News Text Summarization Method based on BART-TextRank Model

Abstract

With the rapid development of computer technology and network technology, the Internet has become the main place for people to release and obtain information. The amount of data on the network is growing explosively. The urgent problem is how to accurately obtain the real valuable information from a large amount of data. News report has always been the most important way for us to understand society and current affairs. Summarizing the main content of each news report in short language can help readers understand the content faster and save time. At present, automatic text summarization technology is used to generate news content summarization, which mainly includes extractive summarization and abstractive summarization. In this paper, we propose an improved strategy to solve the problem of topic deviation in abstractive summarization method. We combine TextRank with BART model. First, we use TextRank and BART to extract and generate summarization from news text. Then we splice the results of two methods to get the new text which improves the weight of the key sentences in the articles and make it more thematic. Finally, we input the above new texts enter the BART model again to get the final summarization. Experimental results show that compared with single BART model, the average recall scores of Rouge-1, Rouge-2 and Rouge-L are improved by 1.5%, 0.5% and 1.3% respectively.