The effect of accent in recognizing dialog act in Chinese

Dialogue-act (DA) classification is a key step for the computer to understand natural-language dialogues. Some prosodic features have been recommended to improve the performance of DA classification; however, most of the prosodic features are based on the pitch variations. In the present work, we propose some new attributes for automatic classification of dialogue acts (DAs) from accent features. We test classification performance on solving a 15-DA classification task with the CASIA-CASSIL Corpus. Promising performance is observed when the accent features are added in training classification models and the misclassification rate drops about 12.57%. Although the accent features may not be sufficient for classification of DAs, it is shown that the accent features can improve the performance of DAs classification to a great extent.