A Two-stage Long Text Summarization Method Based on Extraction-generation

Long text summarization aims to extract key information from lengthy texts and generate concise and accurate summaries. However, due to the complexity of long text information and the limitation of text length, existing summary methods encounter hurdles like semantic degradation and redundant information. To tackle these problems, this article proposes a two-stage long text summarization method based on extraction-generation. This method jointly trains an extractor and a generator, effectively combining the strengths of extractive and abstractive methods. The extractor is utilized to extract key information to tackle the challenge of handling long inputs, while the generator allows the use of dynamic sentence-level attention weights during decoding. This enables the generator to dynamically adjust the importance of sentences based on contextual information, leading to more precise summary generation. Extensive experiments conducted on the arXiv dataset demonstrate the superior performance of our proposed model in the task of long text summarization.