Two Heads May be Better than One: Learning from Computer Agents in Conversational Trialogues
Two Heads May be Better than One: Learning from Computer Agents in Conversational Trialogues
复制标题
两个头脑可能比一个更好：在对话三部曲中向计算机代理学习
作者：
A. Graesser;Carol M. Forsyth;B. Lehman
背景教学代理是计算机化的说话头或具体化的动画化身，其通过执行动作和与学生用自然语言进行对话来帮助学生学习。在AutoTutor和其他具有自然语言对话的智能辅导系统中，对话发生在辅导员代理和学生之间。智能体能够适应学生的行为、语言贡献，在某些系统中，还能适应学生的情绪（如无聊、困惑和沮丧）。研究重点本文探讨了几种设计的trialogues（两个代理人与人类学生互动），已有效地实现了特定的学生，主题和深度的学习。这两个代理人承担不同的角色，但通常作为同行和导师。针对不同班级的学生，有不同的教学目标。例如，学生可以（a）间接观察两个代理的交互，（B）与导师代理匡威，而同行代理定期插话，或（c）教同行代理，而导师拯救有问题的交互。此外，智能体可以就问题相互争论，并询问人类学生对争论的看法。研究设计Trialogues已经开发了系统的实验调查，在几项研究中，测量学生的印象，学习收益从前测到后测的客观测试，以及在学习过程中的认知和情感状态。这些研究比较了不同教学原则的条件，以评估这些原则对学生印象，学习，情绪和其他心理措施的影响。对日志文件中的语言和行为进行话语分析，以评估其对心理测量的影响。这些基于代理的系统的测试表明，在学习收益和学生情绪的系统影响的改善。未来，研究者需要进行更多的研究，以实证评估不同的三重设计对心理测量的心理影响。这些三重设计的范围从学生观察到的代理之间的脚本交互，到学生帮助同伴代理，到学生解决两个代理之间的争论。中心问题是学习经验和结果是否显示出比典型的人机对话（即，一个人和一个导师代理）和常规的教学干预。
Background Pedagogical agents are computerized talking heads or embodied animated avatars that help students learn by performing actions and holding conversations with the students in natural language. Dialogues occur between a tutor agent and the student in the case of AutoTutor and other intelligent tutoring systems with natural-language conversation. The agents are adaptive to the students’ actions, verbal contributions, and, in some systems, their emotions (such as boredom, confusion, and frustration). Focus of Study This paper explores several designs of trialogues (two agents interacting with a human student) that have been productively implemented for particular students, subject matters, and depths of learning. The two agents take on different roles, but often serve as peers and tutors. There are different trialogue designs that address different pedagogical goals for different classes of students. For example, students can (a) observe vicariously two agents interacting, (b) converse with a tutor agent while a peer agent periodically chimes in, or (c) teach a peer agent while a tutor rescues a problematic interaction. In addition, agents can argue with each other over issues and ask what the human student thinks about the argument. Research Design Trialogues have been developed for systematic experimental investigations in several studies that measure student impressions, learning gains from pretest to post-test on objective tests, and both cognitive and affective states during learning. The studies compare conditions with different pedagogical principles underlying the trialogues in order to assess the impact of these principles on student impressions, learning, emotions, and other psychological measures. Discourse analyses are performed on the language and actions in the log files in order to assess their impacts on psychological measures. Recommendations Tests of these agent-based systems have shown improvements in learning gains and systematic influences on student emotions. In the future, researchers need to conduct more research to empirically evaluate the psychological impact of different trialogue designs on psychological measures. These trialogue designs range from scripted interactions between agents being observed by the student, to the student helping a fellow peer agent, to the student resolving an argument between two agents. The central question is whether the learning experiences and outcomes show improvement over typical human-computer dialogues (i.e., one human and one tutor agent) and conventional pedagogical interventions.