- - - From: Alan D. Berfield Mixed Initiative How good is Lookout? Have tests been done to see how people react? Is it annoying like that horrible Office paper clip? Guinn Has any other/further work been done on finding a good synthesis of user modeling techniques? - - - From: Matthew T. Bell Guinn: The concept of using computer-computer simulation for developing human-computer dialog systems is intriguing. I'm curious as to how he would annotate the resulting dialogs for naturalness or acceptability to humans. How would this be done? Ishizaki et al On pg 83 the authors speak of mixed initiative dialogs as ones where a user can respond to an initiating utterence with another initiating utterence. What would be a good example of this where the discourse remains coherent? Is this distinction, that of initiating vs. responding, truly natural or useful? For example, I can imagine a human initiating a new subdialog in response to a request of another human. Hirst essay list: The authors contemplate mixed-initiative systems as systems which entail far more than dialog. To what extent does research in discourse and dialog dominate / inform / define research in mixed initiative? - - - From: hcl@cs.pitt.edu "Mixed-Initiative Interaction" I think control of dialogue vs. control of the task should be added to the list of subjective metrics (p.23). In a computer- to-human setting, shouldn't it be considered a win if the user had the feeling of control, regardless of the initiative details of the interaction? Allen's description of the differences between humans and automated planning, with each bringing its best to the table, is a nice idea, but how is it truly possible for the system to leverage these special human abilities with limited representation of them? How can the line be drawn between trusting the human's decisions (even it is seems to the system to not be correct) and pushing the system's agenda? Ishizaki, et al. I probably missed something, but at a high level, this paper seems to be reminding everyone to remember that non-mixed-initiative dialogue is still a useful goal. Their conclusion doesn't seem incredibly surprising to me... was there a realistic chance that mixed initiative dialogue could turn out to be more efficient? Doesn't this truly depend on the skills of the CPs? - - - From: Andy P. Gaydos Smith and Gordan, p. 159: I found it interesting that expert users developed a ritualistic interaction with the machine. Should a dialog system be designed to have a sense of personality, for example, during the greeting, even if it the user will simply trigger scripted responses? Would this improve the user satisfaction with the system? Would expert users using a system for a real purpose and information develop the same behavior? Ishizaki, Crocker, and Mellish: What was the significance of their result? First, does their computer-computer dialog results apply to human-computer dialog? Second, mixed initiative seems to be a good goal to strive for, based on reasons in Allen's paper and others. Andy Gaydos - - - From: Roy Wilson Smith and Gordon I think the most important statistical aspect of Table 3 is the interaction of mode and subdialogue. I wonder if the authors had enough data to test the statistical significance of the difference of the two matrices in Table 5. Missing Axiom Theory is a very parsimonious way of accounting for computer-to-human initiative: namely, without an explicit user model. Guinn The presentation and discussion of the principles of mixed-initiative dialog is very nicely done, but the definition of dialog initiative seems to be based on the unspecified (at least up to page 112) notion of (perhaps shared) expectation. There is a theory from mathematical sociology that makes it possible to specify initial expectations and to update them based on communication patterns, etc. Finally, it seems to me that the mathematical analysis of efficiency would be more psersuasive and cleaner if it were (assuming it can be) formulated in terms of Bayesian networks (a la Horvitz, et. al). Ishizaki, et al. Although it is useful to generate artificial dialogues, the work in this study is TOO artificial: the authors give no guidance as to how one might apply these results to human-computer interaction (the validity issue mentioned by Chin), nor do they suggest a research path for addressing this issue. "Mixed-initiative interaction" I very much agree with Guinn that simulations are useful in helping formulate dialogue mechanisms, providing data for comparing the behaviors generated by artificial and natuiral dialogues, and provide a testbed for testing dialog theories. I like Allen's presentation, especially, table 2, because it provides empirical support for the intuition claim that planning systems need to handle mixed-initiative. I reading Horvitz's account, I wonder (echoing HCL) whether student-ITS amounts to a joint activity: perhaps (as Ilya may have suggested) a typology of subtypes are needed for both collaboration and/or joint-activity. - - - From: Eric Williams Smith and Gordon This domain seems poorly modeled as task-oriented. Since the participants were electronic circuit novices (and the system is an expert of sorts), this domain seems better modeled as tutorial. Anyone else think so? Allen, Horvitz, and Guinn I'm a bit skeptical of the usefulness computer-computer dialogs. It seems like a catch 22 situation to me. We don't understand human-computer dialog well because we can't easily model how a human will respond to a particular computer dialog system. To get an idea of how a human would behave, we model one as a computer dialog system. The problem is we need to know how a human would behave to accurately model one. Thus we are stuck in a paradox. ---