摘要：

Abstract:

【Purpose/significance】To extract a summary with higher relevance and less redundancy, this article proposes an unsuper⁃

vised extractive text summarization model combined with deep learning.【Method/process】Based on the Maximal Marginal Relevance

model, we calculate the similarity between sentences by the word embedding as well as the sentence embedding, rank sentences by

keywords and location information on the importance of sentences to obtain higher-quality summary. The model is applied to the 2018

Byte Cup of automatic generating the article title task to conduct experiments to test the effectiveness.【Result/conclusion】The experi⁃

mental results show that the ROUGE-L value of extracting single sentence summary is 28.24%, which is higher than that of traditional

extracting text summary algorithms, i.e. CI (17.37% ), TextRank (22.70% ) and MR (23.52% ). The Rouge-L value of extracting

multi-sentence summaries is 37.78% , which is higher than that of traditional extracting text summarization algorithms, i.e. CI

(29.35%), TextRank (34.15%) and MR (31.09%). The experimental results prove that leveraging deep learning can improve the quali⁃

ty of text summary.【Innovation/limitation】The innovation is that we combine the maximal marginal relevance (MMR) with deep learn⁃

ing, and comprehensively consider the three aspects of sentence features to obtain extractive summary. The limitation is that the scope

of our research is limited to extractive text summarization, and the abstractive text summarization model will be integrated into our

model in further research.