【计算机科学】【2018.05】深度学习在审计决策中的应用

在这里插入图片描述

本文为美国新泽西州立大学罗格斯分校（作者：TING SUN）的博士论文，共172页。

本论文的目的是探讨商业沟通文件或社交媒体资讯中的情绪特徵是否能提供相关且可靠的资讯给稽核人员。本文首先考察了盈余会议情绪特征对内部控制重大缺陷预测的增量信息。借助IBM Watson提供的深度学习文本分析工具Alchemy Language API，本文通过这些情感特征与以往文献提出的ICMW的其他决定因素（即Doyle、Ge和McVay，2007a；AshbaughSkaife、Collins和Kinney，2007）获得了文本的总体情感得分和情感“喜悦”置信度得分。结果表明，这些情感特征特别是喜悦分数提高了模型的解释能力和预测精度。

第二部分将深度学习与“词袋”方法进行了比较，论证了基于深度学习的情绪分析在财务错报预测背景下10-K文件MD&A部分的有效性。研究结果包括：（1）情绪特征为财务错报预测提供了有用的信息，主要用于欺诈检测；（2）使用基于深度学习的情绪特征模型通常比使用“词袋”方法提取的情绪特征模型更有效。

第三部分探讨了客户公司微博信息与审计费用的关系。该研究考察了2015年美国上市公司审计费用与客户公司推特性质之间的关系：推特的情绪、推特的数量和推特的流行程度。所有推特信息都是通过IBM Twitter Insights获得的，这是一个Twitter数据分析工具，它依靠深度学习算法提供情感和其他丰富内容。研究发现，对于未发表持续经营审计意见的公司和重述风险处于中等水平的公司，审计费用与负面推文的频率正相关，并且收到较多转发的公司比收到较少转发的公司的这种关联性得到加强。

The objective of this dissertation is toinvestigate whether the sentiment features of business communication documentsor social media information extracted by deep learning techniques deliverrelevant and reliable information to auditors. The first essay investigates theincremental informativeness of sentiment features of earnings conference callsfor the prediction of internal control material weaknesses (ICMW). With thehelp of a deep learning textual analyzer provided by IBM Watson, AlchemyLanguage API, this essay obtains the overall sentiment score of the text andthe confidence score of the emotion “joy.” These sentiment features are thenused as additional predictors along with other determinants of ICMW suggestedby prior literature (i.e., Doyle, Ge, and McVay, 2007a; Ashbaugh-Skaife,Collins, and Kinney, 2007). The results indicate that these sentiment features,especially the score of joy, improve the explanatory ability and the predictionaccuracy of the model. The second essay compares deep learning to the “bag ofwords” approach and demonstrates the effectiveness and efficiency of deeplearning-based sentiment analysis for MD&A sections of 10-K filings in thecontext of financial misstatement prediction. The findings include (1)sentiment features provide insights for financial misstatement prediction,primarily for fraud detection; (2) the model using deep learning-basedsentiment features generally performs more effectively than the model usingsentiment features extracted by the “bag of words” approach. The third essayexamines how the information of tweeting activities about the client company isassociated with the audit fee. It examines the relationship between the auditfee of U.S. public firms in 2015 and the properties of tweets about the clientfirm: the sentiment of tweets, the volume of tweets, and the popularity oftweets. All tweet information is obtained using IBM Twitter Insights, a Twitterdata analysis tool that provides sentiment and other enrichments relying ondeep learning algorithms. It finds that for companies without going-concernaudit opinions and companies with a median level of restatement risk, the auditfee is positively associated with the frequency of negative tweets, and thisassociation is strengthened for companies receiving more retweets than thosereceiving less retweets.