, 26-34 (1993). 6. Ford, N., Chen, S. Y. Individual Differences, Hypermedia Navigation, and Learning: An Empirical Study. Journal of Educational Multimedia and Hypermedia, 9(4), 281-311 (2000). 7. Felder, R. M., Brent, R. Understanding Student Differences. J. Engr. Education, 94 (1), 5772 (2005). 8. Reinecke, K., & Bernstein, A. (2008). Predicting user interface preferences of culturally ambiguous users. CHI '08: CHI '08 extended abstracts on Human factors in computing systems (pp. 3261--3266). New York, NY, USA: ACM. 9. Sharples, M., Taylor, J., Vavoula, G. A Theory of Learning for the Mobile Age, In R. Andrews and C. Haythornthwaite (eds.) The Sage Handbook of Elearning Research. London: Sage, pp. 221-47 (2007) 10. Barbosa, D. N. F., Augustin, I., Barbosa, J. L. V. Proceedings of the Fourth Annual IEEE International Conference on Pervasive Computing and Communications Workshops PERCOMW’06 (2006). 11. Lemlouma, T., Layaïda, N. Context-Aware Adaptation for Mobile Devices, IEEE International Conference on Mobile Data Management, Berkeley, CA, USA, pp. 106-111 (2004). 12. Rosa, G.P.J., Ogata, H. Yano, Y. A multi-Model Approach for Supporting the Personalization of Ubiquitous Learning Applications, IEEE International Workshop on Wireless and Mobile Technologies in Education, pp. 40-44 (2005). 13. Yang, S. J. H., et al. Context Model and Context Acquisition for Ubiquitous Content Access in ULearning Environments”, Proc. of the IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing, vol. 2, pp. 78-83 (2006). 14.MOBIlearn, 2003, MOBIlearn final report, viewed January 2007, 15. Bouzeghoub, A., Do Ngoc, K. A situation based metadata for describing pervasive learning objects. mLearn 2008: 1st International Conference on Mobile Learning, October 8-10, University of Wolverhampton, Ironbridge, UK (2008) 16. Eyharabide, V., Amandi, A. An Ontology-Driven Conceptual Model of User Profiles. In Proceedings of ASAI07, Simposio Argentino de Inteligencia Artificial, Argentina (2007). 17. Abarca, M., Alarcon, R., Barria, R., Fuller, D. Context-Based e-Learning Composition and Adaptation. OTM Workshops (2), On the Move to Meaningful Internet Systems 2006, pp. 1976-1985 (2006). 18. Dockhorn Costa, P., Almeida, J., Pires, L., van Sinderen, M. Situation Specification and Realization in Rule-Based Context-Aware Applications. Distributed Applications and Interoperable Systems, 7th IFIP WG 6.1, DAIS 2007, Paphos, Cyprus, 2007, Proc. pp. 3247 (2007). 19. Shen Yang, S., Huang, A., Chen, R., Tseng, S.-S., & Yen-Shih. Context Model and Context Acquisition for Ubiquitous Content Access in ULearning Environments. IEEE International Conference on Sensor Networks, Ubiquitous, and Trustworthy Computing, 2, pp. 78-83, (2006).
11
20. Eyharabide, V., Amandi, A. Semantic spam filtering from personalized ontologies. JWE Journal of Web Engineering, Rinton Press, 7 (2), 158-176 (2008). 21. De Bra, P., Stash, N., AHA! Adaptive Hypermedia for All. Proceedings of the SANE 2002 Conference, Maastricht, pp. 411-412 (2002). 22. Guzman, J., & Motz, R. Towards an Adaptive Cultural E-Learning System. LA-WEB '05: Proceedings of the Third Latin American Web Congress (p. 183). Washington, DC, USA: IEEE Computer Society (2005). 23. Schiaffino, S., Garcia, P., Amandi, A. eTeacher: Providing personalized assistance to elearning students. Computers and Education, 51 (4), 1744 – 1754 (2008). 24.Suchman, L.A. Plans and Situated Actions: The Problem of Human-Machine Communication. Cambridge: Cambridge Press (1987). 25. Dockhorn Costa, P., Guizzardi, G., Almeida, J., Pires, L., & van Sinderen, M. Situations in Conceptual Modeling of Context. EDOCW '06: Proc. 10th IEEE on International Enterprise Distributed Object Computing Conference Workshops (p. 6). IEEE Computer Society (2006). 26.Chandramouli, K. and Stewart, C. and Brailsford, T. and Izquierdo, E. CAE-L: An Ontology Modelling Cultural Behaviour in Adaptive Education. 2008 Third International Workshop on Semantic Media Adaptation and Personalization, IEEE Computer Society, 183- 188. 27. Blanchard, E.G., & Mizoguchi, R. (2008). Designing Culturally-Aware Tutoring Systems: Towards an Upper Ontology of Culture. Workshop on Culturally-Aware Tutoring Systems (CATS 2008), ITS 2008: Intelligent Tutoring Systems: past and future, pp 23-34. 28. Savard, I., Bourdeau, J., & Paquette, G. (2008).Cultural Variables in the Building of Pedagogical Scenarios: the Need for Tools to Help Instructional Designers. In: Workshop on Culturally-Aware Tutoring Systems (CATS 2008),ITS 2008: Intelligent Tutoring Systems: past and future, pp 83-92. 29. Motz,R., Guzmán, J., Deco, C., & Bender, C. (2005). Applying Ontologies to Educational Resources Retrieval driven by Cultural Aspects, Journal of Computer Science & Technology (JCS&T), vol. 5, no. 4, pp 279 - 284. 30. Sieg, A; Mobasher, B; Burke, R. Representing Context in Web Search with Ontological User Profiles. (2007). LNCS, Modeling and Using Context, v. 4635, pp 439-452, Springer. 31. Hofstede, G. Cultures and Organizations: Software of the Mind. New York: McGraw-Gill, 1991. 32. Bossard, A. Ontology-Based Cultural Personalization in Mobile Applications. Thesis (Master Degree), Department of Informatics, University of Zurich, 2008. 33. Zaharias, P. Cross-Cultural Differences in Perceptions of –leaning Usability: An Empirical Investigation. International Journal of Technology and Human Interaction, v. 4, issue 3, ed. Bernd Carsten Stahl, IGI Global, 2008. 34. IEEE. Draft Standard for Learning Object Metadata. (2002) Available: http://ltsc.ieee.org/wg12/files/LOM_1484_12_1_v1_Final_Draft.pdf
12
Using Immersive Simulations to Develop Intercultural Competence W. Lewis Johnson and Alicia Sagae Alelo Inc., 19210 Culver Bl., Suite J, Los Angeles, CA 90066 USA
[email protected],
[email protected]
Abstract. The Alelo family of learning products is intended to help people to quickly develop the cultural skills they need to be effective in intercultural settings. This includes the knowledge and skills necessary to handle common situations involving intercultural interactions, as well as the adaptability needed to cope with unexpected intercultural interactions. The approach utilizes simulations of intercultural encounters that learners are likely to experience in the course of carrying out their jobs or missions, which help learners develop the skills and confidence necessary to be effective in those situations. The approach also supports the assessment of cultural competence by testing trainees in simulated intercultural encounters. The courses employ a number of innovative technologies, including immersive game and artificial intelligence technologies. Keywords: Computer-assisted learning of culture, computational modeling of cultural contexts and dynamics, cultural adaptation methodologies
1 Introduction Intercultural skills are increasingly recognized as important for many jobs and everyday situations. The workforces of multinational corporations are frequently organized into multinational teams in which people from various cultural backgrounds need to work together effectively. Companies that deliver products internationally need to understand the cultural backgrounds and perspectives of their potential customers in each target market. Professionals engaged in health care, law enforcement, education, humanitarian relief, and peacekeeping operations all require skills in interacting with people with other cultural backgrounds [2], [5], [8], [9], [11]. Unfortunately, relatively few specialists in other professions get the opportunity to devote that much time to cultural training. Thus a major challenge for cultural skills training is finding ways to help people develop the intercultural skills they need, in the time they have to learn. This article describes a simulation-based approach to cultural competency training realized in the Alelo family of training products. It is intended to help learners who may not be cultural specialists to quickly develop cultural skills they need to be effective in intercultural settings. This includes the knowledge and skills necessary to handle common situations involving intercultural interactions, as well as the
13
adaptability needed to cope with unexpected intercultural interactions. Crucially, it utilizes simulations of intercultural situations that trainees are likely to encounter in the course of carrying out their jobs or missions, which helps trainees develop the skills and confidence necessary to apply intercultural skills in those situations. The approach also supports the assessment of intercultural competence by testing trainees in simulated intercultural encounters. The approach makes use of a number of innovative technologies, but most importantly immersive game and artificial intelligence technologies. Immersive game technologies are used to simulate the intercultural encounters that serve as the context for skills training. Artificial intelligence technology is employed to create interactive computer characters that can engage in dialog with the learners, in a culturally appropriate way.
2 Example Courses
Fig. 1. Operational Dari language and culture training system.
Many users of Alelo courses are members of military services around the world, who require training in cross-cultural communication skills so that they can perform missions overseas, such as reconstruction and humanitarian assistance, and work together with coalition military partners. Fig. 1 shows a scene from one such course, the US version of the Operational Dari language and culture trainer. Operational Dari includes interactive lessons that teach task-relevant communication skills and interactive scenarios in which learners can practice those skills. In the example in the
14
figure, the learner, playing the character on the left, is engaged in a meeting with the elders in an Afghan village. He needs to develop rapport with the elders in order to gain their trust and cooperation. To play the scenario, the learner must speak into the microphone in Dari, speaking phrases appropriate for his or her character in the scenario, and the non-player characters respond accordingly. The dialog history is shown in the top center of the figure.
Fig. 2. Exercise from the goEnglish Web site.
This scenario-based approach is being applied more broadly to teach language and culture. For example, Fig. 2 shows a dialog exercise from goEnglish, a Web site developed for Voice of America to teach American culture and language worldwide. The site is designed for learners who have acquired a basic reading knowledge of English in school, but have a limited understanding of American culture, and have had a limited opportunity to develop and practice their conversational skills. It contains a set of lessons that focus on the language and culture of situations that arise in everyday life in the United States, as well as cultural topics of particular interest. To develop the set of topics, we interviewed recent immigrants to the United States to find out what situations they found particularly surprising or challenging. The site contains a set of lessons where they learn conversational language and cultural points relevant to these situations and topics, and exercises in which they practice in simulated conversational situations. In the situation shown in Fig. 2, the learner is getting to know an acquaintance named Malika and is talking about their relationships. The learner has found out that Malika has a boyfriend (see the dialog history in the top center of the figure) and can now ask further questions, such as when and where they met. If the learner is uncertain as to what to say, he or she can consult a menu of hints about conversational topics (top left, in Chinese) and possible ways of expressing each particular topic (center left). The learner’s score for the
15
exercise depends upon the total number of dialog exchanges, and the extent to which the learner relied on hints to get through the scenarios.
3 Course Structure and Development Methodology Developing Alelo courses requires knowledge of the culture being taught. This is used both to develop learning materials that teach about the culture, and to develop animated agents that behave in accordance with the norms of the culture, as in Figs. 1 and 2. Depending upon the target audience and the learning objectives of the course, some courses provide extensive instruction in the local language, and include characters that understand and speak only the local language, while others include a more limited amount of language instruction and engage in conversation in the learner’s native language. It is also important to have a good understanding of the learner’s culture to help ensure that the learning materials and objectives are presented in a style, and with language and phraseology, that match the learner’s expectations. As noted above, goEnglish is available in multiple languages. Likewise, Operational Dari is available in versions intended for use by the military services of the United States, Australia, the United Kingdom, Germany, and other NATO countries. The instructional language of the one for Germany is in German; all the others are in English. Course details vary slightly in each case to reflect the learner’s culture. Differences include uniforms and physical appearance of game characters, artwork, as well as terminology, spelling, and phraseology. Voiceovers and narrations may be recorded using voice actors from the local country. Military services and other large organizations have their own distinctive subcultures with specialized terminology and writing style, which we try to respect where possible. We employ a course development methodology that enables us to identify which types of situations are most important, and what cultural skills are needed to master those situations. This course development methodology, called the Situated Culture Methodology (SCM) [3], [12], works as follows and is illustrated in Fig. 3. The first step, outlined in the top left corner of the figure, is to determine scope of the course. This takes into account the situational contexts in which learners are likely to employ their cultural skills. First, it is necessary to identify the intended scope of the course, in terms of the breadth of the geographic region being covered and the duration of the course. To further focus the course, we work with cultural subject matter experts to identify the particular jobs and missions that learners will likely want to train for. This process identifies typical scenarios that require intercultural skills, situations that learners might encounter during those scenarios, and tasks that learners might need to perform in those situations. Given these situations and tasks, it is then possible to identify specific situated culture learning objectives: knowledge and skills that learners need to acquire in order to be effective in those situations. If the learners must employ the local language in these situations, then a set of situated language learning objectives are implied as well.
16
Fig. 3. Situated Culture Methodology.
Once these situated learning objectives are defined, we go through a process of identifying the specific cultural factors that need to be covered in the course to meet these objectives. These factors are grouped in the following general categories: the physical environment or milieu in which the culture is situated, and the political, economic, and social structures common within the culture. The perspectives that are typical of people within the culture are important, such as people’s attitudes toward time, personal relations, work, and the role of the individual within the community. We also identify relevant cultural practices, including norms of conversational interaction and nonverbal communication. To gather this information we draw on published sources, as well as conduct cultural research using a combination of ethnographic interviews and role-playing with subject matter experts. Once the particular cultural factors are identified, they are used to develop the specific curriculum content to be included in the course. Detailed learning objectives are defined for each relevant cultural factor. These typically comprise knowledge about the culture, intercultural skills, and attitudes that are conducive to success in intercultural settings. The knowledge and skills tend to be interrelated; learners first acquire knowledge about the culture and then acquire the skills necessary to apply this knowledge to intercultural situations. We then identify the target performance standards for each learning objective, e.g., demonstrate the ability to apply the cultural skills in particular simulated scenarios. The result is a computer-based curriculum containing lesson pages that introduce language and cultural knowledge, exercises that reinforce knowledge and develop intercultural skills, and simulated scenarios for practicing and assessing those skills. For example, Fig. 4 shows one of the lesson pages in goEnglish that introduces the language and cultural knowledge that learners apply in the simulated small talk
17
scenario in Fig. 2. This page explains that, in modern American culture, being “single” indicates not only that a person is unmarried, but also that they do not have boyfriend or girlfriend. The page also gives the learner the opportunity to practice asking someone whether they are single, and giving an appropriate reply. The learner speaks the phrase into the microphone, and the quality of their speech is scored. Learners advance through a series of progressively more difficult practice exercises until they are ready to practice in a simulated scenario such as those seen in Figs. 1 and 2.
Fig. 4. A goEnglish lesson page.
4 Cultural Modeling Technologies Immersive simulation technologies, together with other supporting technologies, play critical roles in developing and implementing these courses. Knowledge management tools are being developed to capture and organize the cultural knowledge. Authoring tools have been developed to help authors create interactive lessons and activities for teaching cultural knowledge, skills, and attitudes. Immersive simulation technologies,
18
incorporating intelligent agents, are used to simulate culturally appropriate behavior and assess the learners’ ability to behave in culturally appropriate ways.
Fig. 5. Conversational agent architecture
These technologies come together in the conversational agent architecture used in Alelo courses to generate the behavior of the non-player characters in the simulated scenarios. Fig. 5 gives an overview of the agent architecture common to the learning environments described here, and its interaction with the learners and the immersive game environment. Further details are available in [4]. The most fully developed implementation of this architecture is in the Virtual Role Player (VRP) system, a system that enables trainers to populate immersive environments with artificially intelligent non-player characters that are capable of producing culturally appropriate behavior in 3D multi-player training scenarios. This discussion will focus primarily upon the VRP. All Alelo learning environments, are migrating to this architecture.
4.1 Game Engine Platform Game engines commonly incorporate features for creating virtual worlds, populating those worlds with objects and animated characters, and interface controls that enable users to interact with the objects and the animated characters. We build on these to create a uniform set of software platforms for building and running virtual environments for learning culture, with a common set of capabilities. The game engine platforms provide a conversational interface that allows the learners to choose which non-player characters to talk to, and a set of graphical user
19
interface controls for managing the dialog, as shown in Figs. 2 and 4. These include buttons for starting and stopping speech input (top right corner in the figures), and controls for choosing nonverbal gestures for the learner’s character to perform. The interface provides other information such as a summary of the objectives to complete in the scene, hints about what to say and what to do, and a transcript of the dialog. The game engine platform is also responsible for generating the behavior of each character. We record a set of spoken lines for the characters to say and gesture animations for the characters to perform. The characters are then able to perform any combination of speech and gesture, as directed by the social simulation module, which selects the appropriate speech and gestures. The game engine platform includes a simulation management capability that keeps track of the location of objects, characters, and landmarks in the virtual world, many of which have cultural significance. It controls and tracks the movement of people and objects in the virtual world, and can trigger event notifications in response, e.g., cause a non-player to react when a player approaches within a set distance. 4.2 Sociocultural Simulation Module The Sociocultural Simulation Module is responsible for controlling the behavior of the non-player characters in the simulation, and ensuring that they behave in culturally appropriate ways. The agent dialog models follow an interpretationdecision-behavior cycle, in a variant of the SAIBA agent architecture [13]. The first phase of processing starts with interpreting the learners’ spoken utterances and gestures. Acoustic models are trained on both native speech and the speech of learners of the language, to ensure that the speech recognizer is tolerant of the accents of language learners. The recognizers utilize grammar-based language models constructed from a database of sample utterances in the language. As course developers author learning materials, the example utterances in these materials are entered into the database and then used to create the language models. The speech recognition system runs using Julius speech recognition decoder [6]. Next, the agent evaluates and interprets the communicative intent of the learner’s utterance and gesture. Communicative intents are represented using a library of communicative acts, derived from speech act theory originating in the work of Austin [1], and further developed by Traum and Hinkelman [10]. Each communicative act has a core function, i.e., the illocutionary function of the utterance (to greet, inform, request, etc.), and grounding function, i.e., the role of the utterance in coordinating the conversation (e.g., to initiate, continue, acknowledge, etc.). The grounding functions help to determine the current dialog context, which in turn can influence how subsequent utterances are interpreted. At each point in the dialog, the agent is expecting to hear and respond to one of a set of possible communicative acts, which changes over the course of the conversation. If the learner says something that is not appropriate at that stage of the conversation, e.g., greeting a character at the end of the conversation instead of the beginning, the agent will act as if the learner said something odd that does not make sense. The mapping of the utterances and gestures to communicative acts is specified for each culture. Some gestures have meaning only in certain cultures, e.g., placing the
20
palm of the right hand over the heart in greeting only has meaning in Islamic countries. Some gestures are appropriate only in some social contexts; for example, American culture and Arab cultures differ as to when it is acceptable to shake hands with the opposite sex, or kiss the cheek of someone of the same sex. Depending upon the type of dialog exercise, the agent may not just interpret the the learner’s communication, but also identify and classify the learner’s mistakes. Courses commonly include so-called mini-dialog exercises, in which learners practice individual conversational turns with a non-player character and receive feedback regarding any mistakes they may have made. Detected errors include grammatical errors, semantic errors (e.g., confusing words with similar meanings), and pragmatic errors (e.g., inappropriate use of expressions of politeness, honorifics, etc.). Once the learner’s input is interpreted, the intent planning stage occurs, in which each agent in the conversation decides how to respond. Intent planning is challenging because it must address multiple conflicting needs: accuracy, versatility, authorability, and run-time performance. The agents should choose communicative acts that are culturally appropriate, e.g., that match the dialog examples created in the cultural data development process described in section 3. However the agent models cannot simply follow the example dialogs as scripts, but need to be versatile enough to respond in a culturally appropriate way regardless of what the learners might say. The agent modeling language needs to be powerful enough to achieve such versatility, yet be authorable by instructional designers who lack the computer science background required for sophisticated agent programming languages. It also is important for the intent-planning module to have good runtime performance, so that the intent planning process does not interfere with speech processing or 3D scene rendering, both of which require significant amounts of computing resources. Our most common approach for specifying agent intent planning is through finite state machines. This approach is highly authorable and has been used to create hundreds of dialog models for a variety of different cultures and agents. Unfortunately, it is somewhat lacking in versatility; agent models must be authored specifically for each agent in each scenario. To address this limitation, VRP provides a rule-based intent-planning engine that enables authors to specify cultural behavior rules. Each rule specifies a learner communicative act that the agent should respond to, a set of conditions that must be met, and the effects of applying the rule. Conditions can refer to properties of the learner’s communicative act (e.g., its degree of politeness), the state of objects in the virtual world, and properties of the social relationships between the characters in the scenario. For example, the response of a non-player character such as the village elder in Fig. 1 depends upon the extent to which the learner has developed the village elder’s trust, which in turn depends upon what the learner has said and done up to that point. Once each agent in the conversation chooses an action to perform, the sociocultural simulation module generates behavior to realize the action. This typically includes selecting an utterance that realizes the utterance in the target language, as well as one or more animated gestures. These in turn are passed to the game engine platform for execution. Using the agent behavior authoring tools, we create libraries of reusable agent models. Authors can then create cultural training scenarios by taking a virtual world representing the cultural environment, populating it with non-player characters, and
21
assigning agent models to each character appropriate for that character’s role in the scenario. Example roles include village elders, shopkeepers, and passers-by on the street. Each agent will respond to the learners in a manner appropriate for the culture, the agent’s role in the scenario, and the current state of the unfolding situation. Authors can extend and modify the rule sets of the agents if needed to reflect the specific characteristics of the scenario.
5 Current Work We are now working to extend our sociocultural models and apply them to new types of learning environments. An important area of development is in the area of environments that support multiple learners as well as multiple non-player characters. When learners reach a sufficient level of communicative proficiency, multi-player learning environments can become advantageous, as a way to motivate learners and encourage further practice. We are using the Virtual Role Players (VRP) architecture to create multi-player mission rehearsal environments in which teams of trainees work together on mission rehearsal scenarios. Trainees must interact with non-player characters representing local people in order to complete their mission successfully. We plan to extend the VRP framework so that a player can assume control of one of the local people characters and direct its actions without the learners being aware of it. This makes it possible to offer the best possible combination of human-human interaction and human-agent interaction in intercultural skills training. Another learning game under development, the ISLET Game, takes the concept of multi-player game-based cultural learning even further. The ISLET Game is part of a larger language learning environment called ISLET (Integrated System for Language Education and Training), that is exploring methods for making language learning highly motivating, so that learners will devote their free time to learning a language, just as they would an entertainment game. The ISLET Game incorporates a sophisticated reward system, similar to what is found in multi-player entertainment games such as World of Warcraft, that motivates learners to keep practicing and developing their communication skills. It incorporates engaging multi-player quests which encourage collaboration between learners. It also supports team conversations in which learners work together in conversing with a non-player character and learn from each other in the process. The CultureCom project is developing formal models of the cultural influences underlying dialog and utilizing them to increase the flexibility and realism of the behavior of non-player characters in training simulations. Cultural and linguistic anthropologists are developing validated sociocultural data sets for Afghanistan and other cultures of interest, consisting of annotated dialogs of cross-cultural interactions. Experts in artificial intelligence then use these data to develop logical models of sociocultural behavior in different cultures, based upon a formal ontology of microsocial concepts underlying interpersonal communication. This in turn is being used to create an enhanced version of the VRP architecture in which agent intent planning utilizes explicit validated models of sociocultural reasoning for different cultures, which can be swapped in and out to enable agents to model a variety of
22
different cultural characteristics. For example, this is approach is being used to model the differences in which Americans and Afghans express promises and commitments. Such differences contribute to misunderstandings between American and Afghan military personnel in cross-cultural collaborations. The C-CORE project is developing a set of authoring tools and workflow management tools to facilitate the creation and maintenance of cultural models. It enables culture researchers to incorporate sociocultural data from a variety of sources, including subject matter expert first-person accounts, field reports, hand-authored dialogs and dialog transcripts, and other sources. It supports the process of managing, processing, and utilizing streams of updates to sociocultural data to keep the cultural models up-to-date as new information is obtained or properties of the cultural environment changes. It provides tools for comparing the behavior produced by different character models, and for comparing different versions of the same character model resulting from changes in the sociocultural data. The output of C-CORE will be specifications for cultural simulations, including the cultural environment, scenario logic, non-player characters and their behavior. C-GAME, developed in coordination with C-CORE, is extending the overall approach for cultural competence training, and developing it into an interoperable framework that can be utilized with a variety of game engines and artificial intelligence models. The cultural simulations in C-GAME will serve as rich learning environments in which learners can develop their knowledge and awareness of the culture, and then develop intercultural skills, all within the same environment. The scenario construction toolset will enable trainers to develop their own sociocultural training scenarios through a combination of libraries of reusable sociocultural models and authoring tools. It will include mechanisms for incorporating current sociocultural and human terrain data into the scenarios (e.g., from C-CORE), to keep them up-to-date. It will provide a way to annotate a variety of entities in the virtual world in terms of their cultural significance, so that learners can learn their meaning in the context of the culture. This may include clothing and other examples of the material culture, locations (public, private, or sacred places), food and other items associated with cultural practices, and even nonverbal gestures and other human behaviors. These elements will be drawn from an ontology of cultural entities developed in C-CORE, which in turn will relate back to the dimensions of culture shown in Fig. 3. C-GAME will also provide a way for trainers to define trainee cultural performance standards. Trainers will be able to assess those performance standards themselves by participating in or observing the training scenario, and they can rely on automated cultural skills assessments built into the non-player characters. The C-GAME run-time execution engine will incorporate instrumentation to compare trainee performance to the training objectives and provide automated performance review and remediation capabilities. This includes assessment of trainee performance at each stage in the interpretation-decision-behavior cycle of the virtual role players. In the interpretation phase, whenever learners say or do something that is linguistically or culturally inappropriate, the role players will attempt to analyze and classify the nature of the action. This analysis is similar to the analysis that takes place in mini-dialogs, as described in section 5, but can apply to any immersive scenario. Learners will get feedback on their cultural competence both from the
23
moment-to-moment reactions of the non-player characters and from the performance review and the end of the scenario. Acknowledgments. The author wishes to express his thanks to the members of the Alelo team who contributed to this work, and to LeeEllen Friedland for her help in editing. This work was sponsored by USMC PMTRASYS, Voice of America, Office of Naval Research, and DARPA. Opinions expressed here are those of the author and not of the sponsors or the US Government.
References 1. Austin, J.L.: How to Do Things with Words. Harvard University Press, Cambridge, MA (1975) 2. Earley, P.C.: Intercultural training for managers: A comparison of documentary and interpersonal methods. The Academy of Management Journal 30 (4), pp. 685--698 (1987) 3. Johnson, W.L.: A simulation-based approach to training operational cultural competence. In Proc. of ModSIM 2009 (2009) 4. Johnson, W.L., Valente, A.: Tactical Language and Culture Training Systems: using AI to teach foreign languages and cultures. In AI Magazine 30 (2), pp. 72-84 (2009) 5. Kosoko-Lasaki, S., Cook, C.T., O’Brien, R.L.: Cultural Proficiency in Addressing Health Disparities. Jones & Bartlett, Boston (2008) 6. Lee, A., Kawahara, T.: Recent Development of Open-Source Speech Recognition Engine Julius. In Proc. of Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (2009) 7. MCCLL: Tactical Iraqi Language and Culture Training System. MCCLL Newsletter 4 (8), p. 4 (2008) 8. Schneider, P., Sadowski, D.: The effects of intercultural collaboration strategies on successful PhD education. In Proc. of IWIC 2009. ACM Press, New York (2009) 9. Sorcher, M., Spence, R.: The interface project: Behavior modeling as social technology in South Africa. Personnel Psychology 35(3), pp. 557—581 (1982) 10.Traum, D., Hinkelman, E.: Conversation acts in task-oriented spoken dialogue. Computational Intelligence 8, pp. 575–599 (1992) 11.US Dept. of Health and Human Services (USDHHS): Cultural competency curriculum for disaster preparedness and crisis response, https://cccdpcr.thinkculturalhealth.org/ 12. Valente, A., Johnson, W.L., Wertheim, S., Barrett, K., Flowers, M., LaBore, K., Johansing. P.: A Dynamic Methodology for Developing Situated Culture Training Content. Technical Report, Alelo (2009). 13. Vilhjálmsson, H. and Marsella, S.:Social Performance Framework. In Proceedings of the AAAI Workshop on Modular Construction of Human-Like Intelligence (2005)
24
Creating Virtual Synthetic Cultures for Intercultural Training Samuel Mascarenhas and Ana Paiva IST - Technical University of Lisbon and INESC-ID, Av. Prof. Cavaco Silva, Taguspark 2744-016, Porto Salvo, Portugal {samuel.mascarenhas,ana.paiva}@gaips.inesc-id.pt
Abstract. Currently, interactive virtual environments for inter-cultural training are mostly designed to train in a specific target culture, focusing on the communication aspects of that culture. We are currently pursuing a different approach, one where users learn to broadly cope with culture shock as they interact with virtual agents that can enact different synthetic cultures. To create such agents we propose an architecture that integrates cultural aspects in the way the agent feels and chooses its goals and actions, based on anthropological studies. This architecture was applied in an agent-based educational role-playing game that tries to promote inter-cultural empathy in young teenagers. Key words: Virtual Agents, Cultural Behaviour, Educational Games
1
Introduction
Research in intelligent virtual environments designed for inter-cultural training is strongly emerging. Current technology offers the possibility to create 3D virtual environments in which the visual and auditory aspects of a culture (architecture, clothes, artefacts, language, gestures) can be simulated with a great deal of realism. Moreover, in virtual environments users can safely interact with autonomous embodied agents that can display human-like behaviour, in particular behaviour that is culture-specific. However, the creation of autonomous agents that are able to enact various socio-cultural contexts is still hard and so far the results are limited. As reported in [13], most virtual environments for cultural training, such as TLCTS [10], CUBE-G [20], Elect BiLat [7], VECTOR [3], Second China [6] or Croquelandia [23], focus on training verbal and non-verbal communicative aspects of a culture. Although communicative aspects are a fundamental issue to consider in intercultural training it is also important that users learn skills on how to cope with other differences such as different value orientations. In this paper, we propose an agent architecture that aims to generate such differences, integrating cultural aspects in the way the agents behave, not only in their gestures or communication styles, but also in their goals, emotions, choices, ways of reacting to the environment, among others.
25
An important inspiration for this work comes from the work conducted in developing simulation games for inter-cultural training. Probably the best known example is the BaF a ´ BaF a ´ simulation [22], where participants are randomly assigned to two synthetic cultures (Alpha and Beta) that differ in their core values. One is a more collectivist and hierarchical culture that stresses socialising and touching, and the other, uses a different language, is more individualistic and task-oriented. In the first phase of the simulation, each group is sent into separate rooms in which they are instructed to learn and rehearse their assigned culture. Afterwards, the participants take turns in visiting the other culture, try to gain an understanding of it, and then come back to their own culture to brief the other members about their experience. During the simulation, participants become aware of having a natural tendency towards ethnocentrism as they normally judge the other culture as “weird” compared to their own, despite the fictional nature of both cultures. As in a real cultural shock experience, the simulation often evokes feelings of bewilderment, disorientation, and exclusion. The concept of synthetic cultures was also proposed as an inter-cultural training tool in [9]. In this work, synthetic cultures were also defined as a simplification of real cultures, emphasising differences in values. After years of experience with simulation role-playing games using these synthetic cultures, it was concluded they were a useful tool for learning about cross-cultural communication [9]. In a comprehensive review of various inter-cultural training techniques [12], some of the advantages described for simulation games are that they eliminate the gap between learning and applying, they provide an opportunity to practice new behaviours in a safe haven, are highly versatile and experience stays with trainees. Their disadvantages are that simulations normally require a large number of human participants, consume a lot of time and some people may be too shy to fully participate in these type of activities. Interactive virtual environments can help ameliorate these issues by allowing users to safely interact with virtual agents and thus eliminating the dependency on other human participants. Hence, we argue that by creating an agent architecture that facilitates the creation of various virtual synthetic cultures that can capture the essence of value differences in real cultures, it is possible to develop new fruitful and enjoyable ways of inter-cultural training. The structure of this paper is described as follows. In the next section, the culture theory used to ground our approach is presented. In section 3, the integration of the cultural elements into an agent architecture is discussed. In section 4, an agent-based educational game designed with this architecture is presented along with a more simple scenario used to evaluate the architecture. Finally, we draw some conclusions and present some future work.
2
Background
Edward B. Tylor, often considered to be the founder of anthropology, defined culture in 1871 as “that complex whole which includes knowledge, belief, art, law, morals, custom, and any other capabilities and habits acquired by man as
26
a member of society” [25]. Since then, several other definitions for culture have been proposed. In 1952, a list containing 164 possible definitions of culture was compiled. [11]. Still, no consensus has yet been reached. The particular notion of culture adopted in this work is grounded on the dimensional theory of culture proposed by Geert Hofstede [8]. In this theory, culture is defined as “the collective programming of the mind that distinguishes the members of one group or category of people from another” [8]. These “mental programs” refer to patterns of thinking, feeling, and potential acting that are shared and learnt by members of the same culture. The foundation of Hofstede’s cultural theory is a large empirical study conducted in more than 70 countries. From this study, he was able to derive a set of five bipolar dimensions: Individualism-Collectivism, Power Distance, Uncertainty Avoidance, Masculinity-Femininity, and Long Term Orientation-Short Term Orientation. Each of these dimensions represent fundamental differences in cultural values across nations. Generally, values can be defined as a “broad tendency to prefer a certain state of affairs over others” [8]. They transcend specific situations, guide selection or evaluation of behaviour and events, and are ordered by relative importance [21]. Another particular aspect of values is that they are often unconscious to those who hold them and so they cannot be directly observable. Instead, they have to be inferred from the way people act under various circumstances. However, besides values there are also other types of cultural manifestations [8] that, unlike values, are explicit and more clearly observable: (1) Rituals - essential social activities that are carried out in a predetermined fashion; (2) Heroes real or even imaginary persons that serve as models for the cultural values; and (3) Symbols - words, gestures, pictures, or objects with a special particular meaning. For the moment, our cultural model includes rituals, symbols and two of the five dimensions: Individualism-Collectivism and Power Distance. We chose to start with these two dimensions because they seemed to be the ones more easily recognisable in a short-term interaction and they also seem to be the ones most agreed upon in the literature. In particular, several studies that independently measure Individualism-Collectivism show a reasonable correspondence to Hofstede’s findings [21, 2, 24]. The Individualism-Collectivism dimension indicates the extent to which individuals see themselves integrated into groups. The more individualistic a culture is the more people stress the importance of personal achievements and individual rights, and everyone is more expected to be responsible only for themselves and their immediate family. Conversely, in highly collectivistic cultures, everyone looks out for one another in exchange for unquestioning loyalty. As for the Power Distance dimension, it indicates the degree to which less powerful members of the group expect and accept that power is distributed unequally. In low power distance cultures, people tend to regard others as equals despite their formal status, while in high power distance cultures powerful people have more privileges and like to wear symbols that reflect their status. Note that these behavioural tendencies indicated by the cultural
27
dimensions should never be considered deterministic, since other factors such as personality also play an important role on determining behaviour. The main advantage of using Hofstede’s culture theory is that it was based on a large empirical study of national cultures and it gives a clear and detailed notion of differences in values between them. Even though the theory has received some criticisms [17] such as being based on the supposition that within each nation there is a uniform national culture that remains static, it still serves our purposes, which is to characterise important cultural differences and not to replicate real cultures in a dynamic and exact way.
3
Cultural Architecture
The agent architecture developed so far for creating agents with different cultural profiles is shown in Figure 1. It was implemented by extending FAtiMA, an emotional agent architecture [4, 14] that follows the OCC model of emotions [19] for creating believable virtual characters. In the resulting architecture, there are three kinds of cultural parameters that influence the agents’ behaviour: Symbols, Dimensions and Rituals.
Fig. 1. Cultural Agent Architecture
28
The architecture works in the following manner. When an event is perceived, it passes through a Symbol Translator that translates the meaning of the event according to the culture’s predefined symbols, using a simple association mechanism. For instance, when an agent performs a “thumbs-up” gesture, in one culture it can be associated to an “approval” meaning, while in another culture, it can be associated to an “insult” meaning. To avoid ambiguity, the mapping between physical actions and meanings is a one-to-one relationship, i.e. a particular action has only one meaning and vice versa. This is an assumed simplification of the real world, where in fact the same physical action can have different meanings in the same culture due to different contexts (e.g. a bow can be a form of greeting but also a sign to acknowledge the applause for performing a play). After the event is translated, it is then used to update the Knowledge Base (KB) and Autobiographic Memory (AM). These are the main memory components of the agent. The first one is responsible for storing semantic knowledge such as properties about the world and relations, while the second one stores information concerning past events and the agent’s personal experience. At the same time the memory components are updated, the event is used to update the agent’s Motivational State. Agents have five different motivational needs that are represented by five continuous drives: Energy, Integrity, Affiliation, Certainty and Competence. These general drives are grounded on a psychological model of human action regulation called PSI [5]. To determine how events affect needs, the actions that agents can perform in the environment are authored with positive and/or negative effects on the motivational drives. To determine if other agents have their needs satisfied or not each agent also builds and updates a Motivational State of Others according to the events perceived. This information is inferred from the agent’s perception of the other agents’ actions and of their initial motivational states. After updating the motivational states, the event is appraised in order to determine the emotional response of the agent. There are two main appraisal processes, the Deliberative Appraisal handles emotions related to the achievement of goals (e.g. satisfaction, disappointment), and the Cultural Reactive Appraisal associates appraisal values to the event perceived and then generates the corresponding emotions. The Individualism score, defined in the agent’s cultural profile greatly affects the Cultural Reactive Appraisal. If the culture is defined with a very low Individualism score the more an event that is undesirable for others but is beneficial for the responsible agent will be blameworthy (e.g. stealing something), which will likely make the agent to feel ashamed. Also, the more an event that is good for others but is bad for the responsible agent (e.g. giving food) will be highly praiseworthy, making the agent feel proud of himself. In other words, the more collectivistic an agent is the more it considers self-sacrifice for the well-being of the group as highly admirable and selfish acts as highly reproachable. For more information on how the Individualism dimension affects the agent’s appraisal, please refer to [16]. In the deliberative layer, the event perceived can activate predefined goals, and the agent will have to select between competing alternative goals. In this
29
layer, the Cultural Goal Selection is the process responsible for calculating the expected utility for each active goal, considering the Individualism and Power Distance dimensions, and the expected impact the goal has on the agent’s motivational needs and on the needs of others. The exact equations are described in [16], but the general idea is the more Individualism and Power Distance the culture has, the more agents prefer goals that benefit their own needs or the needs of agents with higher status. Conversely, the less Individualism and Power Distance a culture has, the more the agents will care for the needs of other agents, regardless of their social statuses. To illustrate these differences consider a situation where an agent asks for help to paint his house. If the culture is defined as collectivistic and has a low power distance, all the other agents will likely offer their help. On the other hand, if the culture is highly individualistic with a high power distance, agents will tend to help only if the other agent has an important social status. After the goal with the highest expected utility is chosen, the agent forms an intention of achieving that goal and uses the Planner component to develop and execute a plan. The architecture has also a Ritual Manager for dealing with the activation and execution of cultural Rituals. The model for Rituals was inspired by plan recipes used in traditional BDI architectures with a fundamental difference: traditional plans are based on technical activities (the focus is in the end result), whilst rituals are based on ritual activities (the focus is in the sequence of steps). As such, a ritual has a set of roles associated with it and each role has one or more steps that must be performed following any specified ordering constraints. For more details on how Rituals were implemented, please refer to [15].
4
Case Study
The implemented cultural architecture was used for the development of a serious game called ORIENT (Overcoming Refugee with Empathic Novel Tecnhology) [1]. The game is an agent-based educational role-play, developed in the context of an EU-funded project called eCIRCUS1 . The main purpose of the game is to promote inter-cultural empathy for young teenagers. In this game (see Figure 2), players (assuming the role of space travellers), must interact with an unfamiliar fictional foreign culture whose planet is about to be destroyed by a large meteor. The main objective for the players is to gain the trust of the culture to then save them from annihilation. To gain their trust, players have to become familiar with the culture’s strange customs and gestures. For instance, they must understand that the culture is strongly hierarchical and everyone is highly compassionate and loyal to each other. In order to create this culturally specific behaviour, we applied our architecture to define the culture’s gestures and rituals. Moreover, their culture was parametrised as highly collectivistic and with a high power distance score. 1
www.e-circus.org
30
Fig. 2. Screenshot of the game.
ORIENT also has an innovative approach in terms of user interaction. It allows three users to interact simultaneously, each one controlling one of the following devices: a Dance Mat, a mobile phone and a WiiMote. Each device has a different but essential function: (1) the Dance Mat is used for navigation purposes; (2) the mobile phone is used for verbal communication and object recognition; and finally (3) the WiiMote is used to perform important cultural gestures that are used for instance, in the greeting rituals of the culture. The rationale for allowing a group of users to interact simultaneously was to promote social collaboration. A second objective was to encourage discussion between players about the cultural differences found in the synthetic culture. Also, the use of novel interaction devices was to incite players’ curiosity to play the game and to provide a more engaging experience. After conducting two pilot studies [1], users found the Sprytes to be a very different culture from their own and most users were interested in the storyline, even though it was considered to be too short and too simple. One key issue that emerged from the pilot study was that even though users identified Sprytes to have a very different and strange culture, users felt at ease during the interaction with Sprytes, considering them to be a friendly, peaceful, trusting, happy, relaxed, natural and social culture. As such, unlike simulation games for inter cultural training such as Baf a ´ Baf a ´ [22], ORIENT did not provide a strong cultural shock experience, which would consist in users experiencing some feelings of disorientation, uncertainty and confusion during the interaction. This can be explained by the fact that Sprytes were authored in a way that encouraged the user to like them so the user would become motivated in his mission to save the culture.
31
Furthermore, due to time constraints, we were unable to add other cultures in ORIENT’s story for users to interact with. Hence, we were unable to use it to measure the power of our architecture in creating distinct virtual synthetic cultures. For that reason, we developed a smaller non-interactive scenario, specifically designed for this purpose. The scenario consisted of a dinner party with a simple plot: five virtual agents with different social status arrive at a party location, greet each other, socialise for a while, and then sit together at a dinner table and start to eat. Using our architecture we defined different cultural profiles which were then associated to the same group of agents that enacted the same dinner party situation. The objective was to assess if users could perceive cultural differences between two groups of agents just by exclusively changing their cultural parametrization. Two separate experiments were conducted. In the first one, discussed in [15], two cultures were created that only differed in their rituals (and associated symbols), inspired on the opposite extremes of the Power Distance dimension. For example, a dinner ritual was defined for both cultures with the following differences: in the low power distance culture everyone rushed to the table immediately, not even waiting for the host to finish the announcement saying the dinner was about to start, while in the high power distance culture everyone waited first for the elder to sit before they could sit, and then waited for the elder to start eating before they could eat. In the second experiment, described in [16], the two created cultures only differed in the parametrization of the Individualism-Collectivism dimension. In both cultures the agents had the same available goals to choose from. To exemplify the differences between the cultures, consider the following situation in the scenario. After the characters greeted each other, one of the characters who is sick, reports about his sickness to the other characters. One of the other characters has medicine but is not a friend of the sick character. In the highly collectivistic culture, the character who has medicine will promptly offer it in order to help, feeling pride afterwards. Conversely, in the highly individualistic culture, the sick character is not given any medicine. This particular situation reflects the fact that, as stated in [8], in collectivistic cultures people tend to always look out for one another while in the individualistic cultures people assume that they are only responsible for those they share a close bond. Both experiments had approximately 40 participants (the majority was Portuguese) and both shared a similar methodology . Users observed the two different cultures enacting the dinner party scenario and then were asked to choose from a list of possible values and adjectives the ones they saw as more fitting to describe the behaviour of each group. In the end, users were asked if they found differences between the groups and if so, if those differences were due to the character’s personality or due to the character’s culture. Both experiments gave significant yet different results. We found that the dimensional model was capable of differentiating cultures in terms of their inferred values, yet the differences were not interpreted as cultural. On the other hand, the rituals component was capable of leading to the perception of different cultures, yet few value dif-
32
ferences were identified. These results were somewhat expected because as Hofstede’s theory points out, the values associated to the cultural dimensions are often unconscious to those who hold them and thereby are harder to interpret as cultural by the average person than rituals or symbols.
5
Conclusion and Future Work
In this paper we presented an ongoing work for designing an agent architecture that integrates cultural phenomena not only related to communication aspects such as gestures, but also to more high level behaviour, such as decision-making and emotional appraisal processes. The aim of the architecture is to facilitate the creation of different cultures of virtual agents that are able to enact general cultural differences and can be used in a similar manner as synthetic cultures are used in simulation games for inter-cultural training. The proposed architecture was used to drive the behaviour of an alien culture with strange customs and beliefs, which users learn to cope within a game designed to promote intercultural empathy. Additionally, a smaller non-interactive scenario with different cultures was also built for evaluating the architecture. Two experiments were conducted that showed the architecture was powerful enough to create cultures that were perceived and characterised significantly different by users. As future work, we would like to integrate other important cultural aspects in the architecture as well as improve the existing ones. For instance, as described in [18], there are several other relations between culture and emotions which would be interesting to include. One example is the notion of cultural display rules (how should one act when experiencing certain emotions). Also, we would like to use the architecture in richer interactive scenarios designed for intercultural training. In particular, we are interested in applying the architecture in a scenario where users need to interact with more than one virtual culture, learning to adapt to their different values.
6
Acknowledgements
This work was partially supported by a scholarship (SFRH BD/62174/2009) granted by the Funda¸c˜ ao para a Ciˆencia e a Tecnologia (FCT) and by the European Community (EC) and was funded by the eCIRCUS project IST-4-027656STP with university partners Heriot-Watt, Hertfordshire, Sunderland, Warwick, Bamberg, Augsburg, Wuerzburg plus INESC-ID and Interagens. The authors are solely responsible for the content of this publication. It does not represent the opinion of the EC or the FCT, which are not responsible for any use that might be made of data appearing therein.
References 1. Ruth Aylett, Ana Paiva, Natalie Vannini, Sybille Enz, Elisabeth Andre, and Lynne Hall. But that was in another country: agents and intercultural empathy. In Proceedings of AAMAS 2009), Budapest, Hungary, 2009. IFAMAAS/ACM DL.
33
2. S. Bochner. Cross-cultural differences in the self concept: a test of hofstede’s individualism/collectivism distinction. Journal of Cross-cultural Psychology, 25:273– 283, 1994. 3. J.E. Deaton, C. Barba, T. Santarelli, Larry Rosenzweig an Vance Souders, Chris McCollum, Jason Seip, Bruce Knerr, and Michael Singer. Virtual environment cultural training for operational readiness (vector). Virtual Reality, 8(3):156–167, 2005. 4. J. Dias and A. Paiva. Feeling and reasoning: a computational model for emotional agents. In Proceedings of 12th Portuguese Conference on Artificial Intelligence, EPIA 2005, pages 127–140. Springer, 2005. 5. Dietrich Dorner. The mathematics of emotions. In Prooceedings of the Fifth International Conference on Cognitive Modeling, pages 75–79, Bamberg, Germany, 2003. Bamberg, Germany. 6. J. Henderson, P. Fishwick, E. Fresh, F. Futterknecht, and B.D. Hamilton. Immersive learning simulation environment for chinese culture. In In Proceedings of Interservice/Industry Training, Simulation, and Education Conference, 2008. 7. Randall W. Hill, James Belanich, H. Chad Lane, and Mark Core. Pedagogically structured game-based training: Development of the elect bilat simulation. In In Proceedings of the 25th Army Science Conference, 2006. 8. G. Hofstede. Culture Consequences: Comparing Values, Behaviors, Intitutions, and Organizations Across Nations. Sage Publications, Thousand Oaks, 2001. 9. G.J. Hofstede. Role playing with synthetic cultures: the evasive rules of the game. In Proceedings of the 9th International workshop of the IFIP, pages 49–56. Helsinky University of Technology, SimLab Report no 10, 2005. 10. W. Lewis Johnson, Hannes H¨ ogni Vilhj´ almsson, and Stacy Marsella. Serious games for language learning: How much game, how much ai? In Chee-Kit Looi, Gordon I. McCalla, Bert Bredeweg, and Joost Breuker, editors, AIED, volume 125 of Frontiers in Artificial Intelligence and Applications, pages 306–313. IOS Press, 2005. 11. A.L. Kroeber and C. Kluckhohn. Culture: A Critical Review of Concepts and Definitions. MA: Peabody Museum, Cambridge, 1952. 12. Daniel Landis, Jannet M. Bennet, and Milton J. Bennet. Handbook of Intercultural Training. Sage Publications, Inc, 2003. 13. H. Lane and Amy Ogan. Virtual environments for cultural learning. In Second Workshop on Culturally-Aware Tutoring Systems in AIED 2009 Workshops Proceedings, pages 25–34, 2009. 14. Mei Yii Lim, Jo˜ ao Dias, Ruth Aylett, and Ana Paiva. Improving adaptiveness in autonomous characters. In Helmut Prendinger, James C. Lester, and Mitsuru Ishizuka, editors, IVA, volume 5208 of Lecture Notes in Computer Science, pages 348–355. Springer, 2008. 15. S. Mascarenhas, J. Dias, N. Afonso, S. Enz, and A. Paiva. Using rituals to express cultural differences in synthetic characters. In Proceedings of AAMAS 2009, Budapest, Hungary, May 2009. IFAMAAS/ACM DL. 16. S. Mascarenhas, J. Dias, R. Prada, and A. Paiva. A dimensional model for cultural behaviour in virtual agents. International Journal of Applied Artificial Intelligence: Special Issue on Virtual Agents, 2010. 17. Brendan McSweeney. Hofstede’s model of national cultural differences and their consequences: A triumph of faith - a failure of analysis. Human Relations, 55(1):89– 118, 2002. 18. Batja Mesquita and Nico Frijda. Cultural variation in emotions: A review. Psychological Bulletin, 112:179–204, 1992.
34
19. A. Ortony, G. Clore, and A. Collins. The Cognitive Structure of Emotions. Cambridge University Press, UK, 1988. 20. Matthias Rehm, Nikolaus Bee, Birgit Endrass, Michael Wissner, and Elisabeth Andr´e. Too close for comfort?: Adapting to the user’s cultural background. In HCM ’07: Proceedings of the international workshop on Human-centered multimedia, pages 85–94, New York, NY, USA, 2007. ACM. 21. S. H. Schwartz and W. Bilsky. Toward a theory of the universal content and structure of values: Extensions and cross-cultural replications. Journal of Personality and Social Psychology, 58:878–891, 1990. 22. R.G Shirts. Beyond ethnocentrism: Promoting cross-cultural understanding with bafa bafa. In S. M. Fowler and M. G. Mumford, editors, Intercultural Sourcebook: Cross-cultural training methods, pages 93–100. Yarmouth, Intercultural Press, 1995. 23. Julie Sykes. Learner requests in spanish: Examining the potential of multiuser virtual environments for l2 pragmatic acquisition. Computer Assisted Language Instruction Consortium, 2009. 24. H. Triandis, C. McCusker, and C. H. Hui. Multimethod probes of individualism and collectivism. Journal of Personality and Social Psychology, 21:1006–1020, 1990. 25. E.B. Tylor. Primitive culture: researches into the development of mythology, philosophy, religion, art, and custom. New York Gordon Press, 1871.
35
36
CAMPERE: Cultural Adaptation Methodology for Pedagogical Resources in E-learning Franck Mpondo Eboa, François Courtemanche and Esma Aïmeur Department of Computer Science and Operations Research University of Montreal, Quebec, Canada {mpondoef, courtemf, aimeur}@iro.umontreal.ca Abstract. Intelligent Tutoring Systems (ITS) are increasingly used around the world for distance learning. The popularization of internet made pedagogical resources accessible directly to students within their own environment. Many sociology and psychology studies show that our social environment and life experience have an impact on how we interpret information or given situations. However, most intelligent tutoring systems present the learning content without regard to learner’s cultural background. This paper presents a resource personalization technique for cultural adaptation within ITSs. The approach is based on a collaborative filtering technique using an implicit cultural profile, which is automatically updated using the learner’s interactions with the system. Keywords: Cultural adaptation, collaborative filtering, user interaction.
1
Introduction
Education has benefited from distance learning in many ways [10]: a) ensuring knowledge sharing between geographically distant people, b) expanding educational institutions’ recruitment pool for both students and teachers, c) fighting brain drain and uprooting in countries where high quality education is not widely provided. Despite many great successes, most intelligent tutoring systems for distance learning have an important limitation restricting worldwide scale use: the lack of pedagogical adaptation to the learner’s socio-cultural context [8]. Indeed, although existing systems do personalize the learning content with regards to the learning style and objectives of the learner, most of them display the learning content to different learners without regard to their cultural environment. Existing research [3, 11, 16] highlight the fact that our mental programming – how we act, think, learn and interpret – is conditioned by our social circle, the countries in which we grew up, etc. For instance, the word “football” means American football in North America, while in Europe or South America it stands for the so called Soccer. Resources standardization may therefore lead to misinterpretations causing additional frustrations to the learner. To deal with this issue, cultural characteristics must be considered during the design process of distance ITSs [16]. Designers should focus on elements (examples, images, videos, etc.) which describe a specific reality in learner's socio-cultural context and avoid content that could discourage the learner or generate negative emotion during the learning process.
37
For example, if we talk about “team spirit” in a lesson, the system must adapt the pedagogical resources according to the learner’s hobbies (e.g.: images of basketball, hockey players, or engineers working together). On the other hand, if we talk about “watercourses” it is judicious to use references like “Mississippi”, “Yukon” or “Niagara” rivers for a North American learner. To give the impression to an African learner that the system has been designed for him these examples must come after “Nile”, “Draa” or “Niger” rivers which are among the most famous rivers in the African Continent. The challenge here is to find an effective way of defining cultural profiles associated to each learner in order to select appropriate resources. But what is really a cultural profile? Does it depend on our nationality, level of education, religion, hobbies, profession, etc..? Most studies [9, 11, 14, 16] agree to say that culture is a complex and in tangible concept [14] and that the creation of homogenous cultural groups for cultural classification of learners requires an important and tedious survey process [16]. Given that, we propose a cultural adaptation approach that bypasses this difficulty using a two-step technique: a) A minimal amount of information about the learner is gathered to initialize the adaptation process b) A collaborative filtering technique is used to adapt pedagogical resources using the learner’s cultural profile, which is dynamically updated by his/her interactions with the system. The paper is organized as follows: section 2 defines the concept of culture and refers to some initial works on modeling learner’s cultural profile. Section 3 reveals the objectives and motivation of the paper. Section 4 introduces the collaborative filtering technique. Section 5 presents our cultural adaptation approach which is based on an expert knowledge and collaborative filtering technique. Conclusion section closes the paper, outlining the originality of our approach and further research that still needs to be carried through to improve it.
2
Related Work
The student model is an ITS’s modules representating the learner’s cognitive, psychological and emotional profile. It models a student’s knowledge, styles and preferences in order to allow providing appropriate help, choosing the next step in the learning process and presenting material in a way that matches the learner’s preferences. Culture is a set of artefacts, practices, ideational elements and specific world representations emerging mainly at the group level [3]. As shown in [7], [9] and [17] culture as an impact on our cognitive and psychological processes. Scollon [18] and Kim-Prieto [12] recently showed that classification, frequency and interpretation of some emotions depend on our cultural background. These researches underline the necessity of integrating a cultural variable in a model student to adapt content, structure of the information or learning goals to the learner’s cultural profile. In the literature many works try to adapt the pedagogical resources (example, videos, and images) according to the learner’s cultural background.
38
One of the first suggestions is given by the CAWAS system in [5]. CAWAS generates a cultural profile using Hofstede’s dimensions and associated scores [9]. CAWAS is a rule based system and each rule generated is weighted to reflect the relevance according to a specific cultural group or an individual learner. The weights of the rules are periodically refine by a neural network as users interact with the system. In 2008, [22] proposed a template for extending ITSs’ authoring tools to be culturally aware. Templates are supported by many variables. Each variable is link with the others by an association taken the form “x_depends_on_y” rules. When comes the time to choose a resource for a learner, the system looks for, among the rules, variables whose values correspond to the learner’s information. Many other approaches [3, 4, 15, 16] try to draw knowledge from experts and target people in order to conceptualize, in an ontology, universal aspect describing a person sociocultural profile. Ontology can be then used to infer knowledge and adapt resources to the learner’s cultural profile.
3
OBJECTIVES
Culture is a complex and intangible concept [14]. All attempts to model learner’s cultural profile using universal and tangible statements, criteria or rules are very tedious [16]. Therefore, the main goal of our research is to propose a cultural adaption technique for pedagogical resources, to teach declarative concept, without requiring a large amount of explicit information. As the nature of cultural background is very elusive, our approach implicitly gathers learner’s preferences in order to build a cultural profile. Since a learner’s culture is subject to change overtime, the implicit cultural profile is updated automatically as learners interact with the system. As culture refers to a set of values, beliefs, behaviors and practices that characterize a given group of individuals [11], we consider that learners sharing similar cultural background have the same way to interact with the system and react similarly to pedagogical resources. There isn’t an effective way to conceptualize all universal characteristic describing a learner cultural background. A cultural background can also be influenced by variables that are not directly cultural factors, for example gender or age [14]. So far there are no formulated rules to determine a learner's cultural background; there are only implicit principles very difficult to define clearly. To gather implicit learners’ preferences we use a proven method, called collaborative filtering, working well with high level of subjectivity. Collaborative filtering technique allows us to identify cultural preferences that the learner can’t express himself clearly, or that an expert is unable to model.
39
4
COLLABORATIVE FILTERING
Collaborative Filtering includes all methods designed to recommend items enjoyed by other users presenting similarities in their tastes and preferences with the target user [2]. It is a proven method for item recommendations without requiring explicit preferences or knowledge of the items [6]. Most methods require users’ interactions and retain his/her opinions to build a profile. Users offer feedback on items, which is then used to predict the preferences for unseen items and subsequently for recommending items with the highest predicted relevance [1]. Collaborative filtering usually requires two successive phases. First, the neighborhood phase compares the target profile with all the other users stored in the database to choose, based on a similarity function (Pearson correlation or similarity vector), the most similar users (same evaluation or opinion on a set of items). Second, once the similarity is calculated, the system infers the appropriate item for the target user based on his/her common preferences with the neighborhood.
5
ADAPTATION TECHNIQUE
In our approach, each problem consists of a web page presenting a given topic. The domain knowledge included within a problem is presented using a series of concepts. In a problem a concept can be presented to the learner using different materials (images, texts or videos). We name pedagogical resource a particular material used to present a concept. For example, within a problem about nutrition the concept of “Carbohydrate-rich foods” can be presented as an image of maple syrup (Canada), bok choy (China), potatoes (US) or cassava (Africa). Adaptation process is done in two steps: the first step is based on expert knowledge and allows us to select resources event if we don’t have information about learners’ interactions with the system. The second step uses the collaborative filtering technique, where the learner has to interact with the system, for selecting resources. 5.1 Knowledge based adaptation Each pedagogical resource in the domain knowledge is tagged by a domain expert (arrow 1, Fig.1) in order to establish its relevance according to different country. When logging for the first time, the learner provides his/her adolescence country in order to initialize a temporary cultural profile used to reduce the cold start effect. Therefore, for a new user the choice of pedagogical resources to integrate in a problem is based on his/her temporary profile (arrow 2, Fig.1).
40
INTERFACE
EXPERT
TUTOR
LEARNER
Adaptation modules
Tagging resources
Knowledge based
1
Collaborative filtering
3
4
2
Pedagogical resources
Update
Cultural profiles
Fig. 1 – System architecture
5.2 Collaborative Filtering based adaptation People’s mental programming varies often considerably within a same nation [17, 19]. It is therefore unwise to rely solely on country or nationality to determine a learner’s cultural profile. A cultural background may depend on several factors including [14]: the environment in which one spent most of adolescence, his/her hobbies, family education, religion and degree of worship, language, countries of residence, etc. However, it is difficult to get all this detailed information explicitly from a learner and the expert. To overcome this problem the main adaptation module is based on a Collaborative Filtering (CF) technique (Fig.1). The CF adaptation module allows to implicitly inferring the learner’s cultural preferences over the pedagogical resources. The cultural profile used by the CF module contains the learner’s appreciation of all pedagogical resources associated to every concept in the domain knowledge (Fig. 2). Pedagogical resources
Concepts
C1 C2 C3 : :
R1 1 R5 0 R10 2
R2 R3 R4 3 1 2 R6 R7 R8 0 3 1 R11 R12 1 4
Example C1 = “Carbohydrate-rich foods” R1 = “Maple syrup” R2 = “Bok choy” R3 = “Potatoes” R4 = “Cassava”
R9 1
Fig. 2 – Implicit cultural profile
41
During a learning session the learner can change directly in the problem interface a pedagogical resource that is not culturally suited, from a set of equivalent resources for the same concept (e.g.: changing a maple syrup image for a cassava image). The learner’s cultural profile is then dynamically updated (arrow 3, Fig.1). The newly selected resource’s scoring is incremented by one (R4 = 2+1 = 3) and the changed resource’s scoring (R1) remains the same. As learners sharing the same cultural background understand concepts using the same frame of reference they will have similar resource scoring. The pedagogical resources integrated in problems are selected by the collaborative filtering adaptation module (arrow 4, Fig.1). The collaborative filtering algorithm selects pedagogical resources that were appreciated by learners with the same or close cultural background. A cultural similarity coefficient is computed between a target learner and all other learners using the Pearson correlation defined by the following equation:
Sim (t,u) =
𝑚 𝑖=1(𝑠𝑡,𝑖 −𝑠𝑡 )×(𝑠𝑢 ,𝑖 − 𝑠𝑢 ) 𝑚 2 𝑖=1(𝑠𝑡,𝑖 − 𝑠𝑡 )
2 × 𝑚 𝑖=1(𝑠𝑢 ,𝑖 − 𝑠𝑢 )
(1)
Where st,i is the target learner’s score for the resource i and su,i stands for another learner’s score for the same resource i. The cultural similarity coefficient is computed over the m resources in the domain knowledge. Two users are considered as similar if the similarity (equation 1) is greater or equal to 0.3 (Sim >= 0.3). 5.3 Experimental validation To test our approach, we have developed an e-learning application presenting a set of problems to the learner. Each learner must provide the country where he/she has spent much of his/her adolescence. This information is used by an expert during the resource tagging process. The experimentation includes two steps: the training step which allows us to build the learner’s cultural profile using his/her interaction with the system. The test step is used to evaluate our approach accuracy. To do this, we compare our predicted resource with the resource selected as being appropriate by the learner. The validation of the approach is currently in progress. 5.3.1
Training step
During the training step, problems are presented to learners using his/her temporary profile, that is to say using the expert knowledge (Section 5.1). Figure 3 shows an example of the adaptation of the “watercourses” problem for a North American learner.
42
Fig. 3 – “rivers” concept for a North American learner
Learner can click on the text (Mississipi River, Amazon River, etc.) or on the image to change the resource. After clicked on the resource, the learner sees all the resources related to the concept and can change it (Fig.4).
Fig. 4 – Modification of “The Mississipi River” text resource
Figure 5 illustrates an example of the “rivers” concept for an African learner, rivers examples (Amazon River, Nile, Niger and Draa River) and an image different from a North American learner. According to the expert, Yukon River or Mackenzie River is not relevant for an African learner. In the same way, Congo River or Niger River is not appropriate for a North American learner.
43
Fig. 5 – “rivers” concept for an African learner
5.3.2
Test step
In the test step we present a set of concepts to the learner. For each concept, learner has to choose the most appropriate resource between the available resources for the given concept. For each resource available for the concept, the system predicts the learner score in the collaborative filtering module (Fig.1) using the following equation:
P (t,i) = 𝑠𝑡 +
𝑛 (𝑠 −𝑠 ) × 𝑆𝑖𝑚 𝑢 𝑡,𝑢 𝑢 =1 𝑢 ,𝑖 𝑛 𝑆𝑖𝑚 𝑡,𝑢 𝑢 =1
(2)
Where Pt,i is the target learner’s predicting score for the resource i. Simt,u is the similarity between the target learner t and neighbor u (see equation 1, section 5.2). n represents the size of the target learner’s neighborhood ( in our application n = 3), and Su,i stands for the neighbor’s score for the resource i. To evaluate our approach accuracy, the resource with the highest predicted value is compared to the resource chosen by the learner. At the end, an agreement percentage is then calculated between our prediction and the actual learners’ preferences.
6
Conclusion and Further Work
This paper presents an extension of the learner module within ITSs in order to dynamically adapt pedagogical resources to the learner’s cultural profile. CAMPERE makes it unnecessarily by using learner’s cultural background in an exhaustive way. The adaptation technique based on collaborative filtering allows inferring cultural preferences which the system is unable to retrieve explicitly from the learner or the expert. The implicit cultural profile is automatically updated using the learner’s
44
interactions with the system throughout learning sessions. The update process is transparent for the learner and don’t need to define learner’s cultural background in an exhaustive way. Given that our approach uses the collaborative filtering technique in the adaptation process, it will make a more relevant resources recommendation once it gets enough learners’ information and interactions. To address this limitation, an expert tagged resources to minimize the cold start effect on the system. Another drawback consists of too many resources to change, which could be tedious for the learner since at first the system will suggest resources likely to be changed by the learner. Furthermore, the knowledge used by the system could have been modeled in an ontology to better organise our concepts, resources and the expert knowledge. This point will be addressed in our future work. Indeed we will associate our recommendation system with structured knowledge, making it more usable and more interoperable.
References 1. Adomavicius, G., Kwon, Y.: New Recommendation Techniques for Multicriteria Rating Systems. IEEE Intelligent Systems, 22(3), 48-55, (2007) 2. Adomavicius, G., Tuzhilin, A.: Toward the next generation of recommender systems: a survey of the state-of-the-art and possible extensions. IEEE Transactions on Knowledge and Data Engineering, 17(6), pp. 734-749, (2005) 3. Blanchard, E., Lajoie, P.: Learner-Concerned AIED Systems: Affective Implications when Promoting Cultural Awareness. In: 2nd International Workshop on Cultural-Aware Tutoring System (CATS), pp. 13-23, Brighton (2009) 4. Blanchard, E., Mizoguchi, R., Lajoie, S.P.: Addressing the Interplay of Culture and Affect in HCI: An Ontological Approach, 13th International Conference on Human-Computer Interaction: HCII2009, San Diego:Ca, USA (2009) 5. Blanchard, E., Razaki, R., Frasson, C.: Cross-Cultural Adaptation of Elearning Contents : a Methodology. International Conference on E-learning, Vancouver, Canada (2005) 6. Burke, R.: Hybrid Recommender Systems: Survey and Experiments. User Modeling and User-Adapted Interaction, 12(4), pp. 331-370, (2002) 7. Cassady, J. C., Mohammed A., Mathieu, L.: Cross-cultural differences in test perceptions: women in Kuwait and the United States. Journal of cross-cultural psychology, vol. 35(6), pp. 713-718, (2004) 8. Dunn, P., Marinetti, A.: Cultural Adaptation: Necessity for eLearning. Learning in the New Economy e-Magazine (LiNE Zine), (2004). http://www.linezine.com/7.2/articles/pdamca.htm 9. Hofstede, G.: Culture’s Consequences: Comparing Values, Behaviors and Organizations across Nations (2003) 10. Jeunesse, C., Manderscheid, J. C.: L'enseignement en ligne: A l'université et dans les formations professionnelles Pourquoi? Comment? (2007) 11. Kashima, Y.: Conceptions of culture and person for psychology. Journal of Cross-cultural Psychology, 31(1), pp. 14-32, (2000) 12. Kim-Prieto, C., Fujita, F., Diener, E.: Culture and structure of emotional experience.Unpublished Manuscript, University of Illinois, Urbana-Champaign (2004)
45
13. Manasi, P., Shane, F., Neil, T.: Extending ITS Authoring Tools to be Culturally Aware. In: 1st International Workshop on Cultural-Aware Tutoring System (CATS), pp. 101-105, Montreal (2008) 14. Reinecke, K., Schenkel, S., Bernstein, A.: Modeling a User's Culture. In The Handbook of Research in Culturally-Aware Information Technology: Perspectives and Models, IGI Global (2010) 15. Reinecke, K., Gerald, R., Bernstein, A.: Cultural User Modeling With CUMO: An Approach to Overcome the Personalization Bootstrapping Problem (2007) 16. Savard, I., Bourdeau, J., Paquette, G.: Cultural Variables in the Building of Pedagogical Scenarios: the Need for Tools to Help Instructional Designers. In: 1st International Workshop on Cultural-Aware Tutoring System (CATS), pp. 83-92, Montreal (2008) 17. Scherer, K. R., Brosch, T.: Culture-specific appraisal biases contribute to emotion dispositions. European Journal of Personality, 23, 265-288, (2009) 18. Scollon, C.N., Diener, E., Oishi, S., Biswas-Diener, R.: Emotions across cultures and methods. Journal of cross-cultural psychology, vol. 35(3), pp. 304-326, (2004) 19. Urry, J.: Mobility and Proximity. Sociology, 36, Number 2, pp. 255-275, (2002)
46
Interactive PhrasebookTM – Conveying Culture Through Etiquette Peggy Wu1 and Christopher Miller1 1
Smart Information Flow Technologies, 211 N 1st St. Suite 300, Minneapolis, MN USA 55401 {PWu, Cmiller} @ Siftech.com
Abstract. Computerized training systems may be the only viable solution to accommodate the quickly evolving learning needs of individuals who work in multi-cultural teams. While proficiency in the foreign language is important, it is only part of the equation in effective communication. For an Interactive Training System to convey the cultural nuances, it must possess a scalable, customizable, and computationally tractable model for the code of conduct, or etiquette of that culture. Empirical evidence shows that etiquette in the form of politeness affects scenario outcome, such as compliance, trust, and affect. A computational model that incorporates culture specific social intelligence can allow virtual actors to act and react to human students in ways that are more aligned with the humans they simulate. The authors have adapted a sociolinguistic model of human-human interaction for use in language and culture training, an application we call Interactive Phrasebook. Keywords: Intelligent Training Systems, Language and Culture Training, Cross-Cultural Interactions, Formal Models, Computational Etiquette, Politeness, Socio-linguistics, Interactive Phrasebook.
1 Introduction In a wide range of training domains, high levels of teacher student interactions have been shown to impact a student’s learning experience [1]. To reduce the cost of high human teacher-to-student ratios, intelligent tutoring systems (ITS) are being developed to replace or supplement human instructors. However, current ITSs lack many of the interpersonal facets of interaction that can directly impact learning outcomes including, perhaps especially, the affective engagement that comes with engagement with a human instructor. Many current ITSs revolves around learning style or pedagogical theory [2], thus focusing on the student’s mastery of various skills in order to estimate which problems will be easy or difficult for the student. However, apart from a focus on objective performance against specific problem sets, it is well known that a user’s attitudes and level of engagement have significant effects on learning outcomes. For example, a user’s attitudes towards computers have an important impact on learning, since attitudes influence self-efficacy, a belief in one’s ability to succeed in a given task [3].
47
In the domain of culture and language training, affect not only plays a role in the tutorstudent interaction, but is also an essential part of the lesson content. Language not only conveys semantics, but is also an important channel with which humans detect and express affective states in culturally sensitive ways. Our prior work in examining human-machine collaborations suggest that a theoretically based model of human-human interaction can be used to inform virtual agent behavior so they exhibit some elements of the socially sophistication shown by the humans they are simulating. A computational model can dynamically generate agent behaviors to react to a human learner based on contextual factors that are socially and culturally biased. Further, it may also enable visualizations of how these factors affect decision making and goal progression, making explicit the ethnic influences that normally only occur in our heads. 1.1 Culture and Language Training While immersion may be a successful means of rapidly understanding cultural norms and language, it is costly and impractical for most students. Computerized simulations (such as OneSAF or the Future Immersive Training Environment for military training) may provide simulated environments with ‘indigenous’ virtual actors, creating rehearsal opportunities for the human player. However, they are extremely complex and expensive to build, maintain, and update. Efforts in advancing serious games [4], such as DARPA’s DARWARs effort have resulted in the development of highly regarded games with dramatic effects—but corresponding research [5] shows that games are effective only if they are precisely tuned for their environment and for the material to be taught. We believe that it is possible to create a small scale, low cost, and highly scalable training system by focusing on the critical behaviors that are significant for the intended training objectives. Since detection and expression of facial expressions are believed to be innate [6] and therefore cross cultural, we believe language use and interpretation to be the critical behaviors in culture and language training. Until recently, much of the advancements in serious games have been in advancing the rendering of realistic looking avatars, but physical realism is not the most influential, or even necessary in efficient language and culture training. Reeves and Nass [7] has found that people readily anthropomorphize media including machines. Our own studies [8] have found test subjects to react emotionally (i.e. trust, affect, perceived competence) to text based virtual agents that contain no physical embodiments. Effective computer based culture training requires the capability to model and generate socially believable interactions. That is, considering the situation and beliefs of each actor, the actions and reactions of the virtual agent falls within a set of behaviors that may or may not be expected, but a human might exhibit.
2 Background Since 2001, the authors have focused on a family of sociology theories applied to human-computer interactions, examining the effects of politeness on scenario outcome as
48
well as effects of contextual considerations such as the power difference and social distance between the speaker and hearer. It would be trivial to craft agent behaviors that are consistently rude or overly polite, but most communication strategies occur in the dynamic believable middle ground, e.g. being assertive without being rude or being polite without appearing subservient. This is the danger zone where many miscommunications and cultural misunderstandings occur. We leveraged Brown and Levinson’s politeness model [9] based on a seminal body of work in universal, cross-cultural human-human politeness, to create a computational model that can allows virtual agents to utilize believable politeness strategies [10]. In validating the model [11, 12], we found that subjects’ perceptions of politeness were aligned with the design, even when utterances came from machines with no physical embodiment, suggesting that 1) subjects anthropomorphize machines, even if interactions are limited to voice or text messages and 2) a computational model can be used to manipulate and predict perceived politeness. Follow-on experiments showed that politeness can significantly impact subjective and objective performance metrics [13]. For example, increased politeness directly increased perceived familiarity and affect, decreased perceived workload, and affected compliance in complex ways that interact with the power and social distance between actors. We have also applied this work supporting the DARWARS effort in language and cultural training in the military domain [10]. A product of DARWARS, the Tactical Language Training System (TLTS) [14], was developed at the University of Southern California (USC) and is now marketed by Alelo Inc. The authors applied a computational model of etiquette to provide a universal social framework to enhance scenario realism, while simultaneously advancing the underlying system capability to swap cultural modules, resulting in engineering cost savings. A “socially intelligent” character would take offense believably if not addressed in a culturally appropriate fashion or may appear recalcitrant or ignorant when it is merely trying to follow its culturally-derived notions of polite turn taking in discourse. Human interactions involve highly subjective variables that are not clearly defined and are difficult to quantify. Interactions also occur in complex contexts, with the variables in each context containing many possible interpretations. Currently, agent behaviors are frequently emulated in hand-written scripts and simple, locally-relevant rules. Such an approach is labor intensive, extremely brittle, and incapable of conveying the nuances that make it difficult to learn about cultural differences in the first place. While etiquette is not sufficiently rich or precise to explain all aspects of behavior, we believe it is particularly well suited for culture and language learning. 2.1 A Computational Model of Etiquette We use the term etiquette to refer to the set of expectations about observable behaviors that allow interpretations to be made in a cultural context. Politeness is one aspect of observable behaviors that can be used as a measure of etiquette. While the expression and interpretation of politeness is culturally sensitive, its concept is universal.
49
Brown and Levinson [9] produced a seminal body of work in cross-cultural linguistics. In their study of three unrelated languages (English, Tamil, and Tzeltal), they found that people regularly deviate from Grice’s maxims of efficient language [15]. This is, rather than being completely truthful (Maxim of Quality), concise (Maxim of Quantity), relevant (Maxim of Relation), and clear (Maxim of Manner), people across different cultures habitually insert politeness language in their utterances that are not necessary for efficient and cooperative communication. Brown and Levinson hypothesize that this seemingly unnecessary politeness behavior is in fact required to because any communication exchange implies that the speaker is causing some burden, or face threat, upon the hearer. Goffman [16] proposes that each individual is motivated by two components of face; positive face is roughly the need to be seen as a valuable member of one’s society, and negative face is one’s need for autonomy. By simply speaking, the speaker has demanded the hearer’s attention, regardless of the message’s content. Therefore, the speaker has, at the very least, threatened the hearer’s negative face. The fact that the speaker’s message is, for example, a request for fifty dollars, further adds to the threat. This total face threat is dependent on situational variables, such as the power difference between the actors, their social distance, and the inherent imposition of the request the speaker is making of the hearer. The same request is more threatening if asked by a superior rather than a peer, or a stranger rather than a friend. Brown and Levinson call these interactions Face Threatening Acts (FTAs). To mitigate the threat the speaker imposes, politeness strategies are used. Over years of cross linguistic and cross cultural studies, Brown and Levinson created an extensive catalogue of politeness strategies used by different cultures, ranging from “being apologetic”, to “using honorifics”. Further, they hypothesize that in nominal cases, the face threat of the interaction should be balanced with an equal amount of politeness Brown and Levinson themselves do not operationalize these parameters; instead, they are offered as qualitative constructs. Others (e.g. 17, 18) have created numerical representations for them. Since 2001 and under various sources of funding, SIFT personnel have modified Brown and Levinson’s theories to apply to human-computer and computer mediated human-human interactions. The result is a content coding scheme, a software architecture, and set of algorithms which we call the Etiquette EngineTM. The model takes into account influences at the individual level to calculate how an action may be interpreted in terms of politeness, and therefore whether it follows the social code of conduct. We also modified the original calculation of face threat to include an egocentric view, allowing individual virtual actors to hold its own perceptions of the weight of power difference, social distance, imposition, as well as the perceived level of politeness for different strategies. This allows the Etiquette Engine to model differences within each virtual agent that may be a result of cultural or individual biases. The model abstracts cultural factors to preserve its universal applicability, but allows the incorporation of Hofstede’s cultural dimensions [19] to influence behavior selection and interpretation through cultural content. For example, when creating the content for a high as opposed low power dimension culture (i.e. a culture that places more value on power differences, such as Japan, as opposed to a relatively egalitarian culture, such as the United States), the scoring of perceived politeness value for an utterance spoken by a superior may be scored lower than the same in the high power dimension culture. This is because in a high power dimension culture, power
50
differences between individuals are perceived as larger than those in a low power dimension culture. A superior-subordinate pair has a larger power difference in Japan, for example, than in the U.S. A Japanese superior speaking to a Japanese subordinate is thus expected to be more polite than her American counterpart in order to mitigate the larger perceived power difference. By scoring the perceived politeness of a strategy lower for Japanese actors, the Etiquette Engine algorithms dictate that the Japanese superior must use more politeness language in order to be viewed as at par with the American superior. The same method of reasoning can be extended to other cultural dimensions such as individualism/collectivism and masculinity/femininity in order to represent multiple cultures. Currently, the Etiquette Engine does not provide formal methods to automatically convert the scoring of content from one culture to another, and rely on the scoring of cultural subject matter expert, but the underlying mechanisms are in place for future expansion. Below we describe one implementation of the Etiquette Engine in a language training tool we call Interactive Phrasebook.
3 Interactive Phrasebook Interactive Phrasebook is a tool for rehearsing conversations with virtual agents that represent indigenous actors. It is implemented using C# in the Microsoft .NET framework to operate on Microsoft PocketPC enabled hand held PDA devices and has been tested on an HP iPAQ. Interactive Phrasebook focuses on verbal behaviors and gestures displayed in text (with references available in audio and video clips). Similar to a paper phrasebook, a student first selects from a set of predetermined scenarios (e.g. requesting information or check point interactions). The user then customizes the scenario by editing information about the virtual agents within the scenario, including ethnic culture, relative power and social distance, and occupation which has implications on Figure 1 Interactive Phrasebook power (see Figure 1). These Customizing virtual agents parameters affect each virtual actor’s interpretation of behaviors, and ultimately shape their exhibited behaviors. Currently, the end user has a choice of interacting with one of four virtual characters (three Middle Eastern and one
51
American) within the scope of one test scenario containing 37 possible spoken utterances in English and Modern Standard Arabic, and six possible gestures. To interact with agents, the user proceeds to rehearsal mode (see Figure 2). Based on the progression of the scenario, a selection of phrases and gestures are available. The phrases may be displayed in the foreign language or its English translations, with accompanying audio of utterances and video of gestures available for the user’s reference. When the user selects one of the available phrases, the etiquette engine calculates the appropriateness of that action, i.e., based on the relationship (power and social distance etc.) between the virtual agent and the speaker, the action falls under polite, nominal, or rude. A nominal (most appropriate) action would result in the pale green bar at the bottom of the screen extending to either the left (slightly rude) or the right (slightly polite). Figure 2 shows a moderately, but not overly rude user action, as indicated by the dark green bar to the left. The appropriateness is calculated based on the difference between expected and exhibited behavior. Politeness and rudeness is relative to expectations, which in turn are dependent on the situational contexts. If, for example, the user is in a heated argument, the virtual agent may expect the human to use a “rude” phrase. If the human is overly polite instead, the large difference between expected and exhibited behavior would be shown as inappropriately polite, illustrated as a long red bar extended to the right. This visualization provides instant feedback to the user, akin to reading a facial expression of a smile or frown, but presented in an unambiguous way. The user can immediately refer to
52
Figure 2 Interactive Phrasebook Rehearsal Mode
Figure 3 Interactive Phrasebook Review Mode
the rating to “see” what the virtual character was “thinking” in terms of its reaction to the user’s utterances. The interaction is akin to instant messaging, except that the user selects from a set of pre-existing phrases as opposed to entering free text. This interaction method allows users of various levels of fluency in a language to interact with virtual character, reinforcing both language and culture learning. During or after the scenario, the user may enter the review mode to replay the scenario thread (see Figure 3). The visualization in the review mode is an expansion of that of the rehearsal mode, displaying the expected amount of politeness, the exhibited amount, and the difference. Further, the user may change the observer field, allowing the scenario to be replayed as observed by a different virtual agent. This allows the user the advantage to explore how the same utterance may be interpreted differently by individuals of varying ethnicity, power distance, or social distance. By explicitly showing the student when and how social expectations are violated, Interactive Phrasebook can provide insight into the behaviors and underlying motivations of the virtual agents, and therefore the specific actors that are representative of those the student will encounter.
4 Conclusions The notion of selective-fidelity for simulations [20] places focus on aspects of the real situation that make a functional difference from the human user’s perspective. Many cultural misunderstandings occur because of nuanced differences in the interpretation of actions. While politeness of an utterance by itself does not convey many aspects of social protocols, it can explain how behaviors are interpreted at a basic level that most people can relate to. Politeness is used to convey many nuances that are inherent in cross-cultural interactions, and has been shown to significantly influence interaction outcomes (e.g. trust, affect, compliance)[8]. A computational approach with strong theoretic underpinnings can equip the training mechanism with culturally-biased models. Further by abstracting cultural content biases from the underlying computational model, the system allows different cultural information to be swapped in to minimize engineering and maintenance costs. For anyone from soldiers to salespeople, it is important to know what the most culturally appropriate actions are, but it is just as important to anticipate reactions and avoid unintentional misconduct. Training tools like Interaction Phrasebook not only allows the learner to practice interactions, but also provides a structure for the human user to understand how their actions may be viewed as violations of social expectations, the severity of the violation, and how they may singularly or cumulatively alter the scenario outcome.
5 Acknowledgements The authors would like to acknowledge the support for Interactive Phrasebook by the U.S. Army Research Laboratories under contract # W911QX-07-C-0039.
53
References 1 Graesser, A. C., Person, N. K., and Magliano, J. P.,: Collaborative dialogue patterns in naturalistic one-to-one tutoring. Applied Cognitive Psychology, 9, 1-28. Accessed: http://www.cs.pitt.edu/~chopin/references/tig/graesser_person.pdf.pdf (1995) 2 Ferguson, R., Last, M., Liu, C. and Mertz, J.: Service-Learning Projects: Opportunities and Challenges. Proceedings of the 37th SIGCSE technical symposium on Computer science education, Houston, TX. New York, NY. ACM Press. 127-128. (2006) 3 Torkzadeh, R., Pflughoeft, K., and Hall, L.: Computer self-efficacy, training effectiveness and user attitudes: An empirical study. Behaviour & Information Technology. 18:4, 299-309. (1999) 4 Michael, D. & Chen, S.: Serious Games: Games that educate, train & inform. Florence, KY:Course Technology. (2005) 5 Greitzer, F., Kuchar, O., & Huston, K.: Cognitive science implications for enhancing training effectiveness in a serious gaming context. J. Educational Resources in Computing., Vol. 7, No. 3, Article 2. (2007) 6 Matsumoto, D., & Willingham, B.: Spontaneous facial expressions of emotion of blind individuals. Journal of Personality and Social Psychology, 96(1), 1-10 (2009) 7 Reeves, B. and Nass, C. . The Media Equation: how people treat computers, televisions, and new media like real people and places. University Press, Stanford, California. (1996) 8 Wu, P.: "Computational Models of Etiquette: Implementations and Challenges." In Hayes, Caroline and Miller, C., Human Computer Etiquette. New Jersey: Auerbach. (Forthcoming) 9 Brown, P. & Levinson, S.. Politeness: Some Universals in Language Usage. Cambridge,UK.; Cambridge Univ. Press. (1987) 10 Miller, C., Chapman, M., Wu, P. and Johnson, L.: The “Etiquette Quotient”: An Approach to Believable Social Interaction Behaviors.In Proceedings of the 14th Conference on Behavior Representation in Modeling and Simulation (BRIMS). 16-19 May 2005; Universal City, CA. (2005) 11 Miller, C., Wu, P. and Funk, H.:. A Computational Approach to Etiquette and Politeness: Validation Experiments. In Proceedings of the First International Conference on Computational Cultural Dynamics, August 27-28; College Park, MD. (2007) 12 Miller, C., Wu, P., Funk, H., Wilson, P., Johnson, W. L. A Computational Approach to Etiquette and Politeness: Initial Test Cases. In Proceedings of 2006 BRIMS Conference, May 15-18, Baltimore, MD (2006) 13 Wu, Peggy, Ott, Tammy, and Miller, Chris:. Evaluating the Effects of Culture and Etiquette on Human-Computer Interaction and Human Performance. Proceedings of the AAAI Spring Symposium, 23-25 March 2009. Stanford, Palo Alto, CA. (2009) 14 Johnson, W.L., Marsella, S., Viljalmsson, H.: The DARWARS Tactical Language Training System. Procceedings of the 2004 Interservice/Industry Training, Simulation, and Education Conference (I/ITSEC). Accessed: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.117.601&rep=rep1&type=pdf (2004) 15 Paul Grice, Studies in the Way of Words (Cambridge, MA: Harvard University Press, (1991). 16 Goffman E, Interaction Ritual, Pantheon: New York, (1967). 17 Cassell, J. and T. Bickmore, Negotiated Collusion: Modeling Social Language and its Relationship Effects in Intelligent Agents. User Modeling and User-Adapted Interaction. 13(1): 89-132 18 Johnson, W.L., Marsella, S., Mote, N., Viljhalmsson, H, Narayanan, S., Choi, S.: Tactical language training system: Supporting the rapid acquisition of foreign language and cultural skills. In Proceedings of InSTIL/ICALL Symposium, Venice, Italy. (2004)
54
19 Geert Hofstede: Culture's Consequences: International Differences in Work-Related Values. 475 pages. Beverly Hills CA: Sage Publications, (1980). 20 Schricker, Bradley C., Schricker, Stephen A., and Franceschini, Robert W.: Considerations for Selective-Fidelity Simulation. In Proceedings of SPIE – The International Society for Optical Engineering Volume 4367, p 62-70, Enabling Technology for Simulation Science V, Alex F. Sisti; Dawn A. Trevisani, Eds. (2001).
55
