基于注意力机制的结构化文本自动生成  

Automatic generation of structured text based on attention mechanism

作　　者：王进强 刘金硕[2] WANG Jinqiang;LIU Jinshuo(China Shenhua International Engineering Company,Beijing 100007,China;School of Cyber Science and Engineering,Wuhan University,Wuhan 430072,China)

机构地区：[1]中国神华国际工程有限公司,北京100007 [2]武汉大学国家网络安全学院,湖北武汉430072

出　　处：《武汉大学学报（工学版）》2022年第2期198-203,共6页Engineering Journal of Wuhan University

摘　　要：为了解决手工编写招标文件耗时耗力、技术参数和要求覆盖不全面的问题,以及由于设备升级、行业规范修改所带来的范本自动迭代更新问题,提出了一种基于注意力机制Bert模型的招标技术范本自动生成方法。具体包括:建立标的物范本知识语料库,并使用正则表达式对招标文件进行数据清洗和技术参数提取;然后使用基于注意力机制的Bert模型提取参数文本的语义信息,映射至向量空间;最后通过TextRank算法得到生成的范本内容。以国家能源公司2016—2020年的161种标的物招标材料作为语料,将本方法生成的范本与专家人工编制范本作对比,结果表明,本方法获得了较高的Rouge评分与准确率,验证了其在真实生产环境下的可行性。In order to solve the problems of time-consuming as well as incomplete coverage of technical parameters and requirements in manual preparation of bidding documents, and automatic iterative updating of bidding technical templates caused by equipment upgrading and modification of industry specifications, an automatic generation method of bidding technical templates of Bert model based on attention mechanism was proposed. The specific contents include: establishing the subject object model knowledge corpus, using the regular expression to clean data and extract technical parameters of bidding documents;then, Bert model based on attention mechanism is used to extract the semantic information of parameter text and map it to vector space;and at last, TextRank algorithm is used to obtain the generated template contents. 161 bidding materials of National Energy Corporation from 2016 to 2020 were used as corpus, and the samples compiled by experts were compared. The results show that the Rouge score and accuracy of the proposed method are higher, and the feasibility of the proposed method in real production environment is verified.

关 键 词：结构化文本 注意力机制 正则表达式 自然语言处理 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]