Discourse Marker Detection for Hesitation Events on Mandarin Conversation

The occurrence of hesitation events in spontaneous conversations can be associated with the difficulties in memory recall. One indicator of hesitation in speech in Taiwanese Mandarin is the usage of discourse markers. This paper introduces an approach to the detection of discourse markers that denote hesitation events. We propose a sequential labeling model to detect discourse markers in conversations by taking information on both acoustic level and word level into account. Experimental results show the integration of word-level acoustic feature extraction network significantly enhances the detection performance. Our approach for further applications is also discussed.