TV News Story Segmentation Using Deep Neural Network

TV news programs usually contain multiple stories on different topics, and it is essential to locate the story boundaries for the purposes of video content indexing, search, and curation. With the exponential growth of video content, story segmentation will enable the consumers to view their favorite content effortlessly, and the service providers to provide their customers with personalized services. Given the dynamic range of topics, smooth story transitions in news, and varying duration of individual story, automated news story segmentation is a challenging task. This paper focuses on using linguistic information extracted from closed caption as the initial attempt to tackle this challenge, and our future work will integrate both audio and visual. A convolutional neural network framework with attention mechanism is proposed. The model is trained and tested on TDT2 data set, and it achieves an outstanding F-measure of 0.789 on the validation set and a reasonable F-measure of 0.707 on the testing set.