基于Transformer-TextRank-PGN的文本摘要模型
Text Summarization Model Based on Transformer-TextRank-PGN
吴广硕 1樊重俊 1陶国庆1
作者信息
- 1. 上海理工大学管理学院 上海 200093
- 折叠
摘要
针对文本摘要任务中存在编码器端对文本语义信息理解不足,解码器端生成单词不可控的问题,提出了一种Transformer-TextRank-PGN文本摘要模型,该模型同时保留了生成式摘要和抽取式摘要的优点.在模型编码器端引入Tex-tRank算法增强编码器学习文本语义信息的能力,解码器端引入指针网络指向原文中抽取单词,使用抽取单词概率分布和解码器生成单词的概率分布共同影响最终生成词,使模型可以复现出原文细节和生成OOV词汇.经过在NLPCC文本摘要数据集上的实验结果表明,该模型所生成摘要的准确性和可读性更接近于数据集中所给出的标准摘要.
Abstract
Aiming at the problem of insufficient understanding of text semantic information by the encoder and uncontrollable words generated by the decoder in the text summarization task,a Transformer-TextRank-PGN text summarization model is pro-posed,which retains the advantage of both generative summary and extractive summary.The TextRank algorithm is introduced on the encoder side of the model to enhance the ability of the encoder to learn the semantic information of the text.The decoder side in-troduces a pointer network to point to the extracted words from the original text.The probability distribution of the extracted words and the probability distribution of the words generated by the decoder together affect the final generated words.The model can repro-duce the original text details and generate OOV vocabulary.The experimental results on the NLPCC text summarization dataset show that the accuracy and readability of the summaries generated by the model are closer to the standard summaries given in the data set.
关键词
生成式摘要/抽取式摘要/Transformer/TextRank/指针网络Key words
generative summary/extractive summary/Transformer/TextRank/pointer network引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据