"... In PAGE 3: ... Moreover, a reply can be of two kinds, being either an attacking or a surrendering reply. Table1 specifies how to attack or surrender to another speech act. Table 1 contains the two combined communication languages.... In PAGE 3: ... Table 1 specifies how to attack or surrender to another speech act. Table1 contains the two combined communication languages. The idea is that the why-reject p speech act triggers a persuasion dialogue.... ..."

"... In PAGE 3: ...ppropriate response. A useful approach to this is the identification of DAs. A DA represents the meaning of an utterance or a query a5 . Each query is assigned a unique DA label drawn from a well-defined set (see Table1 ). Thus DAs can be thought of as a tag set that classifies queries according to a combination of pragmatic, semantic, and syntactic criteria a6 .... In PAGE 3: ... Queries are classified into two general categories, question and statement, which are again sub-categorized into primary or secondary, each of which consists of several DAs. As a whole, thirty domain-independent DAs are defined as in Table1 . These DAs enrich the available input for matching a query with a response.... In PAGE 3: ...ialogue Act Categorization module (see Fig. 1) classifies queries into DAs. Only one DA is assigned to the query in case of primary question or statement whereas several DAs are assigned to a query in case of secondary question or state- ment. Each question or statement has several predefined DAs as in Table1 . The Di- alogue Act Categorization module is implemented by automata that are constructed IAT2001: submitted to World Scientific on June 17, 2001... ..."

"... In PAGE 7: ...Table 6, Table7 , and Table 8 show the aver- age accuracy of each learning approach on the 20-fold cross validation experiments for domain action, speech act, and concept classification re- spectively. For DA classification, there were no significant differences between the TiMBL, C4.... In PAGE 8: ... Finally, Table 12 shows the results from two tests to compare the performance of combining the best output of the SA and concept sequence classifiers with the performance of the complete DA classifiers. In the first test, we combined the output from the TiMBL SA and CS classifiers shown in Table7 and Table 8. The performance of the combined SA+CS classifiers was almost identical to that of the TiMBL DA classifiers shown in Table 6.... ..."

"... In PAGE 10: ... 3.1 The combination First the combined communication language Lc is de ned in Table3 . As can be seen, the negotiation language is reformulated in the format of Section 2.... ..."

"... In PAGE 7: ...Table 8: Concept Sequence classifier accuracy. Table 6, Table7 , and Table 8 show the average accuracy of each learning approach on the 20-fold cross validation experiments for domain action, speech act, and concept classification respectively. For DA classification, there were no significant differences between the TiMBL, C4.... In PAGE 8: ... Finally, Table 12 shows the results from two tests to compare the performance of combining the best output of the SA and concept sequence classi- fiers with the performance of the complete DA classifiers. In the first test, we combined the output from the TiMBL SA and CS classifiers shown in Table7 and Table 8. The performance of the com- bined SA+CS classifiers was almost identical to that of the TiMBL DA classifiers shown in Table 6.... ..."

"... In PAGE 26: ... Classification performance is shown for each of the individual classifiers, as well as for the optimized combined classifier. Table1 0: Accuracy of Individual and Combined Models for Seven-Way Classification Knowledge HLD Set DEV Set DEV Set Source true words true words N-best output samples 2737 287 287 chance (%) 14.29 14.... In PAGE 32: ... This difference is likely to be captured to some extent by the pause feature. Table1 1: Feature Usage for Classification of Questions and Statements Feature Feature Usage Type (%) Dur regr dur 0.332 Pau cont speech frames n 0.... In PAGE 36: ...Table1 2: Feature Usage for Main Feature Types in Classification of Yes-No Questions, Wh-Questions, Declarative Questions, and Statements Feature Usage Type (%) F0 0.432 Dur 0.... In PAGE 38: ...outcome is shown in Table 13. Table1 3: Accuracy of Individual and Combined Models for the Classification of Questions Knowledge HLD Set DEV Set DEV Set Source true words true words N-best output samples 1852 266 266 chance (%) 50.00 50.... In PAGE 39: ... The all-features tree is complex and thus not shown, but feature usage by feature type is summarized in Table 14. Table1 4: Feature Usage for Main Feature Types in Classification of Incomplete Utterances and Non- Incomplete Utterances Feature Usage Type (%) Dur 0.557 Nrg 0.... In PAGE 41: ... Table1 5: Accuracy of Individual and Combined Models for the Classification of Incomplete Utterances Knowledge HLD Set DEV Set DEV Set Source true words true words N-best output samples 2646 366 366 chance (%) 50.00 50.... In PAGE 45: ... Since the present task pits backchannels against the longer agreements, an increase in the percentage of shorter backchannels (from training to test, as occurs when testing on the DEV data) should only enhance discriminability for the prosodic trees as well as for the language model. Table1 6: Accuracy of Individual and Combined Models for the Classification of Agreements Knowledge HLD Set DEV Set DEV Set Source true words true words N-best output samples 2520 214 214 chance (%) 50.00 50.... ..."

"... In PAGE 6: ...pinion) account for 83% of the words in our corpus (since e.g. backchannels and answers tend to be short). Table7 shows, however, that using utterance-specific lan- guage models can significantly improve WER for some dialog acts, and hence this approach could prove useful for tasks with a different distribution of utterance types. 5 Conclusions We have described a new approach for statistical modeling and detection of dis- course structure for natural conversational speech.... ..."

"... In PAGE 6: ...pinion) account for 83% of the words in our corpus (since e.g. backchannels and answers tend to be short). Table7 shows, however, that using utterance-specific lan- guage models can significantly improve WER for some dialog acts, and hence this approach could prove useful for tasks with a different distribution of utterance types. 5 Conclusions We have described a new approach for statistical modeling and detection of dis- course structure for natural conversational speech.... ..."

"... In PAGE 6: ...pinion) account for 83% of the words in our corpus (since e.g. backchannels and answers tend to be short). Table7 shows, however, that using utterance-specific lan- guage models can significantly improve WER for some dialog acts, and hence this approach could prove useful for tasks with a different distribution of utterance types. 5 Conclusions We have described a new approach for statistical modeling and detection of dis- course structure for natural conversational speech.... ..."