Interactive Communications

Theory Committee Members

The research problem addressed by this cluster is: What properties of dialogue do or do not promote learning? Educational dialogue has mostly been studied in classrooms (Lave & Wenger, 1991; Leinhardt, 1990) and workplaces (Hutchins, 1995; Nunes, Schliemann & Carraher, 1993). In order to investigate more tractable albeit still complex situations, most of the proposed work will focus on dyadic dialogues, namely dialogues between: (a) a human tutor and a human student, (b) two human students, or (c) A computer tutor and a human student. Although there have been many studies of naturally occurring dialogues of these kinds (Fox, 1993; Graesser, Bowers, Hacker, & Person, 1997; MacArthur, Stasz, & Zmuidzinas, 1990), few have also measured the learning that occur and tried to find properties of dialogue that are associated with learning. That is, our proposed studies will combine observation and manipulation of dialogue processes combined with detailed measurements of cognitive and motivational changes.

The PSLC theoretical framework, which was introduced in section 3.3, suggests that the impact of dialogue on robust learning can be understood in terms of sense-making and foundational skills building. On the foundational skills side, dialogue affects the features that are present during learning. Varying the features present at an interactive communication event affects the likelihood that the student will recall the knowledge component later. On the sense-making side, dialogue affects the likelihood that the desired knowledge events will occur. In particular, interaction between a student and another agent affords two kinds of opportunities to students. On the one hand, dialogue gives the student the opportunity to avoid knowledge events by having the agent do it instead, and this should reduce robust learning. On the other hand, the agent can help the student apply knowledge components that the student is missing or weak on and thus increase the set of beneficial knowledge events and robust learning. Whether a particular instructional process helps or hurts learning depends strongly not only on the instruction itself, but on the students' motivations and learning strategies.

The rest of this section sketches both the research plans of the individual projects in the dialogue cluster and how they fit into the PSLC theoretical framework. Each description is preceded by the project leaders' names as an identifier for the project.

• In an activity where students solve a design problem collaboratively, students who are trained to elaborate on each others' ideas learn more than students who are trained to offer constructive criticism. The constructive criticism leads to longer negotiations about what design to try next, so fewer designs are tried, and fewer appropriate knowledge events occur. (Hausmann & Chi)
• When students ask a tutoring system for help, shorter sequences of hints should produce more learning. Longer hint sequences discourage students from asking for help when they need it; they prefer to guess instead. Thus, even though they might guess correctly, they do not apply the appropriate knowledge, and thus learn less, since guessing is unlikely to help achieve feature validity - it is more likely to lead to overly general knowledge (Aleven, McLaren, Roll & Koedinger, study 1).
• A help-seeking tutoring system (Aleven, McLaren, Roll & Koedinger, study 2) and a collaborative learning paradigm (Aleven, McLaren, Roll & Koedinger, study 3) are being developed that explicitly teach students to monitor their familiarity with knowledge components and their success at knowledge events so that they can ask for help when and only when they need it. If a student can recall a knowledge component and apply it without help, even if it is a bit of a struggle to recall and/or look it up, the encoding features more closely match the retrieval contexts than those created by first getting hints, which are never available during retrieval at test time. Thus, asking for help when you really don't need it hurts learning. On the other hand, failing to ask for help when you do need it hurts learning by reducing the number of appropriate knowledge events. Thus, if explicit instruction on help-seeking succeeds in getting students to use a better learning strategy, it should increase robust learning. The help-seeking skills acquired through this instruction should also help students in becoming more self-supervised learners.
• After students have solved a problem, they can be asked reflection questions about the solution, such as identifying its critical steps or indicating how the solution changes if the given information is changed slightly. Some reflection questions, if answered fully, elicit knowledge events that do not occur during problem solving. For example, some reflection questions challenge students to apply fundamental knowledge that they have not yet had to retrieve and therefore only exist as "inert" or unused knowledge. Such questions therefore stimulate the robust learning process of rederivation. Other questions, particularly "what if" scenarios that alter the given situation prompt generalization or adaptation of overly specialized knowledge to new situations. Finally, reflection questions model an important learning strategy-namely, the ability to generate comprehension-monitoring questions. Therefore, they potentially promote metacognitive skills needed for self-supervised learning. However, it is difficult to get students to fully engage in such reflective knowledge events. Three methods for increasing student knowledge events during post-practice, reflective activity will be compared: Interacting with a natural language tutoring system that encourages mixed-initiative interaction; interacting with the same tutoring system but without the mixed-initiative encouragement; or merely reading a monologue instead of participating in a typed dialogue. All these should cause more learning than a control condition, where students are not given reflection questions. However, it is unclear which method of engagement will elicit the most knowledge events that facilitate robust learning, as measured in terms of long-term retention and transfer. Long-term retention will be measured with respect to student performance on course exams administered several weeks after the training intervention. Transfer will be measured by test items that require application of knowledge targeted by the intervention to novel situations. (Katz)
• A scripted approach to learning may help students develop good meta-cognitive strategies and quickly and thoroughly learn materials. Compared to unscripted collaboration or working alone, scripted peer tutoring should improve learning and help learning occur more rapidly. For instance, in the study (McLaren, Rummel, Kalchman & Spada, study 1) scripted peer tutoring involves two students taking turns being tutor or tutee. The tutor first prepares by solving individually a problem using an intelligent tutor, and then coaches the tutee through the same problem. Afterwards, their joint solution is scored by a tutoring system and their errors are discussed by it. They switch roles and continue. The scripting should encourage the student playing the tutor role to be more engaged in problem solving (as opposed to guessing) and thus have more knowledge events.
• A jigsaw method of problem solving has students prepare for collaboration by solving individually different subproblems. The solutions of these subproblems are then combined, with modifications, during collaborative solution of a larger problem. Because the stakes are higher for solving a subproblem, and the solution may have to be explained or modified, students should be more engaged in its solution than in normal, unscripted collaborative problem solving. The higher engagement causes more relevant knowledge events and thus more robust learning (McLaren, Rummel, Kalchman & Spada, Study 2). This study is also an explicit exploration of the complementary benefits of foundation building and sense making in that it attempts to combine the benefits of foundation building (i.e., the more "drill-and-practice approach" of individual tutoring) with sense making (i.e., the explanations and dialogue inherent in collaboration).
• Worked examples are a form of interaction which is completely controlled by the student. The student can ignore the next step presented by the example, work out the step mentally, and then see what the example says (anticipatory self-explanation). Alternatively, the student can just try to memorize the steps for use later. The students' learning strategy for the example strongly determines the knowledge events that will done while studying it, and thus the amount of robust learning. On the other hand, solving problems with a tutoring system that will, if asked enough times, reveal steps is also a form of interaction that is at least partially controlled by the student. Two studies (McLaren, Yaron & Koedinger, study 1; Ringenberg & VanLehn) are comparing worked examples to tutoring.
• When students study film clips selected to illustrate uniquely French cultural practices, watching the clips should cause less learning than receiving attention-directing instruction (such as stopping the clip to predict the next event) and feedback while watching the same film clips. Appropriate knowledge events should be increased asking the students to retrieve their knowledge of French culture in the context where it could be applied, and to hypothesize about and characterize the values and behaviors in the film. Such knowledge events are likely to lead to higher quality of peer discussion of the current clip, thus leading to greater preparedness for future learning. (Ogan, Aleven & Jones).

As seen in the explanations above, the theory redefines the problem of explaining robust learning to be a problem of explaining why knowledge events vary, but the theory does not yet provide strong predications about which kinds of interaction elicit more engagement and more knowledge events. This is an area where the PSLC is poised to make advances.