面向招标文件的半结构化文本自动生成

Automatic Generation of Semi-Structured Texts for Bidding Documents

招标文件范本总结了历史招标文件的核心内容。针对现有文本摘要方法无法有效提取文本主题结构、忽略时间特征对文本重要性的影响等问题,提出一种基于多头图注意力网络的半结构化文本自动生成(SGMG)模型。在模型预处理层中,利用BERT预训练模型进行文本嵌入,通过白化操作增强文本向量的表达能力。在主题抽取模块中,利用异质图建立主题、段落及文档之间的语义关系,应用多头图注意力网络加强图节点之间的交互,提高模型学习能力。在中心句抽取模块中,通过融合时间特征及语义相似关系的有向图计算文本中心性,防止重要时间信息的丢失。在句子选择模块中,结合三元词过滤法,提升生成内容的多样性。将国家能源集团2016年至2020年的招标文件作为语料进行实验,结果表明SGMG模型能够有效提取文本主题结构,生成文本内容与人工编制的范本接近,ROUGE-L评估指标相比于TextRank、LexRank等现有文本摘要方法提升了4.3个百分点以上。

The bidding document template summarizes the core content of historical bidding documents.As current text abstraction methods can not effectively extract text topic structures and ignore the impact of temporal features on the importance of text,we propose a Semi-structured automatic text Generation model based on Multi-head Graph attention network(SGMG).In the model preprocessing layer,the pre-trained model Bidirectional Encoder Representations from Transformers(BERT)are used for embedding text and enhancing the expressions of the text vector through a whitening operation.In the topic extraction module,a heterogeneous graph is used to establish semantic relationships between topics,paragraphs,and documents.Graph nodes interact through a multi-head graph attention network that improves the feature-learning ability of the model.In the central sentence extraction module,text centrality is calculated by fusing a directed graph of time features and semantic similarity.In the sentence selection module,the Trigram Blocking method is used to improve the diversity of the generated content.After using the bidding documents accumulated by CHN ENERGY from 2016 to 2020 as the corpus,the experimental results show that the proposed SGMG model can effectively extract the text topic structure and that the generated text content is similar to that of a manual template.Under the Recall-Oriented Understudy for Gisting Evaluation Longest common subsequence(ROUGE-L)evaluation index,the text abstraction generated by SGMG model performs 4.3 percentage points better than existing methods such as TextRank and LexRank.