Active learning for story segmentation of spoken documents

In story segmentation, it is often difficult to gather the segmented data to train a new model for the purpose of supervised learning. Therefore, how to gather the useful data and to reduce the human effort on segmenting stories is an important issue. We apply active learning to selecting the most informative examples to train a supervised model more efficiently. Active learning aims to minimize the number of segmented examples by automatically selecting the examples that are most informative for the story segmenter. By this method, we can decrease the labor effort in boring story segmentation and get the same or better performance in automatic story segmentation. We also consider another problem, namely whether training data in story segmentation can be reusable or not and how the different special structures of the language influence the performance of active learning for story segmentation. The experimental results show that active learning can reduce the number of segmented examples to reach a given level of performance and that reusable training data in story segmentation still contain information that improves the performance. It is a satisfactory result