Spoken document understanding and organization

Spoken documents associated with the network content are critical for retrieval and browsing. An overview is given of various technology areas reaching towards this goal in a unified context. Technology areas covered include named-entity (NE) extraction, segmentation, and information extraction for the spoken documents as well as automatic summarization, title generation, and topic analysis and organization. The relevant problems and issues, general principles, and basic approaches for each area are briefly reviewed. A framework for properly integrating all these different technology areas is proposed, in which four different levels of processes are defined and bottom-up and top-down relationships are discussed. An initial prototype system for such purposes has been developed by National Taiwan University. The resultant system used broadcast news in Mandarin Chinese as example spoken documents.