Susan W. McRoy and Syed S. Ali A Practical, Declarative Theory of Dialog

[Full Text] [send contribution] [debate procedure] [copyright]

C1. Richard McConachy and Ingrid Zukerman (23.2.00):

This paper attempts to cover a lot of ground in a comparatively small amount of space. As such we found it difficult to get a good feel for how several of the component pieces worked and how they are integrated. This is particularly salient in light of the fact that the components that are actually described, i.e., the parser, RRM, ARGUER and YAG, are treated quite separately, and the only thing they seem to have in common is the knowledge representation. The few rules that are actually presented give an inkling of the operation of the system, but their generality is unclear, as is their ``knowledge intensiveness'' -- how many rules are there of each type, what sort of coverage do they have? How does this compare to other dialogue systems?

One of the main contributions of the paper, viz the fourth and fifth levels of the discourse model, is described in a very cursory manner, even though these are the components that enable the higher functions of the system. You state that these components ``are needed to allow the system to interpret context-dependent utterances or to deal with unexpected utterances''. How are they actually used? Likewise, it is unclear how a representation such as that in Figure 4 is obtained. Could you please describe the process that yields this representation? Also, it is not clear whether the described system is implemented as a whole (or what is the stage of the implementation), or whether these are separate (but related) research efforts that will be eventually pulled together.

We believe that to answer these concerns, a longer submission is required which details the different components of the system and describes how they actually work. In addition, we suggest that the descriptions of ARGUER and RRM (which are not really complete enough to be helpful) be removed in favor of more a detailed description of the operation of the dialogue framework and system.

Additional (more detailed) questions

• You indicate that the input and output can be in several modalities and media. However, the architecture seems to account for text only. Can you please elaborate on the components that handle the other media?
• You also say that your architecture builds on your prior work. In what way? How is it different from the system described in [McRoy et al. 1997]?
• This is both a personal and a general question. In page 14 you say ``In contrast to [.... ], we use a general computational method for establishing relations between propositions''. What is not general about using Bayesian networks? [Zukerman et al. 1998] Likewise, we don't see any generality problems with IACAS [Vreeswijk 1997] or HERMES [Karacapilidis and Papadias 1998]. Could you please explain in which ways your argumentation approach is more general than those of these systems? Also, what do you mean by ``establishing relations between propositions''? Are we talking about defining and expressing the underlying relationship between two propositions, or showing how they interact with each other for some particular purpose, e.g., in some particular argument?
• How is your realization system different in functionality from FUG or the systemic framework? How does it work so that its speed is independent from the number of template types? Anyway, do you mean template or template types? If the former, then please explain, if the latter, then why does the number of types matter?

References
Karacapilidis, N. and Papadias, D. (1998), Hermes: Supporting Argumentative Discourse in Multi-Agent Decision Making. In AAAI98 Proceedings -- the Fifteenth National Conference on Artificial Intelligence, Madison, Wisconsin, pp. 827-832, AAAI Press.

McRoy, S., Haller, S. and Ali, S. (1997), Uniform knowledge representation for NLP in the B2 system. Natural Language Engineering , 3(2), 123-145.

Vreeswijk, G. (1997), Abstract Argumentation Systems. Artificial Intelligence 90 (1-2), 225-279.

Zukerman, I., McConachy, R., and Korb, K. (1998), Bayesian Reasoning in an Abductive Mechanism for Argument Generation and Analysis. In AAAI98 Proceedings -- the Fifteenth National Conference on Artificial Intelligence, Madison, Wisconsin, pp. 833-838, AAAI Press.

Richard McConachy and Ingrid Zukerman, University of Monash

A1. Susan McRoy and Sy Ali (26.3.00):

Thank you Richard and Ingrid.  We appreciate the care that you put into your comments

Comments:

We are trying to do two things in this paper: Firstly, to describe an overall architecture and secondly, to describe how  dialog tasks (which are typically specified only procedurally) can be capured in a declarative framework. We agree that the presentation is a bit terse and, with the editors approval, we will expand our final version to make things clearer. To answer the question, our approach is knowledge intensive. Although we are still working on the representations, for each step there are typically at least 5 rules for each interpretation type at each level of analysis. We deal with this complexity, in part, by partitioning the semantic network representations, although this introduces a cost for accessing the partitions. We are not aware of any other declarative approaches to dialog, so we do not believe there is a comparison to be made of the sort suggested.

Comments:

We will certainly provide more detail about the fourth and fifth levels of the discourse model and how they would be obtained. Although the first 3 levels are fully implemented and integrated, the last two are implemented but not fully integrated.  There is an integrated system that uses the argumentation schema but with a subset of the rules that we have written for levels 1 and 2; we are revising our work on misunderstandings and have not yet integrated it into our current declarative model. This will still be the case when we submit our final paper.

Comments:

We agree -- we will either delete material that does not stand alone or expand upon it.  Given its current state of development we are likely to expand ARGUER and omit RRM.

Additional (more detailed) questions /Replies

• You indicate that the input and output can be in several modalities and media. However, the architecture seems to account for text only. Can you please elaborate on the components that handle the other media?
• The core of the architecture is that all modalities and media are the same at the level of communicative act.  It is within the input manager that non-text input is mapped onto a propositional form (roughly a locutionary act) such as "select(user,storyProblem)" when the corresponding interface token is moused.  We have not yet implemented multimedia output although our generator could produce it if we implemented non-text primitives. (For example, an implementation that uses the same knowledge representation framework, SNePS, but not same content or dialog structure, has implemented simple speech and graphics type animations in VRML.)
• You also say that your architecture builds on your prior work. In what way? How is it different from the system described in [McRoy et al. 1997]?
• The archtecture extends the notion that the interpretation of communicative acts is driven by expectations and these expectations are based on a combination of knowledge including linguistic, intentional, and social conventions.  The latter two involve nonmonotonic reasoning--plausible, but possibly unsound inferences that we make based on the information that is available at the time and which are revised when contradictory evidence is received. Both the current and earlier work use schema to specify these plausible rules.  The new architecture extends the idea of schemas to handle argumentation and integrates the reasoning components into an end-to-end (parser to acting system to generator) system.
• This is both a personal and a general question. In page 14 you say ``In contrast to [.... ], we use a general computational method for establishing relations between propositions''. What is not general about using Bayesian networks? [Zukerman et al. 1998] Likewise, we don't see any generality problems with IACAS [Vreeswijk 1997] or HERMES [Karacapilidis and Papadias 1998]. Could you please explain in which ways your argumentation approach is more general than those of these systems? Also, what do you mean by ``establishing relations between propositions''? Are we talking about defining and expressing the underlying relationship between two propositions, or showing how they interact with each other for some particular purpose, e.g., in some particular argument?
• The sentence you allude to is poorly worded on our part. (We will reword it in the final paper.)  No implication about the generality (or lack) of the mentioned work was intended, rather that ARGUER's schemata (and their use) are based on standard first-order logical reasoning and, thus, could be utilized by any implementation of a reasoner.
  The schemata allow ARGUER to establish "attack" or "support" relationships between propositions (although, in principle, other relationships are possible), so we're clearly interested in how propositions interact for the purposes of determining how an utterance's propositional represention might be part of an argument in context.
• How is your realization system different in functionality from FUG or the systemic framework? How does it work so that its speed is independent from the number of template types? Anyway, do you mean template or template types? If the former, then please explain, if the latter, then why does the number of types matter?
• Our realizer, YAG, differs from previous ones in two respects: first its functionality is geared toward use by knowledge based systems. It accepts underspecified syntactic and semantic representations. The latter fits well with the semantic representations that we use for reasoning about dialog, in contrast to systems that expect a nearly complete description of syntax. The second key distinction is that our template based approach is fast and robust---so it can keep up with real-time dialog. Unlike a unifier or systemic grammar, realization time does not depend on size of the grammar because the application or the knowledge engineer specifies which template to use. (If the application does not specify a template, then the realizer will retrieve it from a hash-table that includes an entry for each type of proposition; the entry includes the name of the template to use and the mapping from arguments of the predicates to slots in the template.) Thus, the speed of YAG depends on the complexity of the particular template that is selected and not on the number of template types.  Short utterances are always produced more quickly than longer ones.