Interdependencies among dialog acts, task goals and discourse inheritance in mixed-initiative dialogs

This paper describes a study of the interdependencies among dialog acts, task goals and discourse inheritance in mixed-initiative dialogs in the restaurants domain. Our study is based on 199 dialogs, with disjoint training (169 dialogs) and test sets (30 dialogs). The training set is annotated manually in terms of task goals and dialog acts and tagged automatically in terms of semantic and syntactic categories for each request (from the customer) and response (from the waiter). Based on observations from the process of annotation, we have written a set of category inheritance and refresh rules, which constitute our selective inheritance strategy. We compared the selective strategy with two control strategies -- (i) no categories are inherited; and (ii) all categories are inherited throughout the dialog session. Comparison is based on the automatic identification of task goals and dialog acts. The selective inheritance strategy outperformed the two control strategies and identified the correct task goals for 92.6% of the dialog turns and the correct dialog acts for 97.8% of the utterances in the test set. We have also developed a discourse inheritance procedure which correctly handled 95.9% of the dialog turns in the test set.