Design Principles for Gamified Pedagogical Conversational Agents

Pedagogical conversational agents (PCAs) such as chatbots and voice assistants can support learners in their studies. However, interactions with PCAs are often perceived as less motivating. Gamifying PCAs has been proposed as one approach to counteract this issue and increase learners' engagement. However, there is currently little prescriptive knowledge on how to design gamified PCAs. To address this, we conducted interviews with learners and reviewed relevant literature to derive 18 meta-requirements, five design principles, and 20 design features for gamified PCAs. We then applied our design knowledge to create a conceptual prototype, which we validated through an experiment with 76 participants. The experiment results demonstrate that our design knowledge can positively influence motivation and enjoyment in learning with PCAs.


INTRODUCTION
Learners have difficulties in motivating themselves to learn [73,84,114].Digital technologies can be used by higher educational institutions to motivate learners and promote their learning success.One of these technologies are pedagogical conversational agents (PCAs), i.e., intelligent dialog systems that aim to support learners by interacting with them using natural language options such as text or voice [45].This allows them to guide learners individually [117].They can, for instance, teach learners content through dialog, help them manage their time, motivate and mentor them, or encourage collaborative group work [54,117].However, many PCAs are inadequately designed and thus rejected by their users [34,48].In particular, interactions with PCAs are often not very stimulating, so that they cannot fulfill their potential of motivational support [12,98].This leads to learners discontinuing the PCA use [ibid.].Gamifying PCAs has been proposed as one way to counteract this issue [12].Gamification has been showing positive learning effects for years, such as promoting engagement [97] or knowledge gain through quizzes and challenges [92].Studies also determined that gamification is a critical success factor for long-term PCA use [78,102].However, researchers emphasize that in order to achieve positive effects through gamification, it is vital to design gamified systems adequately [97].If gamified learning applications are not adequately designed, this can lead to negative consequences such as competitive pressure or frustration [42,107].Designing gamification appropriately is a complex and difficult process, esecially due to the large number of possible design options [91,97].Accordingly, providing designers with recommendations for the creation of gamified PCAs is crucial.Researchers already contributed frameworks for gamified systems in general [e.g., 21,85,86,97], but frameworks that apply gamification in conversational systems such as PCAs are rare [53].Since the effects of gamification can vary in different contexts and application systems [42], it is useful to establish prescriptive recommendations for gamified PCAs.Benner et al. [12], for instance, provide a first approach to establishing design knowledge by contributing a conceptual framework for gamifying PCAs.However, this framework still lacks the formulation of design principles (DPs), i.e., prescriptive rules for a proposed design [37].These DPs would be useful to guide researchers and practitioners in designing innovative artifacts [38].We address this research gap with the following research question (RQ): How to design gamified PCAs in order to foster students' motivation to learn?
We focus on motivational empowerment because recent studies show that learners particularly desire technology support in term of motivation [10,84] and because gamification can address this issue by boosting learners' motivation [23,65].Researchers have highlighted that higher motivation leads to learners acknowledging the value of the PCA and consequently being more willing to use it regularly [83,95].Consequently, we contribute knowledge to solving the initially described problem of PCA rejection.To answer our RQ, we derive design knowledge for PCAs based on interviews and kernel theories, illustrate it in a low-fidelity prototype, and evaluate it in an experiment.

RESEARCH BACKGROUND 2.1 Gamified Pedagogical Conversational Agents
PCAs interact with humans using natural language, either as chatbots or as voice assistants [45].PCAs are popular in digital education because they are always accessible, can support learners in their learning processes by giving advice, and enable human-like interactions [45,113].In addition, continuous progress in artificial intelligence (AI) enables PCAs to understand their users' concerns better and respond to them [27].Therefore, PCAs are increasingly able to build relationships with their users by acting as virtual companions [78,103].Furthermore, PCAs can take on different roles to support learners depending on their functions and behaviors [12,54,117].For instance, they can act as a tutor by teaching knowledge, as motivators by encouraging learners to study, as mentors by accompanying their learning progress, as organizers by assisting them in everyday study tasks, or as moderators by facilitating group work [54].However, interactions with PCAs are often less stimulating [12].Gamification can counteract learners' low engagement because game elements motivate users during their activities and foster long-term engagement [8,25].We speak of gamified PCAs when PCAs are enriched with individual game elements [12,53].For example, Benner et al. [11,12] developed the PCA "Micromate" to foster learners' motivation with rewards or progress visualizations.Overall, different game elements can be used to achieve motivational effects.Schöbel et al. [12] provide a taxonomy of game elements already applied in the context of PCAs.Accordingly, the following game elements can be used in PCAs for motivational purposes: points, badges, virtual goods, level, ranking, progress bar, feedback/information, avatar, goals, time pressure, narratives, and reminder.We use this taxonomy in this paper to align our proposed design of gamified PCAs with it.

Kernel Theories for the Design of Gamified Pedagogical Conversational Agents
We incorporate scientific "kernel theories" from the disciplines of motivation, learning, and human-computer interaction (HCI) to rigorously design our artifact [38,72].In the following, we will explain the reasons why we include theories from these different disciplines.We include motivational theories because we found in a recent PCA literature review that many of the existing PCAs are insufficiently designed based on motivational theories, which could be a reason why many do not reach their full potential [54].In this paper, we focus on motivation as the key variable of the design science research project and use gamification as an approach to enhance motivation, which is why we include motivational theories that are established in the gamification literature.However, since our literature review revealed that PCAs should be designed in an interdisciplinary manner, we add learning and HCI theories that are already established in the PCA context [ibid.].In that way, we aim to provide a broad theoretical basis for the holistic design of gamified PCAs.Furthermore, we aim to contribute knowledge for the design of PCAs that addresses both self-regulated learning (independently determining the learning process) [123] and social learning (learning in a group, collaboratively and competitively) [7] as recommended by researchers [36].This allows us to show how gamified PCAs can be applied in different learning contexts.In the following paragraphs, we will explain each of the kernel theories.We rely on the self-determination theory (SDT), according to which motivation arises from fulfilling human needs of competence, autonomy, and social relatedness [89].Game elements should, for instance, encourage learners' competence through rewards, provide learners with the opportunity to make autonomous decisions, and promote social relatedness through collaborative learning [89,91].Moreover, we incorporate flow theory, which suggests that being fully immersed in an activity facilitates learning [24].Furthermore, we consider the ARCS model, which defines different motivational factors, such as providing challenges or that learners perceive the content to be relevant [52].In addition, learners should be able to pursue specific goals and receive feedback regarding their goals [62,64].This leads to positive learning outcomes according to goalsetting theory [62,64].In addition, we consider helping learners gain confidence in their ability to manage challenges on their own (self-efficacy) [6].Self-efficacy can be fostered by regularly applying learning content (e.g., in quizzes) [109].Furthermore, we use the social comparison theory, according to which people strive for comparison with others at a similar ability level [33].Consequently, this should be considered when integrating competitive elements to avoid negative consequences such as frustration [61,91,101].Rewards should also be equitable so that, according to equity theory, there is a balance between the performance and the reward received to avoid frustration [1,93].
In terms of applying learning theories, the PCA could provide learners with tasks having different levels of difficulty (e.g., remembering and applying) to address both lower and higher levels of learning goals according to Bloom's taxonomy [4,12,13].In this regard, according to the ICAP framework, it is also crucial that learners not only passively absorb content but that the PCA fosters constructive and interactive learning [19].The PCA can take this into account both by assuming the role of a learning partner as well as by encouraging interaction with fellow students [19,44,112].To support learners, the PCA should provide assistance while maintaining a "zone of proximal development." For instance, the learning tasks should slightly exceed the ability level of the learner [111].In this way, learners can progress and receive step-by-step assistance in solving the challenges through hints by the PCA (so-called "scaffolding") [115,116,119].In addition, repetition is vital to process the content actively, and repetition intervals should become progressively longer following the "spacing effect" [16,87].
Regarding the HCI, the PCA should act somehow human-like to foster users' trust according to the computers are social actors (CASA) theory [75].According to social agency theory, the PCA can take a specific human-like role (cf.Chapter 2.1), so interaction with the PCA can be made more instructive if it is interpreted as social communication [12,66].

METHODOLOGY
We follow the design science research framework by Kuechler and Vaishnavi [60].This framework has several sequential steps to systematically derive design knowledge, starting from the awareness of the problem to the proposal of the design knowledge, its implementation, and the conclusion about the findings (see Figure 1).In this paper, we report the results of our first design cycle.We used this framework as it is process-oriented and thus allowed for a step-by-step research procedure [60], and because design science researchers have repeatedly emphasized that this is an established model for deriving design knowledge [e.g., 9,38,104].
First, we interviewed eight potential PCA users (four male and four female) to get aware of the problem space and gain insights regarding the solution space [15].For this purpose, we used a semistructured interview guide (see Appendix A.1) [105].We mainly asked questions regarding problems and challenges in learning and desired solutions for the gamified PCA.Regarding the PCA design, we included questions about motivating as well as demotivating functions and behaviors of the PCA, the preferred roles of the PCA (cf.Chapter 2.1), and the implementation of the individual game elements according to the taxonomy by Schöbel et al. [97].We analyzed the interviews qualitatively [67].For instance, we created a coding guide for the two superordinate categories.The further sub-categories emerged inductively, and we finalized our coding in a revision cycle [67].Based on the results, we subsequently formed user stories (USs) according to the scheme described by Moser [74] and assigned them to the associated game elements.We complemented the results with the kernel theories (cf.Chapter 2.2) and supporting literature on gamification and education when deriving the DPs [71].Based on the synthesized findings from interviews and literature, we formulated meta-requirements (MRs), DPs, and design features (DFs).According to Möller et al. [71], we visualized them in a mapping diagram.We formulated all DPs using a unified schema, according to Gregor et al. [37], consisting of the goal, implementer, user, context, mechanisms to fulfill the DP and rationale.Since PCAs can take different roles and consequently show different functions [12,54,117], we specified the PCA role in the "context" part of each DP.We illustrated all DPs in a conceptual prototype, "Ben, " to instantiate our design knowledge [38].To evaluate our DPs, we formulated testable propositions (TPs) that corresponded to the objectives of each DP and which can be measured by constructs [39].We evaluated each DP by validating our TPs regarding whether learners perceive our expository instantiation in compliance with the DPs to be better than a baseline artifact that does not consider the DPs [110].In the next design cycle following this paper, we develop Ben as a mature PCA and test it with real users (see Figure 1).

DERIVATION OF DESIGN PRINCIPLES
In the following, we will elaborate on the five DPs.Since the implementer is identical for each DP (PCA designer), we do not repeat this information in the formulation of each DP.The corresponding USs are visible in the Appendix A.2 and the mapping diagram is visible in the Appendix A.3.
The participants expressed that visualizing the learning progress helps to be able to evaluate one's own skills and knowledge (US1).According to goal-setting theory, this feedback positively affects learners' performance [62,64].Consequently, a gamified PCA should visualize learning progress to highlight learners' achievements (MR1).In addition, the learners wished to get rewards when they achieve special outcomes (US2).These rewards should take into account the time and work needed to succeed.For instance, learners should receive special rewards if they reach long-term milestones and fulfill complex tasks (US3) [32].Rewarding learners increases their perception of competence [89,91].Rewards provide learners with feedback on their positive performance, and when they are awarded performance-related, it encourages learners' motivation [90].Hence, the PCA should reward positive performance (MR2) and use different game elements for this purpose, considering the students' work effort (MR3).To implement the MRs, the PCA could consistently award points for correct answers and accomplishments (DF1), distribute virtual goods such as coins for medium-term In addition, the PCA should use levels per topic that also increase based on points earned (DF5).These levels lead users to perform better due to seeing their progress [68].Hence, we derive DP1, according to which the PCA should act as a motivator that rewards students in a performance-related manner (Figure 2).
The following DP contains mechanisms of the PCA acting as a tutor at the beginning of the learning process and as a mentor later.It is crucial that the PCA performs a clear role since (AI-based) agents better perform their function when they are in a particular relationship with the user according to social agency theory [12,66].This is because people behave human-like toward these agents and apply similar expectations to them as they have towards humans, according to CASA theory [75,100].In the following, we first elaborate on the mechanisms of the PCA in a tutor role and then on the mechanisms of the PCA in a mentor role.First, users desired to be quizzed at the end of learning units to check their understanding of new learning content (US4).This continuous practice could increase learners' self-efficacy [6,109].Consequently, the PCA should test new content after each lesson so that users can review what they have learned and build confidence regarding their skills (MR4).In addition, learners wished to practice even when time is limited (US5).Therefore, the PCA should offer short quizzes so learners can study in small time slots (MR5).The ability to take quizzes in small time slots can increase learners' motivation, for instance, they have the chance to repeat the lesson on their way to the university [11,120].Students perceived it as important that the PCA acts as a tutor at the beginning of the learning process by teaching content and training lower-level learning goals such as remembering facts [4,13].This role should be communicated clearly (US6).Furthermore, learners need to be able to repeat the knowledge (US7).Therefore, designers should implement the spacing effect, according to which the recall of knowledge in increasingly longer intervals favors its long-term memorization [16,87].Overall, the PCA should repeat learning content using summaries and quizzes, and the PCA should allow for spaced repetition learning so that students can recall the knowledge confidently (MR6).The PCA should adapt its role later by acting as a mentor that builds a friendly relationship with learners [55].In doing so, the PCA should address higher-level learning goals, such as the application and transfer of knowledge [4,13].Hints provided by the PCA can help learners find solutions for complex learning tasks [116].Thus, learners could build multiple levels of learning competencies (MR7).Building competencies on multiple levels is crucial as it allows learners to achieve learning success [4,13] and strengthens their sense of competence, which promotes self-efficacy and motivation [6,55,89,109].To implement the MRs, the PCA should first take on the tutor role to exercise learners at the end of a chapter (DF6).To do this, the PCA should, for example, access the calendar to identify opportune times to study and send reminders (DF7) [55,80].In addition, the PCA should repeat the content and offer summaries.Afterward, the PCA should start a quiz to recall what students have remembered (DF8) [2,11,80].This quiz should be repeated at increasingly longer intervals.Furthermore, the PCA should act as a mentor by providing them with more complex learning tasks (e.g., open questions) [117].In these tasks, the PCA could lead learners to apply knowledge and develop solutions for specific challenges.Within the mentoring role, participants explicitly wished the PCA to act as a classmate from a higher semester.In that way, the PCA can establish a co-equal relationship with their users [55,103] and act as a student that has already passed the exam so that learners perceive its help as credible [56] (DF9).Hence, we derive DP2 regarding knowledge acquisition and application (Figure 3).
Interviewees expressed the desire to align their learning process with self-defined or suggested learning goals (US8).Specific and challenging goals can help improve learning performance [62,64].Furthermore, if learners can influence the learning goals themselves, this can lead to a higher acceptance of the learning goals according to the goal-setting theory [62,64].Own or influenceable learning goals can be perceived as more relevant [62,64], contributing to users feeling autonomous and hence fosters their motivation [52,89].This enables users to learn self-regulated, i.e., to independently determine the learning process [14,94].Consequently, the PCA should allow learners to set their own learning goals so that they can incorporate their interests and perceive the learning goals to be relevant and valuable [94] (MR8).At the same time, the PCA should support the students if they have difficulties setting up their own goals or formulating them according to the SMART concept as specific, measurable, attainable, realistic, and time-related [29] (MR9).The PCA's support of goal-setting is useful because some students lack self-regulation skills, making it difficult for them to set goals on their own [118], negatively affecting their learning success [121].For instance, many students set unrealistic goals, which leads to disappointment for them and they hence perceive learning to be a burden [99].Setting clear and realistic goals makes learners more likely to complete the course successfully and is perceived positively by students [43,95].The learners wished the PCA to ensure regular reflection on students' learning progress [17,43,117] (US9).Such self-reflection leads students to train themselves to plan ahead for their learning and to learn from their mistakes [124].This can enable students to be satisfied with their learning progress and to be confident about achieving their learning goals [59] (MR10).Being satisfied when learning contributes to motivation [52].In addition, students wished to divide goals according to their time horizon (US10).Accordingly, the PCA should highlight and reward the achievement of short-, medium-, and long-term learning goals [59] (MR11).These successes foster that learners perceive themselves to be competent [89].The MRs can be implemented by having the PCA establish learning goals with the user at the beginning of the learning process based on the user's ideas and desires [43] (DF10).If needed, the PCA should make recommendations for learning goals based on the user's past performance to present individualized and realistically proposed goals (DF11).The PCA should furthermore support setting goals for each topic or section (DF12).For each (subordinate) goal, the PCA should moreover present the progress via progress bars and feedback [97] and reflect on it together with the student [117] (DF13).Finally, the PCA should reward the achievement of short-, medium-, and long-term learning goals [91].Learners should receive the highest rewards when achieving long-term goals, so that the reward is performance-based and that they value the long-term goal as well as the rewards that come with it [77] (DF14).Hence, we derive DP3 of goal-setting and reflection (Figure 4).
Furthermore, narratives can be used to apply knowledge in reallife situations [35] (US11).The narrative's content should align with the user's learning goal to be perceived as relevant [3,52].The narrative should involve the learner constructively and interactively, as  according to the ICAP framework, this favors learning success [19].The story should enable the user to demonstrate problem-solving skills [26,35,63].In this way, learners can practice higher-level learning goals to enhance their skills [4,13] (MR12).The role of the PCA in the narrative was also important to the participants (US12).They desired the PCA to initiate and accompany the narrative but not be part of it to maintain its credibility.Thus, according to social agency theory, the PCA can better fulfill its function as a narrator [66].The PCA should support the students in their learning process by giving hints, so the PCA needs to have a higher level of knowledge than the user.Learners expressed that if the PCA had a participatory role, they would perceive this as a violation of the PCA's role.For example, if the PCA pretends to be unaware of tasks, users may find this untrustworthy.Furthermore, we assume that users' self-efficacy could increase by completing the narrative independently [6,82].Therefore, we suggest the PCA to initiate and accompany the narrative without being involved to create a clear role (MR13).Furthermore, participants desired the PCA to divide the narrative into multiple challenges [57] (US13).According to goal-setting theory, these challenges represent clear objectives, so that the user can continuously be aware of his/her progress [62,64].The PCA should adapt these challenges to the capabilities of the learner.This creates a state in which the user is neither under nor over-challenged, leading to a complete absorption in the narrative according to flow theory [24] (MR14).The MRs can be implemented by having the PCA introduce narratives that are life-relevant and whose progression depends on the users' responses [26,77].In doing so, the PCA should act as a narrator that presents users with a problem situation to overcome [50] (DF15).By challenging learners to solve problems on their own, and by encouraging them to achieve success in solving problems, learners are empowered in their self-efficacy, thereby increasing their motivation [6].In addition, the PCA should subdivide the narrative into achievable subgoals (e.g., using quests) adapted to the user's abilities so that the student feels that he/she is making successive progress [59] (DF16).Hence, we derive DP4, according to which the PCA should act as a learning story narrator (Figure 5).
The following DP refers to mechanisms fostering collaboration and competition among peers because both tend to motivate students [31].The students wished that the PCA supports them in finding learning partners (US14).This is consistent with further recent literature which states that many learners desire social interaction with their peers (esp. in digital education) [55,76,84].The PCA should only suggest learning partners that are at the same level of knowledge to enable horizontal comparison following social comparison theory [33,91].This is crucial because competition, if not adequately designed, can lead to frustration among players [91].Horizontal comparison allows learners to realistically assess their own skills and recognize that they can only win the competition by putting in effort [61,91,101].This ensures that the collaboration among learners is based on an equal relationship (MR15).They have to solve specific tasks like quizzes and joint narratives [12,96].The PCA should take a moderator role in collaborative learning to accompany the exchange between the learners [54] (US15).According to the zone of proximal development, the PCA should support the learners to varying degrees [111,116].If the users are still at the beginning of their learning process, the PCA should clearly point out errors to them.If the learners have already reached a higher level, the PCA should only subtly point out errors.For instance, the PCA could ask learners to explain unclear content to each other and give some hints.In this way, the PCA could help learners apply the learning techniques of "elaborative interrogation" and "self-explanation," which educational researchers assume to be very effective [30].These learning techniques are based on solving self-made questions and explaining circumstances to classmates [30].In summary, the PCA as a moderator should ask learners to explain the learning content to each other in their own words, support them when needed, and highlight errors (MR16).This allows learners to help and correct each other.Overall, joint learning might strengthen learners' social relatedness [28,89].Learners also wished to get rewarded by the PCA when learning collaboratively (US16, US17).However, in social contexts, rewards must correspond to the expended effort of an actor; otherwise, this is perceived as unfair according to equity theory [1,93].Thus, the PCA should equally reward learners actively participating in joint learning to recognize the performance of all.In summary, the PCA should present students with collaborative challenges and reward all players equally to recognize their effort and promote social relatedness (MR17).Furthermore, learners wished to compete against each other.In this context, the PCA should recognize all players for their participation but give a higher reward to the winning party for its better performance [1,93] (US17).This helps to adequately acknowledge the learners' individual performance and to promote learners' enjoyment (MR18).To connect students, they could either search for peers by themselves or the PCA could match students with each other [55] (DF17).The PCA should use "scaffolding" to guide students when problems arise and encourage them to explain the content to each other [111,116] (DF18).For collaboration, learners should receive common challenges such as quizzes or narratives to solve them together.Upon completion of the challenge, the PCA should reward students using points, virtual goods, or badges.In competitive learning, the PCA should allow users to compete against each other in a quiz, either alone or in teams [122].All students should be rewarded upon completion, with the winning party receiving a higher reward (DF20).Hence, we derive DP5 regarding the challenging moderation of learning with peers (Figure 6).

INSTANTIATION AND EVALUATION OF DESIGN PRINCIPLES
To illustrate our design knowledge, we created a UX prototype using the tool "Figma" (Figure 7).We created a variant with the compliance of the DP for each of the five DPs, and a baseline artifact.The baseline artifact implements the basic features of gamified PCAs but does not fully satisfy the mechanisms of the DPs.The chatbot has a human-like name, "Ben, " and accompanies students in learning English.We chose the subject of English learning to create a scenario that is understandable to many learners regardless of the specific content.This allowed us to focus on evaluating the design rather than the content.In line with the DPs, the PCA includes both tasks that address low-level learning goals, such as single-choice and multiple-choice quizzes, and tasks that address more complex learning goals, such as open-ended tasks and learning stories [4,13].
The quiz tasks are, for example, grammar questions to test students' knowledge of tenses.Open tasks include the application of tenses, such as creating one's own sentences in English based on the previously learned tense.Learning stories, on the other hand, represent practical tasks.For example, the chatbots contains the learning story "The Roman Colosseum," in which the learner has to solve various quests, such as ordering a taxi to the Colosseum or asking for a discount on student tickets.In addition, the chatbots contains aspects that address self-regulated learning in the sense of DP1-DP4 and social learning in the sense of DP5 (both collaborative and competitive).Figure 7 shows excerpts of the conversation with Ben (for the baseline artifact and for the artifact with DP compliance).In the following, we will explain the differences between both versions.In DP1, the PCA awards the rewards performance-related, i.e., the PCA considers that they relate to the time horizon of the goal (from short-term to long-term goals).In the baseline artifact, the rewards are not aligned with the learning goals and their time horizon.Instead, the PCA only congratulates the learners verbally.Regarding DP2, the PCA acts as a fellow student from a higher semester that supports like a real classmate, for example, with specific hints.In the baseline artifact, the PCA acts as a higher-level tutor and cannot give hints.With respect to DP3, learners can autonomously formulate goals and visualize their progress towards these goals.In the baseline artifact, the PCA instead specifies the goals themselves and learners are only informed of their progress via message but cannot view it visually.For DP4, the PCA introduces the learning story and divides it into different quests without being directly involved in it, while in the baseline artifact, the PCA has a participatory role, and the story is not divided into quests.In DP5, the PCA connects the learners with further students at the same level, moderates the collaborative and competitive learning, • Performance-related rewards (points for answering individual questions, coins as virtual goods for medium-term successes, badges for special successes such as the error-free passing of a quiz question in the first attempt).
• Visualization of learning progress through levels and progress bars (per topic) DP2 • Representation of Ben as a more senior professor who has been teaching the course for several years • Pure teaching of learning content • Representation of Ben as a student from a higher semester • Tutor and mentor functions (in addition to the pure teaching of learning content, e.g.: summaries of the content, correction of wrong answers, application tasks including hints, spaced repetition learning) DP3 • No autonomous and individualized formulation of learning goals (users have to achieve predefined learning goals) • Users can set their own learning goals and receive ideas from Ben on how to formulate them if needed • Support for formulating learning goals with different time horizons (from an overarching vision to sub-goals for individual courses or lessons) DP4 • Role of Ben as a facilitator, that is part of the story and asks the learner for help, but cannot give any hints • No communication of the overall progress and no subdivision of the story into subgoals • Role of Ben as the narrator of the learning story, who introduces it (e.g., "Imagine... ") and can give tips and hints when learners make a mistake and get stuck • Subdivision of the learning story into sub-challenges, which are presented as quests DP5 • No matching of learners on the same level • The PCA does not intentionally encourage collaboration (e.g., learners can vote on individual quiz questions without exchanging ideas) • The PCA awards rewards only for the winning team and coins are individually distributed to single team members based on their individual performance • No hints during the collaborative processing of the tasks • Ben matches learners at the same level with each other to have teams of the same strength competing against each other.
• Ben supports collaboration among participants (e.g., by asking them to explain and discuss mistakes to each other) • Ben rewards all teams and their members for collaboration • Hints for students when they work on tasks collaboratively and the learners can exchange in the team.The competition is designed in a way that the chatbot lets two teams with the similar level compete against each other and presents them with tasks such as open-ended questions and learning stories in which the learners can exchange ideas in the team about the correct answer.DP5 is designed to encourage team members to work together.In the baseline artifact, there is no moderation but a passive accompaniment, and no exchange within the team is possible.Table 1 summarizes the differences between both versions.We evaluated our DPs in an online survey lasting approx.30-35 minutes.For each of the DPs, the participants could watch two video scenarios (with DP compliance and baseline artifact) and were asked to rate them afterward.We randomized the order of the items as well as the PCA variants.We set up testable propositions (TPs) and identified constructs that correspond to the mechanisms and objectives of the DPs, as common in design science research for evaluating design DPs [39,70].In the following, we will elaborate on each of the TPs, indicating the source of the construct at the end of each sentence.According to DP1, we assume that a gamified PCA that rewards students through various game elements and visualizes their progress increases learners' perceptions of competence [47] (TP1).According to DP2, a gamified PCA that supports students in knowledge acquisition and transfer could foster students' perceptions of friendship with the PCA [69,103] as well as their self-efficacy [58] (TP2).According to DP3, a gamified PCA that supports students in setting goals and reflecting on them, could contribute to the perceptions of usefulness [88], relevance/goalorientation [18,52], as well as self-efficacy [58] (TP3).According to DP4, a gamified PCA that initiates narratives and accompanies them during the narrative could increase users' trusting beliefs [81], as well as the perceptions of relevance and goal-orientation regarding the challenges to be solved [18] (TP4).According to DP5, a gamified PCA that promotes collaborative and competitive learning could increase learners' social experience [46] and their enjoyment and satisfaction in learning [18,52] (TP5).We used a 7-point Likert scale for all items (1 = strongly disagree to 7 = strongly agree).76 learners participated in our study (39 female, 35 male, and two diverse).The average age was 26 years.Cronbach's alpha was > 0.9 for all constructs, and the corrected inter-item scale correlation was always > 0.3 [41,79].Confirmatory factor analysis confirmed the suitability of the data set, as all factor loadings were > 0.5 [20], explained variances were > 0.5 [5], and KMO values were > 0.5 [49].We calculated the mean values and standard deviations for each TP.Moreover, we performed a paired Student's t-test to check whether the mean values of both variants differ significantly for the individual constructs ( = 0.05).Participants rated the version in compliance with the DPs very positively.All values are above the scale mean and even > 5.The baseline artifact was rated worse for each construct, with mean values ranging from 2.85 to 4.85.The t-test revealed a significant difference for all constructs, confirming our TPs (p < 0.01).For some TPs, this is a strong effect (Cohen's d > 0.8), and for the remaining ones, this is a medium effect (0.5 < Cohen's d < 0.8) [22].Table 2 shows the study results.

DISCUSSION 6.1 Implications of the Findings
First, we combine the current emerging approach of PCAs as an interactive way of learning [45,54] with the established approach of gamification as one way for motivational boost [25,97].Our approach aims to exploit the benefits of both forms of technologyenhanced learning.The combination of PCAs with gamification is quite new, as literature reviews have shown that few researchers have yet contributed knowledge in this area [12,53].The gamified PCA approach differs from current chatbots such as ChatGPT, which are mostly pretended to answer users' questions and thus impart knowledge [51].Instead, gamified PCAs go further than just imparting knowledge by helping learners to achieve learning goals in an interactive way, from knowledge acquisition to higher learning goals such as knowledge application [4,13], as also called for by other PCA researchers [55,95].This was highlighted positive by the study participants, for instance: "The interactive learning stories are a great idea.The feature to self-formulate answers seems to increase the learning success, since you have to formulate the correct answer on your own and not just pick it out of given possibilities.The open communication with the chatbot Ben does not give the feeling that you are learning with an app, but rather as if you are actually working together with a human teacher." Overall, we aim to increase learners' motivation not only in the short-term, but instead by guiding them to be motivated in the long-term, for instance, by encouraging them to set own goals and to keep track of their progress.This finding is supported by recent studies which highlight that learners desire long-term motivational support through PCAs [11,55,84,84].Study participants, i.e., addressed that the visualization of their learning progress contributes to their long-term motivational support.For example, one participant addressed: "The badges are very handy and help you capture the current learning status, i.e., the "ready for the exam" badge." Second, our design approach incorporates approaches to selfregulated learning as well as collaborative and competitive learning.For example, with the support of the PCA, it is possible to set own learning goals and work on learning stories independently, as well as complete learning tasks in teams and compete against other teams.This variety of learning methods was emphasized by the survey participants in their answers to the open-ended questions.For instance, one student said, "I liked the progress indicators and mechanisms for goal-setting, as well as the learning stories which included tips from Ben.That motivates me to learn and leads me to gain more knowledge from the learning sessions." At the same time, participants addressed, "I really liked the opportunities to learn together with learning partners and interact with them.They are way better designed as compared to the alternative prototype." Third, we show how to design PCAs to be valuable for their users to counteract the practical problem that many PCAs implementations are unsatisfactory and fail [78].By combining PCAs with gamification, we create a motivational incentive to encourage long-term PCA use which helps learners recognize the value of the PCA as outlined by further researchers [12,55,95].So far, we have been able to show the effectiveness of our design approach for the context of English language learning.However, study participants criticized whether this approach is transferable to more complex content, with one respondent saying, "Ben appears to only be viable for simple structured subjects.For more complex subjects like mathematics, physics or similar, its use is probably not that helpful." Especially in the context of English language learning, learners may have been "positively biased" because they are used to learn through mobile and gamified language learning applications such as "Duolingo." Existing literature on gamified PCAs has also shown that language learning is a popular use case for gamified PCAs [53], however, researchers have already revealed positive results on gamified PCAs in other contexts such as computer science [2], information literacy [35], or environmental engineering [40].Previous literature also mentioned this limitation for gamified PCAs [53].For example, learning soft skills could be more difficult when applied in a chatbot.We are therefore working on using gamified PCAs in full-fledged games to benefit from the advantages of an existing game environment.The success of gamified PCAs could also depend on many variables, for example, the extent to which learners already use digital tools within their learning strategy and are open to such tools, or whether learners already have a minimum level of motivation or enthusiasm for their subject.Furthermore, we are currently working on transferring our approach to other application contexts such as business information systems education to validate if it is suitable in other contexts as well.To implement a satisfactory solution for more complex issues, we see potential in leveraging the emerging trend towards large language models, such as the one used by ChatGPT, which allow developers to better evaluate students' responses to open-ended tasks provided by the PCA [51].Hence, we plan to include an interface of our PCA to large language models in subsequent development iterations.
Fourth, our design knowledge provides an overview that can guide designers and developers to implement gamified PCAs.Our DPs and the associated MRs and DFs are prescriptive recommendations that can enable them to transfer these scientific findings to practice.

Limitations
We admit the limitations of our paper.So far, the evaluation is only based on a conceptual prototype and the participants did not interact with the PCA.We chose this approach to first test the effectiveness of our design proposal before technically implementing the prototype.In our second design cycle, we plan to validate the results with a mature prototype and real user interactions.Second, our sample size (n = 76) is quite small.We chose this evaluation procedure to show a first proof-of-concept and because researchers emphasize that a t-test as in our evaluation is already considered meaningful with a sample size of n ≥ 30 [41].In contrast, we plan to conduct an evaluation with the mature prototype and a higher sample size to show a proof-of-value as well.

CONCLUSION AND OUTLOOK
Gamification of PCAs helps to make interactions with them more engaging, enabling to better guide learners.Since there is a lack of design knowledge regarding gamified PCAs, we have derived 18 MRs, five DPs, and 20 DFs.We exemplified this design knowledge in a prototype, and our first experimental results show the efficiency of our DPs.Current trends such as ChatGPT show that learning with AI-based PCAs is becoming more present in digital education.In addition to the technical design of PCAs, researchers consider it essential to also contribute knowledge about how to increase learners' motivation [11,55].The artifacts presented in this paper make a first contribution to research and practice by showing how such a realization could look like.We are currently working on implementing the PCA.For this purpose, we use the framework "Rasa" and have integrated the PCA into the tool "Slack" to allow for collaboration as suggested by DP5.In addition, we created an interface from Rasa to a language checking tool to check the correctness of the sentences entered by the learners in open tasks and to give hints and suggestions for improvement in case of an incorrect answer.We plan to conduct another experiment in the next design cycle, in which learners assess the interaction with the PCA regarding the identified constructs.In this way, we hope to show the effectiveness of our DPs in a real-world learning scenario as well.In addition, we plan to add more subjects to our PCA to validate our design knowledge's transferability to other contexts.

Key questions
Follow-up questions Checklist for further follow-up questions Role of the PCA (10-15 minutes) -What should a PCA absolutely have or be able to do in order to promote your motivation?-What should a PCA avoid doing in order for you not to be demotivated?-Imagine a PCA personally supporting you in your learning: How should the PCA act towards you?-How should the PCA not perform?-If the PCA could take on one of the roles defined in the material, which one would help you and why? -How could the PCA be designed so that it can take on multiple roles?-Identify preferences for the role of the PCA -Do learners prefer to have multiple PCA roles for different application scenarios?-What game dynamics are already implied by the interviewees?
Gamification design for PCAs (20 Minutes) -What difficulties do you have when you want to motivate yourself to learn?In which situations do you have difficulties and why? -How could game elements in PCA be used to motivate you? -How could game elements be used in PCA, to make you feel competent?to evoke a sense of achievement in you? to grab your attention?to keep your attention?to make you feel satisfied?-How should the game element [X according to Schöbel et al. [97]  A.2 User Stories Mapped to Game Elements According to the Taxonomy of Schöbel et al. [97]

Figure 1 :
Figure 1: Design Science Research Procedure

Figure 3 :
Figure 3: DP2 of Knowledge Acquisition and Application

Figure 4 :
Figure 4: DP3 of Goal-Setting and Reflection

Figure 6 :
Figure 6: DP5 of Challenging Moderation of Learning with Peers

Figure 7 :
Figure 7: Excerpts of the Conversation with the PCATable 1: Differences between the Baseline Artifact and the Artifact with Design Principle Compliance work/be designed?Why? -To what extent do the game elements you mentioned fit the role of the PCA you mentioned?And how could they be used in the context of the role?-To what extent could game elements be combined with each other?And how should this combination be designed?-What elements should definitely not be included or would demotivate you? -How might others be motivated by this?And how should the PCA consider this? -What would have to be done for you not to be demotivated by them?-Identify potentially motivating game elements -Ask reasons for users perception toward the game elements -How should the game elements work in the learning context?-Is the combination of gamification elements helpful?-Identify poor game design options -Which game elements fit to which role of the PCA? Social, playful learning (10 Minutes) -How should a PCA learn together with you playfully?-How could a PCA support the playful learning together with your peers?-How could a PCA support the playful learning together with your peers?-How should a PCA learn together with you playfully?-How should the PCA help you to learn together with the PCA playfully?-How should the PCA help you to structure your learning?-How do you use learning techniques and what are your experiences with them?-How might a PCA support you in the application of learning techniques?-What role should the PCA play here?-Which game elements would support and/or motivate you in social learning?

Table 2 :
Study Results Mean Value; SD = Standard Deviation; p = Significance; T = t-Value; d = Effect Size; TP = Testable Proposition