Masahiro Araki and Kazunori Komatani and Taishi Hirata and Shuji DoshitaA Dialogue Library for Task-Oriented Spoken Dialogue Systems |
[Full Text] [send contribution] [debate procedure] [copyright] |
| No | Comment(s) | Answer(s) | Continued discussion |
|---|---|---|---|
| 1 |
16.2.2000 Arne Jönsson |
24.3.2000 Masahiro Araki et al |
|
C1. Arne Jönsson (16.2.00):
This paper presents interesting ideas on a dialogue library for dialogue systems, however, it lacks details to really contribute to the research field. The paper merely describe three abstract tasks and the corresponding task libraries in psedu code. The code describe the overall structure of interaction and few probelamtic cases are discussed. The dialogue libraries presented are so general that it is hard to see how they work. Just one example, in figure 14 the authors have the the statement if(understand). What does "understand" mean, i.e. how does the system knows that it has understood? Furthermore, is this procedure general and applicable to all systems or is it assumed that each system needs its own "understand" procedure. If so, does the library only include fairly obvious and abstract descriptions of dialogue moves and interaction patterns or is there more behind the scene?
There is nothing on how focus is handled. How is this supposed to be incorporated in the dialogue library. How can different focussing startegies be adopted and used?
Concerning the list of discourse tags. How does the list of discourse tags in Figure 11 relate to other such lists, e.g. the one developed within the DRI project and what tags are necessary for various types of abstract tasks presented in the paper. Furthermore, the evaluation seems to me unrealistic. As I understand it the authors have themeselves created dialogues for their application and evaluated their approach based on that. I do not think that it is possible to do that without omitting those phenomenon that might be probelamtic and thus this evaluation is not realistic.
To summarise. The paper must include more details on the elements of the dialogue library, how they are identified and used and how focus managament is incorporated in the approach. The approach must also relate to other approaches to dialogue management. If space is a problem, the introduction can be shortened as most of what is in there, i.e. section 2, is fairly well known and need not to be examplified or elaborated on. Instead I would like to have more details on how the dialogue library elements work and how they handle various problems such as interuptions and control. For instance, how is the dialogue in figure 10 handled, what information is included in the dialogue library elements when this dialogue is processed and how does the dialogue manager selects which task to pursue etc.
Arne Jönsson, Linköping University
C1. Araki et al (24.3.00):
Dear Arne Jönsson,
Thank you for your questions and comments.
Comments:
The dialogue libraries presented are so general that it is hard to see how they work. Just one example, in figure 14 the authors have the the statement if(understand). What does "understand" mean, i.e. how does the system knows that it has understood?The authors reply:
In database search task, we define "understand" as identifying the database which is to be searched.
Comments:
Furthermore, is this procedure general and applicable to all systems or is it assumed that each system needs its own "understand" procedure.The authors reply:
There are two types of "understand" procedure. One is for identification of database described above. The other is for identification of which part of XML document is to be explained. We would like to modify the "understand" procedure in Figure 14 and 15 to more specific one.
Comments:
There is nothing on how focus is handled. How is this supposed to be incorporated in the dialogue library. How can different focusing strategies be adopted and used?The authors reply:
In our design, focus is handled in semantic analysis phase. Most recent salient objects in previous utterances are remained and are used for ellipsis or anaphora resolution. That is, focusing is independent to dialogue library. We realized it causes problem in e. g. embedded dialogue. It remains our future work. We will add our focus handling strategy in chapter 3.
Comments:
Concerning the list of discourse tags. How does the list of discourse tags in Figure 11 relate to other such lists, e.g. the one developed within the DRI project and what tags are necessary for various types of abstract tasks presented in the paper.The authors reply:
This tag scheme is one dimensional. Therefore, all the category are necessary for describing dialogue library. On the other hand, DRI scheme is multi-dimensional. It requires indeterministic rules in using library description.
Comments:
Furthermore, the evaluation seems to me unrealistic. As I understand it the authors have themselves created dialogues for their application and evaluated their approach based on that. I do not think that it is possible to do that without omitting those phenomenon that might be probelamtic and thus this evaluation is not realistic.The authors reply:
Our dialogue library is designed by top-down form the abstract task point of view. On the other hand, hand written test dialogue is described by bottom-up method form the task domain point of view. In evaluation, we intended to compare how these two approaches matched.
Comments:
To summarise. The paper must include more details on the elements of the dialogue library, how they are identified and used and how focus management is incorporated in the approach. The approach must also relate to other approaches to dialogue management. If space is a problem, the introduction can be shortened as most of what is in there, i.e. section 2, is fairly well known and need not to be exemplified or elaborated on. Instead I would like to have more details on how the dialogue library elements work and how they handle various problems such as interruptions and control. For instance, how is the dialogue in figure 10 handled, what information is included in the dialogue library elements when this dialogue is processed and how does the dialogue manager selects which task to pursue etc.The authors reply:
We would like to modify our paper according to your comments.
Thank you for your contributions.