================================================================== Amy Soller Cahn & Brennan - How can "conversational tasks" be completely domain independent? They are defined as tasks that help to clear up misunderstandings, but this implies that the user is having trouble understanding some domain material. It seems like the point at which a presentation is rejected (or postponed, as in the case of Heeman et al.) might be a good point to reconsider the initiative distribution, as we discussed last week in class. So, perhaps one could model grounding efforts, and use the resulting model to strengthen one's case for an initiative swap. Along these same lines, I like the idea of using the recognition certainty to alter the grounding strategy - i.e. producing a more controlled or less structured interaction. =================================================================== Antonio Roque - I may be misunderstanding the ideas in Heeman et al., but they seem to be only pushing the grounding and contribution behavior into the taskbox design phase (from the dialogue model design phase), rather than simplifing the design in any way. What exactly do they mean by "factoring out"? - Clark and Schafer's ideas in "Contributing to Discourse" seem to fit so well with what NLP needs in terms of overcoming problems from speech recognition and natural language understanding that I wonder what inspired these ideas - he claims to be inspired by models of discourse from philosophy, linguistics, ai, and psychology, but it almost seems born from a specific project. ==================================================================== Theresa Wilson Heeman, Johnston, Denney, and Kaiser - "Beyond Structured Dialogues" This is obviously more of a proposal paper, representing a very early stage in the authors' work. One concern I have is that the authors are presenting "taskboxes" as their method of partitioning the dialogue. Doesn't this seem like just another name for a frame or template? ---------- Cahn and Brennan - "A Psychological Model of Grounding and Repair in Dialog" I think this was an interesting paper, and the "exchange" structure that the authors introduce into Clark and Schaefer's contribution model seems an intuitive for representing adjacency pairs. One question that I have though is how well their prototype will do as a more general purpose conversational agent? As the authors say, identifying the speech acts from their users is not difficult and there is little or no ambiguity that their system must resolve. The system can easily tell if an utterance is being grounded just by identifying the speech act. In a general purpose conversational agent, identifying speech acts becomes much difficult and contains much more ambiguity. ---------- Matheson, Poesio, and Traum - "Modelling Grounding and Discourse Obligations" This paper was also very interesting, but I feel that results of future work with evaluations are needed to determine how well their representation does in handling grounding, thus how useful it will be. I am curious as to exactly how much the dialogue on page 6 was cleaned up (note the footnote). This might have been informative as to how well their current system does that is presented in this paper. =================================================================== Alan Berfield Heeman et al. Do they have any experimental evidence to support their claim that the taskbox idea increases user-friendliness? Have they done any further work on their taskbox idea (e.g., adding hierarchial knowledge)? Matheson et al. What advantages would the more sophisticated approach to realising intentions (mentioned in 3.4) have? Cahn and Brennan They admit that they have not done any user tests. Is this still the case, or have they done some since the paper was written? ====================================================================== Matt Bell Contributing to Discourse: I love the hypothesis of this paper -- that human conversational behavior is more complex and involved than the other models discussed -- but wonder if the complexity of the model described does not present a question to us: How can we use this model, or models of like complexity, when having a computer dialog with a human? People don't seek to have relationships with their computers (unless they are disturbed or anti-social); it seems that much of the added complexity would be for the purpose of supporting far longer discourses than those which dialog systems typically target. On the other hand, the notion of confirming understanding and repairing misunderstandings would be vital for a successful system. For that, however, do we really need to go to Clark's model? ---------- Heeman et al Hmmn. I take it all back: Someone did try directly to implement Clark's model. The funny thing is, they seem to view it as a simplification rather than an added complexity. If simplified, however, how does this differ at all from simply taking a dialog strategy of *always confirm*? ---------- A psychological model of grounding and repair in dialog This seems to be a more detailed implementation of a prototype system than that discussed in Heeman et al, going a tiny step beyond Clark to posit that negative evidence causes embedding and positive does not (although this conclusion I saw as implicit in Clark's paper). Because this step taken beyond Clark is so small, and possibly even taken into account in his paper implicitly, my question is how does this model differ from or add to his? ---------- Matheson, Poesio, Traum: The theory here discussed seems to address the question of moving information from the set of ungrounded propositions to the set of grounded propositions, but does not seem to deal with rejected uncrounded propositions, nor how it is that a proposition becomes marked as ungrounded in the first place. It seems that the only way an utterence's content could be ungrounded would be if it had been rejected, not heard correctly, etc. (c.f. Clark's paper), in which case there really would be no strong reason to maintain its content at all -- or would there be? Perhaps there would be reason, if there was a strong chance that that information would become accepted later. If such is the case, how long/great is the store for such conditionally accepted/rejected information? ========================================================================= Chad Lane Clark & Schaefer: I'm not sure, but I think their theory assumes the CPs want to establish common ground at all times. If true, then the intentional act of not providing any acceptance evidence (e.g., purposely ignoring a question) will not be captured by the theory. Is this true? They attack the "RWRT" perspective on dialogue, but if you only consider automating dialogue behavior, and put user satisfaction as the top criterion, isn't it enough? Related... is RWRT a product of a model of contributions, or is that a model of contributions is necessary for RWRT? Cahn & Brennan: The idea of an exchange is interesting and potentially very useful, but is the define/execute distinction the right one? Looking at the right half of figure 5, A says "Thanks" and B replies "You're welcome." What is being defined by A and what is being executed by B? It seems to me that both are executing half of task, and this suggests a more general notion of exchange is needed. Matheson, et al: It's cool how an IS makes it easy to distinguish between grounded vs. ungrounded knowledge. However, as Cahn & Brennan point out, both CPs will have different views on the common ground. What is the position of these authors on this point? Is their IS just a representation of the system? If so, did they make it possible to "downgrade" facts from grounded to ungrounded when misunderstanding arises? =================================================================== Mihai Rotaru For all papers I found the contribution model one of the most simple and easy to understand models for dialogs. It seems to be more natural compared with others (rational agency, belief-intention-actions). However, its drawback seems to be its intensive use of semantics. Are there any successful implementations of this model? Did Peter Heeman proposal work? ================================================================== Eric Williams Clark and Schaeffer Interesting. The authors mention contribution to discourse and reference Grosz and Sidner as incomplete and/or oversimplified. While I agree with their assessment of the idealizations, I am somewhat skeptical of this representation's ability to work within the stacks mentioned by Grosz and Sidner, a representation I find very appealing and useful. In other words, the authors' notion of contribution seems too "local" and lacking in long-term collaboration such as is found in task-oriented and tutorial dialogs. Heeman, et al I'm thoroughly unimpressed. This seems like nothing new; just a rehash of frames and planners. Also, the sequential nature of data retrieval from users results in "canned" dialog. ==================================================================