基于三元组信息指导的生成式文本摘要研究  

Research on abstractive text summarization based on triplet information guidance

作　　者：张云佐 李怡 ZHANG Yunzuo;LI Yi(School of Information Science and Technology,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;Key Laboratory of Hebei Province on Electromagnetic Environmental Effects and Information Processing,Shijiazhuang 050043,China)

机构地区：[1]石家庄铁道大学信息科学与技术学院,石家庄050043 [2]河北省电磁环境效应与信息处理重点实验室,石家庄050043

出　　处：《北京航空航天大学学报》2024年第12期3677-3685,共9页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(61702347,62027801);河北省自然科学基金(F2022210007,F2017210161);河北省高等学校科学技术研究项目(ZD2022100);中央引导地方科技发展资金(226Z0501G);石家庄铁道大学在读研究生创新能力培养资助项目(YC2022058)。

摘　　要：针对当前生成式文本摘要模型在解码时对文本事实性信息利用不充分的问题,提出一种以事实三元组为指导的文本摘要模型SPOATS。该模型基于Transformer结构搭建具有事实提取能力的双编码器和融合事实特征的解码器。构建LTP-BiLSTM-GAT (LBiG)模型,并设计最优事实三元组选择算法,从非结构化中文文本中提取最优事实三元组,并获取事实性信息的特征表示;利用改进的S-BERT模型对原文进行句子级向量表示,获取语义丰富的句子编码;设计基于注意力的事实融合机制,融合双编码特征来提高模型在解码阶段对事实性信息的选择能力。实验结果表明:在LCSTS数据集上,所提模型相比于基线模型ERPG的R1值提升了2.0%,摘要质量得到明显提升。This study introduces the text summarizing model SPOATS, which is driven by fact triples, to solve the issue that the current abstractive text summarization models do not entirely exploit the factual information of text in decoding. The model is based on a Transformer structure, which contains a double encoder capable of extracting facts and a decoder for combining factual features. To begin, the LTP-BiLSTM-GAT (LBiG) model is built and paired with the optimal factual triple selection technique suggested in this paper. The optimal factual triples are then extracted from the unstructured Chinese text to acquire the feature encoding of factual information. Then, the improved S-BERT model is used to represent the original text at the sentence-level vector to obtain the semantic rich sentence encoding. Finally, an attention-based fact fusion mechanism is designed to fuse the dual-encoding features, which can improve the ability of the model to select factual information in the decoding stage. The experimental results show that the proposed model has improved the value of ERPG by 2.0% compared to the baseline model on the dataset LCSTS, and the summary quality has been significantly improved.

关 键 词：事实融合机制 三元组 事实一致性 TRANSFORMER 生成式文本摘要 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]