摘要:In this paper, we study the task of template building in automatically generate NBA match reports from NBA live text. As a preliminary study, we collect and process the historical reports compiled by the editors and get different kinds of sentences. Our innovative proposal is to divide the NBA match reports into 11 categories, which covering almost all cases. We use different machine learning methods to classify sentences. Each class finally constructs a template library to service the next automatic writing. By comparing different methods, we get a higher accuracy classification structure. The evaluation results show that our method does construct a template library.