【作者】 刘越；

【导师】 黄少滨；

【作者基本信息】 哈尔滨工程大学 ， 计算机科学与技术， 2018， 硕士

【摘要】 经济责任审计是对党政主要领导干部和国有控股企业领导人员在任职期间相关经济活动进行的审计,经济责任审计对腐败治理起着重要作用。随着国家逐步开展轮审和同步审计,被审计人员对经济责任审计相关知识获取的需求日益强烈。然而经济责任审计相关知识分散,通过搜索引擎检索相关知识需要人工对知识进行审核和筛选,获取困难。在这种情况下,本文针对经济责任审计这一领域构建领域知识问答系统,方便被审计人员获取相关知识,减少腐败发生。本文完成了领域问答库构建和领域知识问答答案检索。在问答库构建部分利用网络爬虫技术从互联网百科全书、相关机构网站和问答网站获取知识,并通过改进基于逻辑回归的领域词判定算法来保证抓取结果的可用性和合理性,完成了领域词库和FAQ库的构建。答案检索部分设计了多维度问句纠错和基于卷积神经网络的答案检索。其中问句纠错运用贝叶斯方法和N-Gram语言模型从拼写、词序维度对问句进行纠错,在实现过程中针对编辑距离的计算方法从拼音音节角度入手进行了优化。答案检索算法主要基于卷积神经网络来展开工作,设计了包含输入层、卷积层和池化层的神经网络结构,在抽取词汇特征的过程中创新性地提出了词汇网络权重技术,从词汇引用关系强弱入手来丰富模型特征。最后,通过对比实验验证了本文所用算法和模型的准确性和有效性,说明本文所构建的知识问答系统能够提供可靠的经济责任审计领域知识问答服务。更多还原

【Abstract】 Economic responsibility audit is an audit of the relevant economic activities of the leading cadres of the party and government and the leaders of the state-controlled enterprises during their tenure of office.Auditing economic responsibility plays an important role in the governance of corruption.As the round of trial and simultaneous audit are gradually carried out,the auditors have an increasingly strong demand for knowledge acquisition related to economic responsibility audit.However,knowledge related to economic responsibility audit is fragmented.Searching related knowledge through search engine requires manual examination and screening of knowledge and is difficult to obtain.Under such circumstances,this paper builds a domain question answering system in the field of economic responsibility audit,which can help the people to obtain relevant knowledge and reduce corruption.This paper completed domain knowledge base construction and answer retrieval.In the knowledge base building part of the use of web crawler technology from the Internet encyclopedia,related agencies and Knowledge quiz website access to knowledge,and by improving the algorithm of domain word decision based on logistic regression to ensure the rationality of the crawling results,the final construction of domain thesaurus and FAQ library was completed.In the part of answer retrieval,multidimensional question error correction and convolution neural network neural network based answer retrieval are designed.Among them,error correction is made by using Bayesian method and N-Gram language model to correct the question from spelling and word order dimension.The calculation method of edit distance is optimized from the point of phonetic syllables.Based on convolution neural network,the answer retrieval algorithm designs the neural network structure including input layer,convolution layer and pooling layer,In the process of feature extraction,this paper proposes a new vocabulary weighting technique to enrich the model features.In the end,the accuracy and validity of the algorithms and models used in this paper was verified through comparative experiments and can provide the question answer ability in the field of economic responsibility audit.更多还原