Towards explaining effective tutorial dialogues
Towards explaining effective tutorial dialogues
复制标题
解释有效的教程对话
DOI：
--
复制
发表时间：
2009
期刊：
影响因子：
--
通讯作者：
S. Ohlsson
中科院分区：
文献类型：
--
作者：
David G. Cosejo;Barbara Maria Di Eugenio;Davide Fossati;S. Ohlsson
为了解释有效的辅导对话，美国伊利诺伊大学芝加哥分校计算机科学系Barbara Di Eugenio和Davide Fossati{bdieugen，dfossa1}@uic.edu，美国芝加哥分校Stellan Ohlsson和伊利诺伊大学心理学系David Cosejo{Stellan，dCosej1}@uic.edu。摘要我们提出了一项在核心计算机科学领域对人类辅导对话的研究，该研究集中于双辅导课程，而不是对比不同类型的导师；使用多元回归分析将这些课程的特征与学习输出相关联；并强调了两种迄今尚未被深入研究的辅导举措的效果，即直接指导和正反馈。关键词：人工智能；教育；课程设置；交互行为；人体实验；智能教学系统。与其他教育环境相比，一对一辅导已被证明是一种非常有效的教学形式。二十多年来，研究者一直致力于揭示在学生中产生学习的辅导互动的特征(Fox，1993；Graesser，Person，&Magliano，1995；Lepper，Drake，&O‘Donnell-Johnson，1997；chi，2001；Moore，Porayska-Pomsta，Varges，&Zinn，2004；Evens&Michael，2006；Litman et al.，2006；Di Eugenio，Fossati，Haller，Yu，&Glass，2008)。这项研究的部分动机是寻找有效辅导策略的计算模型1，这是智能辅导系统(ITSS)对话界面的重要组成部分。从如此广泛的工作中，已经出现了许多Find-Ing，但还没有出现统一的解释，部分原因是教程交互如此丰富，以至于很难完全描述它们。我们以及其他人过去一直在追求的一条路线是试图理解专家导师所做的事情(Lepper等人，1997；Evens&Michael，2006；Di Eugenio，Kershaw，Lu，Corrigan-Halpern&Ohlsson，2006；Cade，Copeland，Person，&D‘Mello，2008)。然而，目前还不清楚专家导师应该从什么开始；专业知识并不一定意味着“拥有丰富的经验”。另一种做法是用我们从教育学中了解到的知识来灌输专业知识的概念，并将专家导师等同于理想化的苏格拉底人类导师。然而，正如Graesser和他的合作者在一系列论文(Per-Son，Graesser，Magliano和Kreuz，1994；Graesser等人，1995)中指出的那样，理想化的苏格拉底人类导师很难我们指的是来自计算机科学的算法意义上的计算，而不是来自认知心理学的心理计算。过来吧。在我们自己的教程对话收集和分析过程中(Di Eugenio等人，2006,2008)，我们开始相信，即使是许多有经验的教师，在大多数情况下也不会表现得很正常。尽管如此，学生们仍然会与他们一起学习，有时但不一定比他们从缺乏经验的导师那里学到的东西更多。正如我们在Ohlsson等人中所讨论的那样。(2007)，这使我们倡导一种不同的方法来进行教程对话分析。我们避免先验地定义谁是专家家教，或者将一种类型的家教与另一种类型的家教进行对比，并分析它们之间的差异。相反，我们将所有辅导教师的数据汇集在一起，并通过运行回归分析，将这些辅导课程的特点与学习结果相关联，专注于个别辅导课程的有效性。下一点是关于辅导课程的哪些特征应该进入回归分析。我们上面描述的工作主体提出了各种编码方案来注释教程对话。在Ohlsson等人中。(2007)，我们注意到，大多数情况下，这种编码方案的动机是教学原则、言语行为的语言学理论，以及辅导对话本身的语用。许多计划缺少的是通过对人们如何学习的认知洞察力来获得信息。在这篇文章中，我们提出了一种语料库分析，它体现了我们早先的建议：它将一项研究中收集的辅导课程汇集在一起，这项研究最初是对经验更丰富和更缺乏经验的教师的对比；因为它的动机是什么在认知上是可信的，所以它专注于迄今尚未深入研究的两个特征：直接教学和正反馈；并使用回归分析。特别是，我们证明了正反馈与学习相关，并且我们的方法允许我们以受控的、打印的方式进一步注释数据。我们还非常简要地讨论了我们已经建立的智能交通系统，部分基于这些结果。计算机科学中的人类辅导我们感兴趣的领域是计算机科学(CS)，特别是引入性数据结构，如链表、堆栈和二叉搜索树，以及操作它们的算法。这个领域的部分动机是我们中的一些人的利益，作为CS的教育工作者。基本信息
Towards explaining effective tutorial dialogues Barbara Di Eugenio and Davide Fossati {bdieugen, dfossa1}@uic.edu Department of Computer Science, University of Illinois at Chicago, USA Stellan Ohlsson and David Cosejo {stellan, dcosej1}@uic.edu Department of Psychology, University of Illinois at Chicago, USA Abstract We present a study of human tutorial dialogues in a core Computer Science domain that: focuses on indi- vidual tutoring sessions, rather than on contrasting dif- ferent types of tutors; uses multiple regression analysis to correlate features of those sessions with learning out- comes; and highlights the effects of two types of tutor moves that have not been studied in depth so far, direct instruction and positive feedback. Keywords: Artificial Intelligence; Education; Dis- course; Interactive Behavior; Human Experimentation; Intelligent Tutoring Systems. Introduction One-on-one tutoring has been shown to be a very ef- fective form of instruction compared to other educa- tional settings. For more than twenty years, researchers have worked on uncovering features of tutorial inter- action that engender learning in students (Fox, 1993; Graesser, Person, & Magliano, 1995; Lepper, Drake, & O’Donnell-Johnson, 1997; Chi, 2001; Moore, Porayska- Pomsta, Varges, & Zinn, 2004; Evens & Michael, 2006; Litman et al., 2006; Di Eugenio, Fossati, Haller, Yu, & Glass, 2008). This research is partly motivated by the search for computational models 1 of effective tutoring strategies, which are an essential component of dialogue interfaces to Intelligent Tutoring Systems, or ITSs. From such an extensive body of work, many find- ings have arisen, but a unifying explanation has yet to emerge, partly because tutorial interactions are so rich that it is hard to characterize them fully. One line of at- tack, pursued by us as well as by others in the past, has tried to understand what expert tutors do (Lepper et al., 1997; Evens & Michael, 2006; Di Eugenio, Kershaw, Lu, Corrigan-Halpern, & Ohlsson, 2006; Cade, Copeland, Person, & D’Mello, 2008). However, it is not clear what expert tutors are to start with; expertise does not nec- essarily mean, ’with extensive experience’. Another ap- proach is to imbue the notion of expertise with what we know from pedagogy, and equate expert tutors with idealized, Socratic human tutors. However, as Graesser and collaborators pointed out in a series of papers (Per- son, Graesser, Magliano, & Kreuz, 1994; Graesser et al., 1995), the idealized Socratic human tutor is difficult to We mean computational in the algorithmic sense coming from computer science, rather than in the sense of mental computation coming from cognitive psychology. come by. In the course of our own tutorial dialogues collection and analysis (Di Eugenio et al., 2006, 2008), we have come to believe that even many experienced tu- tors do not behave socratically most of the times. Still, students learn with them, sometimes but not necessarily more than what they learn with inexperienced tutors. As we discuss in Ohlsson et al. (2007), this brought us to advocate a different approach to tutorial dialogue analysis. We eschew defining who an expert tutor is a priori, or casting one type of tutor against another, and analyzing differences between them. Rather, we pool the data from all the tutors together, and we focus on the effectiveness of the individual tutoring session, by run- ning regression analyses that correlate features of those sessions with learning outcomes. The next point concerns which features of the tutor- ing session should be entered in the regression analysis. The body of work we described above has proposed a variety of coding schemes to annotate tutorial dialogues. In Ohlsson et al. (2007), we noted that most often such coding schemes are motivated by pedagogical tenets, by linguistic theories of speech acts, and often by what hap- pens in the tutoring dialogues themselves. What many schemes are missing is being informed by cognitive in- sights into how people learn. In this paper, we present a corpus analysis which em- bodies those earlier proposals of ours: it pools together tutoring sessions collected in a study that had originally started as a contrast between a more experienced and a less experienced tutor; because it is motivated by what is cognitively plausible for learning, it focuses on two fea- tures that have not been studied in depth so far, direct instruction and positive feedback; and uses regression analysis. In particular, we show that positive feedback correlates with learning, and that our methodology al- lows us to further annotate the data in a controlled, prin- cipled way. We also very briefly discuss the ITS we have built, partly based on these results. Human tutoring in Computer Science Our domain of interest is Computer Science (CS), specif- ically, introductory data structures such as linked lists, stacks and binary search trees, and the algorithms that manipulate them. This domain is partly motivated by the interests of some of us, as educators in CS. Basic