Document summarization based on semantic representations

We present a novel extractive summarization method based on semantic vector representation. The new representation extends the word embedding, and represents words, phrases, sentences, paragraphs and documents in same vector space, which is used to measure the semantic similarity between sentences and document. Then we use greedy search algorithm to extract the summary sentences. The proposed method is evaluated on DUC01 dataset and employ F1-measure and ROUGE-N as metrics. The results show the proposed method outperforms comparison methods. The ROUGE-N is much higher than others.