Integrating acoustic and lexical features in topic segmentation of Chinese broadcast news using maximum entropy approach

This paper studies how to integrate multi-modal features in automatic topic segmentation of Mandarin broadcast news. The multi-modal feature integration problem is formulated within the Maximum Entropy (MaxEnt) scheme for topic boundary classification by maximizing the entropy and respecting all known constraints (i.e., multiple features contributions). We particularly consider two types of features: (1) acoustic features, which reflect the editorial prosody of broadcast news, including pause duration, speaker change and speech type; and (2) lexical features extracted from speech recognition transcripts, which capture the semantic shifts of topics, including two local cohesiveness features and a new boundary indicator based on overall cohesiveness. Compared to local lexical features, the new overall cohesiveness feature maximizes the lexical cohesiveness of all topic fragments and reflects the fact that topic transitions in broadcast news are smooth and the distributional variations are subtle. Experiments show apparent performance improvement in topic segmentation of Chinese broadcast news by fusing acoustic and lexical features within the MaxEnt scheme.