【作者】 张瑾；

【导师】 万里；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2019， 硕士

【摘要】 伴随着互联网和人工智能的飞速发展,现今已经进入“知识互联”的时代,对于知识图谱的关注逐渐升温。知识图谱根据其包含的知识范畴被细分为通用领域知识图谱和垂直领域知识图谱。本文重点聚焦于垂直领域知识图谱构建与答案排序研究,旨在将行业知识引入问答系统,改善问答系统的用户体验,为实现并优化基于垂直领域知识图谱的问答系统提供思路。本文的主要研究内容包括三个方面:(1)以“建筑安全”为例,研究垂直领域知识图谱构建方法。针对目前在建筑领域还没有一个开放的中文安全知识图谱,而目前使用广泛的都是百科类的通用知识图谱,本文提出一个构建建筑安全领域知识图谱的框架。首先,通过爬虫获取建筑领域专业术语。然后,采用CNN和RNN模型区分出建筑安全领域文本,使用Bi-LSTM+CRF模型完成常规实体抽取,通过人工干预结合近义词扩展的方式完成建筑专业实体的抽取。紧接着使用依存句法分析完成关系抽取。最后,将拼接好的三元组数据导入开源图数据库Neo4j中,完成建筑安全领域知识图谱的构建工作,可将其作为知识库模块应用于问答系统中。(2)针对实体链接和关系链接独立处理会丢失信息且效率不高的问题,提出基于连接密度的实体关系联合链接算法。该算法引入连接密度的概念,在已有实体关系候选项的条件下,计算出每个候选项成为最优候选项的概率,从而解决实体关系链接问题。(3)为了优化答案排序效果,改善用户体验,本文提出引入主观评价的答案排序学习算法。首先,通过构造答案对提供给众包人员标注,将人类用户的主观判断引入答案排序学习算法中,然后,使用词向量进行答案表示,将排序学习算法转化为一个二分类问题,即判断两个答案的质量优劣,最后,依据质量得分完成答案排序。该算法对比了经典的pairwise排序学习算法RankNet,经实验证明引入人类的主观评价提升了排序的效果。更多还原

【Abstract】 With the rapid development of the Internet and artificial intelligence,it has entered the era of "knowledge interconnection",and the attention of the knowledge graph is gradually heating up.The knowledge graph is subdivided into a general domain knowledge graph and a vertical domain knowledge graph according to the knowledge categories it contains.This thesis focuses on the vertical domain knowledge graph construction and answer ranking research,aiming to introduce industry knowledge into the question and answer system,improve the user experience of the question and answer system,and provide ideas for implementing and optimizing the question and answer system based on the vertical domain knowledge graph.The main research content of this thesis includes three aspects:(1)Take “construction security” as an example to study the construction method of vertical domain knowledge graph.In view of the fact that there is currently lack open Chinese security knowledge graph in the field of construction,and the widely used ones are the general knowledge graph of encyclopedias.This thesis proposes a framework for constructing knowledge graph in the field of construction security.First,the terminology of the construction field is obtained through reptiles.Second,the CNN and RNN models are used to distinguish the texts in the field of construction security,and the Bi-LSTM + CRF model is used to complete the conventional entity extraction.The artificial professional entities are extracted by means of manual intervention and synonym expansion.Third,the relationship extraction is completed using dependency syntax analysis.Finally,the spliced triple data is imported into the open source database Neo4 j to complete the construction of the knowledge graph in the field of construction security,it can be used as a module with knowledge in the question and answer system.(2)In order to solve the problem that the information is lost and the efficiency is not high in the entity linking and the relation linking when processed independently,a joint entity relation linking algorithm based on the connection density is proposed.The algorithm introduces the concept of connection density,and calculates the probability that each candidate as the optimal candidate under the condition of existing entity relationship candidates,which solving the entity relation linking problem.(3)In order to optimize the answer sorting effect and improve the user experience,this thesis proposes an answer sorting learning algorithm that introduces subjective evaluation.Firstly,by constructing the answer to the crowdsourcing personnel,the subjective judgment of the human user is introduced into the answer sorting learning algorithm,and then the word vector is used to represent the answer,and the sorting learning algorithm is transformed into a binary-classification problem.Finally,the answer is sorted according to the quality score.The algorithm compares the classic pairwise learning to rank algorithm RankNet,and the results show that putting the subjective evaluation into experiment could improve the ranking effect.更多还原