ORAC

ORAC项目的目的是研究使用机器学习，特别是神经网络，技术来显着提高支出分析系统中数据分类的准确性。支出分析通过提供详细的全方位的采购支出洞察，能够有效地降低成本和管理供应商。目前，此类系统中自动数据分类的最先进水平是50%-70%的准确率（不限于机器学习方法，但通常使用多种技术的组合），需要相当多的人为干预才能达到100%（这假设分类能够解决广泛的市场领域）。初步研究表明，使用神经网络对数据进行预处理可以将自动准确率提高到95%。该项目的目标是：1)确定可采用的商业神经网络技术（b谷歌的技术是目前的首选解决方案）；2)定制所选择的神经网络（NN），以支持用于支出分析的特定文本结构；3)创建一个概念验证演示器，用于表征新分类系统的功能（新神经网络与当前的Spend360解决方案相结合）；4)评估基于神经网络的解决方案的能力和约束，并确定需要作为原型评估活动的一部分来解决的问题。从这种新的创新方法中获得的主要好处是，它将大大减少人类对分类过程和相关数据质量保证的参与。这将大大减少完成分类的时间，提高准确率。总之，这将降低解决方案的成本，从而使发现当前服务成本过高的一系列组织能够采用该解决方案。

The aim of the ORAC project is to research the use of machine learning, specifically neuralnetworks, techniques to significantly improve the accuracy of data classification within aspend analysis system. Spend analysis enables effective cost reduction and suppliermanagement by providing detailed all-round insight into procurement spend. Currently thestate-of-the art for automated data classification in such systems is 50%-70% accuracy (notlimited to machine learning approaches but typically using a combination of techniques) withconsiderable human intervention required to achieve 100% (this assumes the classification iscapable of addressing a wide range of market sectors). Initial investigation has indicated thatthe use of neural networks to pre-process the data could increase the automated accuracy to95%.The objectives of this project are to: 1) identify the commercial neural network technologiesthat can be adopted (Google’s technology being the preferred solution at present); 2) tailor theselected Neural Network (NN) to support the specific text structures used in spend analysis; 3)create a proof-of-concept demonstrator that will be used to characterize the capabilities of thenew categorization system (the new NN combined with the current Spend360 solution); 4)evaluate the capabilities and constraints of the NN-based solution and identify the issues thatneed to be addressed as part of a prototype evaluation activity.The primary benefit that will accrue from this new innovative approach is that it willsignificantly reduce the human involvement in the classification process and associated dataquality assurance. This will significantly reduce the time to complete the classification andimprove the accuracy. Together, this will result in cost reduction of the solution therebyenabling adoption by a range of organisations that find the current service cost too high.