计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (7): 163-170.DOI: 10.3778/j.issn.1002-8331.2208-0167

• 模式识别与人工智能 • 上一篇    下一篇

多注意力机制的文本摘要事实一致性评估模型

魏楚元,张鑫贤,王致远,李金哲,刘杰   

  1. 1.北京建筑大学 电气与信息工程学院,北京 102612
    2.北方工业大学 信息学院,北京 100144
  • 出版日期:2023-04-01 发布日期:2023-04-01

Factual Consistency Assessment Evaluation Model for Text Summarization Based on Multi-Attention Mechanism

WEI Chuyuan, ZHANG Xinxian, WANG Zhiyuan, LI Jinzhe, LIU Jie   

  1. 1.School of Electrical and Information Engineering, Beijing University of Civil Engineering and Architecture, Beijing 102612, China
    2.School of Information Science, North China University of Technology, Beijing 100144, China
  • Online:2023-04-01 Published:2023-04-01

摘要: 文本摘要事实一致性是摘要内容与源文档内容的信息一致。最近的研究表明,文本摘要模型生成的摘要存在较多与原文事实不一致的问题,设计能够检测并评估出事实不一致错误的方法至关重要。目前基于自然语言推理的方法存在对源文档内容提取简单,推理信息交互不充分等问题。提出多注意力机制的文本摘要事实一致性评估模型,利用预训练模型微调的sentence-BERT模型挑选源文档中的关键句,然后将摘要句与关键句组合成句子对,输入BERT模型编码获得向量表示结合ESIM进行句子对的推理,利用图注意力网络完成推理信息的聚合,提高文本摘要事实一致性评估模型的准确率。实验结果表明,该算法与多个典型算法在在领域内常用的数据集进行实验比较,其可行性和有效性得到验证。

关键词: 文本摘要, 事实一致性, BERT, 图注意力网络, 交互注意力

Abstract: Factual consistency of text summarization is consistent with information of the source document and the summarization. Recent researches have shown that there are large number of factual inconsistencies existed in the outputs of abstractive summarization models. It is important to design a method that can detect and evaluate the error of fact inconsistency. Most of existing methods based on natural language inference have insufficient ability to extract key content of source document and infer the information of content. This paper improves the accuracy of consistency assessment model for summarization by multi-attention mechanism. Firstly, the key sentences are selected according to sentence-BERT, which is fine-turning on pre-trained language model. Each evidence and claim are formed into separate sentence pairs and sent to the BERT encoder to obtain the vectors of the sentence pairs representation, then it achieves evidence reasoning by the vectors combined with ESIM. Finally, the graph attention network is used to complete the aggregation of inferential information to obtain the fact consistency assessment result. Experimental results show that this algorithm is compared with several typical algorithms in common datasets in the field, and its feasibility and effectiveness are verified.

Key words: text summarization, factual consistency, BERT, graph attention network, interactive attention