Abstract

In this Master Thesis the focus is on doing the Topic Segmentation task. A specific Topic Segmentation tool will be developed for the Spoken Broadcast News (BN) Retrieval demonstrator system that Philips Research in Aachen Germany is working on. A working prototype has been implemented in this project. The main focus will be on the television BN streams, such as CNN. Topic Segmentation is still an unsolved problem, but there are already some great ideas available that provide reasonable Topic Segmentation results. Different solution approaches in different areas are analyzed, and a new adapted Topic Segmentation approach that fits the system architecture of this demonstration system has been developed. In general, there seems to be only three main categories of features for identifying topic boundary positions. They are text-based, audio-based and TV/video-based feature cues. But not all available feature cues are usable at the moment or in the near future. The standard and most important ones are combined in the new adapted solution approach, the BITS approach. Some test experiments have been performed by running the developed prototype as a standalone module in this system architecture. The most important tests are for finding the optimal values for parameters used in this Topic Segmentation tool, and for performance measurement when improvements are added to the BITS approach. With the results from the test experiments other people can continue building newer improved Topic Segmentation tool versions.