The dialogue library for slot-filling task is constructed by refering the telephone shopping dialogue corpus. In order to evaluate the portability of the dialogue library, we examine whether this slot-filling library suit the another task. We selected target task is on-line banking. Because we do not have on-line banking dialogue corpus, we have written down the possible dialogue of on-line banking (48 utterances, 16 interaction unit). In principle, one interaction unit (typically, system's initiate and user's response) corresponds to one element of library.
As a result of the evaluation, the library covered 13 interaction unit (81%) but failed 3 interaction unit which is shown in Figure 16.
Fig. 16: Failed interaction unit
Case 1 and 2 are derived from the difference of application structure. Telephone shopping usually has one slot form and all slots are filled by user. On the other hand, on-line banking has some slot form (set up a fixed deposit, buy trust fund, money transfer, etc) and has slot dependencies. This pattern of slot-filling can be regarded as the most complex one. Therefore, by tuning our library to such type of task, we think that many slot-filling domain can be covered.
Case 3 is derived from another problem. As a matter of fact, we cannot write down all the error-and-recovery situation. This is a defect of our approach. However, if we develop learning method which can aquire various pattern of interaction from tagged dialogue corpus, the library gradually covers several pattern of dialogue.