基于深度学习的建筑规范问答系统的研究与实现

黄子韦

摘 要

我国已经形成了较为完整的建筑标准体系，该体系能对建设项目的设计、施工、验收等各个阶段进行规范。目前，建筑从业者使用建筑标准的方法主要是通过阅读查找，或者从搜索引擎中搜索，但是传统的搜索引擎并不能快速、准确的让用户获得自己想要的答案，因此，需要为建筑从业者提供一个专业的系统，该系统可以快速、方便、准确的回答用户提出的问题。

本文以建筑工程质量验收规范为数据源，提出了一个面向建筑工程质量验收规范的问答系统方法，该系统集成了信息检索与自然语言处理的深度学习模型，主要做了以下几方面的工作：

1、根据建筑工程规范类文件专业性强的特点，构建了建筑工程质量验收规范知识库，这是建筑工程规范类文件在问答系统研究方面的首次尝试。

2、为了能够更好构建面向建筑领域的问答系统，将BERT模型用于建筑工程质量验收规范问答系统，将基于BERT模型与传统的TF-IDF算法相结合实现问答系统的构建，与传统的模板驱动的问答系统不同，本文将重点放在文本文档和可扩展性较低的小型语料库上，所提出的方法具有强大的特征表示能力和学习能力，可以解决用户提出的建筑工程施工验收问题。

3、将BERT模型与传统的TF-IDF算法相结合，实现了建筑工程质量验收规范自动问答系统。同时，开发了原型系统聊天机器人，并将聊天机器人接入微信公众号，使建筑从业者能方便、高效地学习建筑领域相关知识。

通过系统测试结果可知，本文提出的基于深度学习的建筑工程质量验收规范问答系统效果较好。作为工程规范类文件在问答系统上的首次尝试，该系统减少对人工的依赖，自动化程度较高，有较强的通用性强，值得做进一步研究。

关键词：质量验收规范；问答系统；深度学习；BERT；TF-IDF

Abstract

China has formed a relatively complete building standard system, which can regulate the design, construction, acceptance and other stages of construction projects. At present, construction practitioners mainly look for construction articles by reading, or search for construction codes from search engines, but traditional search engines cannot quickly and accurately allow users to obtain the answers they want. Therefore, it is necessary to provide construction practitioners with a professional system that can quickly, conveniently and accurately answer questions raised by users.

This paper takes the construction engineering quality acceptance code as the data source, and proposes question answering system method for the construction engineering quality acceptance code. This system integrates the deep learning model of information retrieval and natural language processing. It mainly does the following work:

1. According to the professional characteristics of construction engineering specification documents, a knowledge base of construction engineering quality acceptance codes is constructed. This is the first attempt of the construction engineering specification documents in the research of question answering system.

2. In order to better build a question answering system for the construction field, the BERT model is used in the question answering system for the construction engineering quality acceptance code, and the construction of the question answering system based on the BERT model and the traditional TF-IDF algorithm is combined with the traditional template. The driven question answering system is different. This article focuses on text documents and small corpora with low expandability. The proposed method has strong feature representation and learning capabilities, which can solve the construction acceptance problems raised by users.

3. Combining the BERT model with the traditional TF-IDF algorithm, an automatic question answering system for construction engineering quality acceptance code is realized. At the same time, a prototype system chat robot was developed, and the chat robot was connected to the WeChat public account, so that construction practitioners can learn the relevant knowledge in the construction field conveniently and efficiently.

According to the system test results, the effect of the question answering system proposed in this paper on the quality acceptance codes for construction engineering based on deep learning meets the requirements. As the first attempt of the engineering code file on the question answering system, this system reduces the dependence on labor, has a high degree of automation, and has strong versatility, which is worthy of further research.

Key words：Codes for Quality Acceptance; Question Answering; Deep Learning; BERT;TF-IDF