The PSLC Interactive Communication cluster
The studies in the Interactive Communication deal primarily with learning environments where there are two agents, one of which is the student. The other agent is typically a second student, a human tutor or a tutoring system. Both agents are capable of doing the instructional activity, albeit with varying degrees of success. They communicate, either in a natural language or a formal language, such as mathematical expression or menus. The main variables are:
- What part of the work is done by which agent? On one extreme, the student does all the work while the other agent watches. On the other extreme, the student watches while the other agent does all the work. In the middle, the two agents collaborate somehow.
- Who makes the choice about which work is done by which agent? The student, the other agent or a fixed policy of some kind?
Our hypothesis is that learning by doing is the best, except that as the student takes on more work or more challenging work, the error frequency or the time to recover from errors may begin to interfere with learning. Communication also can interfere when learning, in that it takes time and cognitive resources, and that it is never perfect. Thus, learning can be optimized by somehow balancing the work done by the student, the work done by the agent and the work done by both in communicating.
Background and Significance
Instructional dialogue has mostly been studied in classrooms (e.g., Lave & Wenger, 1991; Leinhardt, 1990) and workplaces (e.g., Hutchins, 1995; Nunes, Schliemann & Carraher, 1993). In order to investigate more tractable albeit still complex situations, most of our work focuses 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. Moreover, the dialogue are task-oriented (Grosz & Sidner, 19??) in that the participants are working together on a task rather than simply conversing with no shared goals or with opposing goals.
Given that many studies of the structure of dyadic instructional dialogue exist (e.g., Fox, 1993; Graesser, Person & Magliano, 1995; MacArthur, Stasz, & Zmuidzinas, 1990), we are focusing on what properties of interactive communication promote robust learning. Earlier studies (e.g., VanLehn, Graesser et al., in press; Katz, Connelly & Allbritton, 2003; Evens & Michael, 2006; Cohen, Kulik & Kulik, 1982) found surprisingly mixed results. Although most studies showed that interactive communication was more effective than less interactive instruction, it was not always better. Thus, the next step in this important line of research is to determine when different types of interactive communication are effective and why.
To be developed, but will probably include:
- Agent: Something that can perform the instructional activity. Typically a student, a tutor, a tutoring system or a simulated student. In the extreme case, an agent can be a passive medium, such as text or a video, that presents a performance of the activity. For instance, if the instructional activity is solving physics problems, then a worked example, such as the ones shown in a textbook, is an agent.
- Initiative. This measures the ratio of the work initiated by the two agents. A dialogue with lots of student initiative is one where the student spontaneously initiates work on the activity. A dialogue with lots of tutor initiative is one where the tutor either does the work or requests (in the speech act sense of “request”) the student to do the work. The “initiative” term comes from linguistics, whereas a synonymous distinction, learn control vs. teacher control, comes from education.
- Zone of proximal development. When instruction is laid out on a scale of difficulty from easy to hard, there is a region where the instruction is too hard for the student to learn effectively from it without help, but still just easy enough that the student can learn if given help, typically from a second agent. This region is called the zone of proximal development (ZPD), a term from developmental psychology.
The research problem addressed by this cluster is: What properties of interactive communication promote robust learning?
- Some studies in the Interactive Communication cluster examine the impact on robust learning of different types of interactive communication, by contrasting two forms of interactive communication. Examples include compare scripted vs. unscripted peer collaborative problem solving.
- Other studies compare instruction with and without specific kinds interactive communication, e.g., by having students work alone or in pairs, or comparing self-explanation done alone to dyadic, interactive explanation generation.
- A third class of manipulation holds most of an interactive communication constant and varies only a small part of it. For instance, a video may be viewed with and without interactive prompts inserted at key points. It should perhaps be noted that, as in the physical sciences, this study-it-in-isolation strategy is risky. Just as a heart extracted from an animal doesn’t behave exactly like one that still resides in the animal, the process studied in isolation may not behave exactly like the one that occurs in interactive communication. Nonetheless, significant progress has been made in the physical sciences by using this isolation strategy, so it may help the science of learning as well.
Measures of normal and robust learning.
When student engage in collaborative learning with another agent where the collaboration somehow appropriately balances the work done by the agents and their communication, then learning will be more robust than it would if the learning environment had just the student and not the second agent.
Assuming a control condition where the student works alone or with only limited interaction with the second agent, there are 3 cases:
- If the instruction is in the students’ zone of proximal development (ZPD), then a second agent’s help can increase learning compared to a control condition.
- If the instruction above (more difficult than) the ZPD, then the student makes too many errors and/or requires too much communication with the second agent, which thwarts learning. Thus, learning is equally ineffective in the two conditions.
- If the instruction is below (more easy than) the ZPD, then the student can learn just as much working alone as when working with the second agent. That is, learning is equally effective in the two conditions.
This idea can be rephrased in terms of the PSLC’s general hypothesis. Robust learning should occur under two conditions. First, the instruction should be designed to have the right paths, which means that there is a target path that involves the student doing almost all the intellectual work (learning by doing) and many alternative paths where in the second agent does most of the work. Second, the student should choose the paths so that they take the learning-by-doing path by default, and take the other paths when the learning-by-doing path is too difficult for this particular student at this time. Moreover, the choice of taking an alternative to the learning-by-doing path should take into account the overhead and reliability of communication, which is generally higher on the alternative paths.
- Collaborative Extensions to the Cognitive Tutor Algebra: Scripted Collaborative Problem Solving (Rummel, Diziol, McLaren, & Spada)
- Collaborative Extensions to the Cognitive Tutor Algebra: A Peer Tutoring Addition (Walker, McLaren, Koedinger, & Rummel)
- Tutoring a meta-cognitive skill: Help-seeking (Roll, Aleven & McLaren) [Moved to Refinement and Fluency, Was in Coordinative Learning]
- Does learning from worked-out examples improve tutored problem solving? (Renkl, Aleven & Salden) [Was in Coordinative Learning]
- Visual-Verbal Learning (Aleven & Butcher) -- Elaborated Explanation condition is the relevant manipulation
- The self-correction of speech errors (McCormick, O’Neill & Siskin) [Was in Fluency and in Coordinative Learning]