基于混合深度学习算法的工程规范知识抽取

Construction specification knowledge extraction method based on hybrid deep learning algorithm

摘要

工程规范是工程建设过程中常用的重要标准文件之一。面对这些非结构化工程规范文本，高效、准确地从中抽取相关知识，并将这些知识以可视化形式呈现，对于提高知识的利用效率、提升管理人员对工程规范文本的理解效率有着重要的作用。针对典型的工程规范文本，提出一种基于深度学习的工程规范知识抽取方法，融合ALBERT(A Lite Bidirectional Encoder Representation from Transformers)、BiLSTM(Bi-directional Long Shot-Term Memory)和CRF(Conditional Random Fields),建立工程规范实体识别模型，增强文本语义特征，获得工程规范中的实体；融合Attention机制和BiLSTM提取工程规范中的关系，根据所提取出的知识构建工程规范知识图谱。以《给水排水管道工程施工及验收规范》为典型实例对该方法进行了验证，结果表明，工程规范实体识别的F1值为78.18%,优于传统模型；关系抽取的F1值为98.35%。利用所抽取知识建立了工程规范知识图谱，通过基于知识图谱的全局信息展示、特定信息检索，提升工程规范的利用效率，辅助工程现场施工。

Abstract

Engineering specifications are one of the important standard documents commonly used in the construction process. Faced with these unstructured engineering specification texts, efficiently and accurately extracting relevant knowledge and presenting this knowledge in a visual format plays a significant role in improving knowledge utilization efficiency and enhancing management personnel′s understanding of engineering specification texts. A deep learning-based method was proposed for extracting knowledge from typical engineering specification texts, integrating ALBERT(A Lite Bidirectional Encoder Representation from Transformers), BiLSTM(Bi-directional Long Short-Term Memory), and CRF(Conditional Random Fields) to establish an entity recognition model for engineering specifications. The model enhances the semantic features of the text to identify entities within the engineering specifications. Additionally, it employs the Attention mechanism and BiLSTM to extract relationships from the engineering specifications and constructs an engineering specification knowledge graph based on the extracted knowledge. Using the “Construction and Acceptance Specifications for Water Supply and Drainage Pipeline Projects” as a typical example, the method was validated, yielding an F1 score of 78.18% for entity recognition, which is superior to traditional models, and an F1 score of 98.35% for relationship extraction. Leveraging this knowledge, an engineering specification knowledge graph was established. Through a knowledge graph-based global information display, specific information retrieval, the efficiency of utilizing engineering specification knowledge was improved, assisting with on-site construction.

关键词

工程规范 / 知识抽取 / ALBERT预训练模型 / BiLSTM / CRF / 注意力机制

Key words

engineering specification / knowledge extraction / ALBERT pre-training model / BiLSTM / CRF / attention

Author summay

邓旭方(1987—), 男, 工程师, 水工金属结构主任师, 学士, 主要从事水工建筑物运维和大坝安全管理领域的技术研究工作。Email: [e-mail]

基金资助

中国长江电力股份有限公司科研项目(Z212302036)