Abstract
Extractive document summary is usually seen as a sequence labeling task, which the summary is formulated by sentences from the original document. However, the selected sentences usually are high redundancy in semantic space, so that the composed summary are high semantic redundancy. To alleviate this problem, we propose a model to reduce the semantic redundancy of summary by introducing the cluster algorithm to select difference sentences in semantic space and we improve the base BERT to score sentences. We evaluate our model and perform significance testing using ROUGE on the CNN/DailyMail datasets compare with six baselines, which include two traditional methods and four state-of-art deep learning model. The results validate the effectiveness of our approach, which leverages K-means algorithm to produce more accurate and less repeat sentences in semantic summaries.