================================================================== Date: Tue, 15 Jan 2002 20:12:09 -0500 From: Amy Soller To: Matt Bell Subject: q's for class sadek & de mori: Suppose, given the framework presented in section 5.3, that an agent wanted to ask a question in order to learn something new. Then, is it the case that the agent would have to know what it is that it does not know (and desires to learn)? If, so, how is this possible? grosz & sidner: Grosz & Sidner's theory was designed to study written discourse. Can it be applied to dialog as well? For example, can the hearer recognize the speaker's intention (in verbal communication) in the same way as the OCP recognizes the DSP in discourse? What happens when the hearer fails? ================================================================== Date: Tue, 15 Jan 2002 21:40:49 -0500 (EST) From: Andy P. Gaydos To: mbell@cs.pitt.edu Subject: Dialog Systems Question Grosz's theory of discourse structure handled several interuptions - true interuptions, flashbacks to fill in missing places, and digressions. Many systems seem unable to recover from a misinterpretation of the ICP's discourse, even if the ICP tries to say that the current dialogue is not leading toward the ICP's desired results. Could Grosz's theory handle interuptions made by the ICP to correct misinterpreted data as well as missing data? Could the system then recover and continue toward the goal from where the error was made? Andy Gaydos ================================================================== Date: Wed, 16 Jan 2002 00:14:56 -0500 (EST) From: hcl@cs.pitt.edu To: Matthew T. Bell Cc: H. Chad Lane Subject: 3710 questions for 1/16 Grosz & Sidner: G&S make it clear that they are looking for an account of the structure of discourse, and not meaning. It certainly makes sense that a theory of discourse meaning requires a theory of discourse structure. I have to wonder, what would a "theory of discourse meaning" look like? Is it good enough to just string together semantic representations of the parts of a discourse to constitute the meaning of the whole? Is it necessary to seek a theory of discourse meaning at all? Will intentional/attentional structural analysis suffice? I can easily imagine a monologue or dialogue that is mildly ill-formed, but well within the realm of acceptability. For example, someone goes on to a new topic, but wants to rewind briefly for some reason. Shouldn't there be a "trash can" somewhere for recent focus spaces that allow them to be "recycled" (I apologize for the MS-tainted language)? ================================================================== Date: Wed, 16 Jan 2002 01:37:22 -0500 (EST) From: Antonio Roque To: mbell@cs.pitt.edu Subject: CS 3710 discussion questions 1. How do we evaluate the effectiveness of theoretical assumptions? Last class we talked a bit about the difference between AI and Cognitive Science; the latter being focused on actual human cognition, the first not necessarily so. An example of this is in Sadek and De Mori's mention of intentionality (p. 540), and the uses of assuming that something is intentional if doing so helps produce something coherent and useful. The claim is that the validity of that assumption is irrelevant, what is important is the utility of the assumptions. It may be useful to assume intentionality and to do similar things such as describe user-friendliness as conviviality (p. 524) and make it an emergent property of "intelligent" features such as contextual interpretation and flexibility of interaction.(p. 528) But how do we evaluate this usefulness? We may evaluate the effectiveness of the resulting system using metrics such as PARADISE, but in what way do we evaluate the assumptions behind that system, specifically, intentionality, or the ways in which the intelligent features required of the dialogue system were defined? Are the assumptions exactly as good as the systems they produce? 2. How can isolated subsystems help manage computationally-difficult systems? Sadek et al describe the Natural Language Input and Output components of ARTIMIS as "isolated subsystems" that "are meant to be eventually integrated as logical theories in the rational unit" (p. 1033). In what way could such an approach be used to solve the problems of intractability that we discussed last class? ================================================================== Date: Wed, 16 Jan 2002 02:11:35 -0500 From: Noboru Matsuda To: mbell@cs.pitt.edu Subject: Q on Sadek98 [ The following text is in the "iso-2022-jp" character set. ] [ Your display is set for the "US-ASCII" character set. ] [ Some characters may be displayed incorrectly. ] Hi Matthew, Sadek, D. and Mori, R. De, "Dialogue Systems" 1. It takes only half an engineering-month to port French version of ARTIMIS/AGS into English version? Does it means that they have had technologies to build machine translation system? If not, what is the difference between machine translation and dialogue management system that seems like language independent? 2. The Dialogue Rational Agent seems to be more reliable and robust than the agents that are based on simple finite-state diagrams. However, DRA still needs a set of beliefs, uncertainty, and intentions written in FOL that could require many engineering-months? Why is the DRA still superior to other models? Thanks -- Noboru Matsuda (http://www.pitt.edu/~mazda) Intelligent Systems Program, University of Pittsburgh 3939 O'Hara St. LRDC #810, Pittsburgh, PA 15260-5189 USA Voice. (412) 624-2662 Fax. (412) 624-9149 ================================================================== Date: Wed, 16 Jan 2002 08:21:56 -0500 From: Roy Wilson To: mbell@cs.pitt.edu Subject: Questions Matt, Here they are. (1) In "Dialog(ue) Systems", it is implicitly claimed that the formulation of intention in "Defining a conversational agent" is weak. Yet, both papers rely on the B(elief), I(ntention), etc operators of modal logic to analyze communicative action. Do Sadek/De Mori implicitly contradict themselves and, if so, why? (2) In the "Discourse modeling" section of the earlier "Discourse and dialog(ue)" chapter, it is suggested that GST is relatively better than RST because it provides an explicit treatment of intentionality. Does the paper "Dialogu(ue) systems" provide an explicit account of intentionality and, if so, how? (3) In "Dialog systems" (p. 551), it is stated that "the efficiency of the implmented mechanisms is generally a function of the degree of genericity of the relationship between natural language and 'mental' representation". Are efficiency and genericity positively/negativatively/not correlated and why? Sic transit gloriam mundi? Roy -- Roy Wilson AI Programmer/Natural Language Generation rwilson@pitt.edu CIRCLE Group LRDC #701 Learning Research and Development Center 412-624-7464 University of Pittsburgh ================================================================== Date: Wed, 16 Jan 2002 08:52:40 -0500 (EST) From: Alan D. Berfield To: Matthew T. Bell Subject: ai questions Dialogue Systems (ch 15) - The mention of "conviviality" got me thinking about emotion in dialogue. Has anyone tried to identify and reason about emotion, or generate it appropriately in responses? As a side note, this then made me think of Douglas Adams: overly happy doors, depressed (and depressing) robots, and irritatingly helpful spaceship AI... Attention, Intentions, and the Structure of Discourse - In the sections about interruptions, they seem to assume there will be a "return" to the previous topics/DSP? I don't think that is always the case in human dialogue. Sometimes when a more interesting topic is started, old topics are simply dropped, even if their DSP has not been satisfied. ================================================================== Date: Wed, 16 Jan 2002 09:06:59 -0500 (EST) From: twilson@cs.pitt.edu To: Matthew T. Bell Subject: questions for class In Dialogue Systems, section 15.3.3, the authors talk about the importance of a user being able to contest something said by the system in supporting interaction flexibility. In the last paragraph of the section it says, "the possibility to contest is directly related to the global capability of a system to revise its beliefs." This seems like a very strong statement, and I was wondering if there is no other way to robustlly handle user contests than a highly developed belief representation system? ------------- This is a shorter version of my question from yesterday. I wonder how or how well the belief models and communication act models that we've read about handle miscommunication? Is it just a matter of detecting that in the past a miscommunication occurred, and correctly updating one's knowledge base accordingly? Theresa ================================================================== Date: Wed, 16 Jan 2002 09:37:40 -0500 From: Stefanie Bruninghaus To: mbell@cs.pitt.edu Cc: steffi@pitt.edu Subject: Dialog Seminar questions Hi Matt - here's the questions: Sadek/Mori paper: Overall, my question is whether the presented approach really works for building a usable system, or whether this is more of a theory without much practice? Here's some arguemnt why: It seems to me as if the NLU part of the system is not sophisticated at all (it's keyword extraction and then some heuristics to put these keywords together). And, the type of dialog is on the simple side too. The system seems to work quite OK as described in the paper - but what would happen when one does real NLP? The model does not account for (1) ambiguity in interpreting the caller's utterance, and (2) an interpretation of more complex utterances, which may contain expressions of preference, doubt - and hey, my favorite, negation! The agent metaphor is probably the appropriate tool here, but I am not convinced that this whole thing can be robustly scaled up. Another factor that makes me believe this is the nature of the dialogs presented - they look as if they are very system initiative. What would happen if the hearer starts to present something the model was not prepared for, or what if the called wants to go back? Furthermore, the system does not seem to account for errors - which I expect to be an inherent element of using ASR. Also, there does not seem to be any sort of evaluation. That should make anybody in LRDC doubt the approach! Grosz/Sidner paper (this is from the last meeting): I had a really hard time with this paper, and I don't think I really got a handle on it. However, after the Herb Clark talk, I was wondering what is the realtion between the Clark work and the Grosz/Sinder paper - both talk about different aspects of a joint activity and a dialog, respectively, and both line up the dialog and task to be accomplished and the speakers' intentions, respectively. To me, that seemed very similar - or did I seriously misinterpret one? See you later - STeffi. -- -------------------------------------------------------------------- Stefanie Bruninghaus Learning Research and Development Center Mail: steffi@pitt.edu University of Pittsburgh Web: www.pitt.edu/~steffi 3939 O'Hara Street Phone: (412) 624 - 6748 Pittsburgh, PA 15260-5159 -- USA Fax: (412) 624 - 9149 -------------------------------------------------------------------- ================================================================== Date: Wed, 16 Jan 2002 10:15:07 -0500 From: Vincent Aleven To: mbell@cs.pitt.edu Subject: questions Hi Matthew, Sorry for the late submission. My main question about the Allen Ch. 17 (which you may or may not be dicussing to day) is the same as raised before: The theory presented in Allen, chapter 17, makes many good points, for example that to understand dialog(ue), you need to look at the beliefs and intentions of the participants. Also that this raises problems of representing beliefs, beliefs about beliefs (e.g., representing that someone knows something without representing what it is exactly that she knows). What role does this kind of theoretical analysis play in the implementation of systems? (I suppose this will become more clear as we read more papers.) Also, how deep does one need to dig into this theory (representation of beliefs, planning communicative acts, etc.) in order to understand the literature on dialog(ue) systems? (I suppose this one will also become more clear as we read more papers. Vincent