Developing machine reading comprehension methods to automate financial report audit

在过去的几年里，神经网络模型推动了NLP研究的边界，提高了各种任务的最先进性能，包括自然语言理解和机器阅读理解。自然语言处理的主要挑战之一是处理人类语言表达的丰富性和多样性。我们的工业合作伙伴（Ark Paradigm）的目标是开发一种自动验证公司财务报告完整性的方法。由于错误和遗漏，随着每年立法要求的增加，人工或算法验证信息的成本继续变得更加繁重，使该过程更加复杂，耗时和昂贵。为了改进和建立他们现有的产品，该公司试图打破当前机器阅读理解方法在速度、成本和质量之间的普遍权衡。在这个提议中，我们将自然语言理解问题作为一个机器阅读理解问题来处理。机器学习中的阅读理解包括训练一个模型，该模型能够处理文档，根据给定的文档和外部知识进行推断，并回答有关文档内容的复杂问题。该项目有望为机器阅读理解的研究做出新的贡献，特别是在金融领域。金融领域提供了独特的挑战，因此在现有阅读理解数据集上工作的方法可能不适用于金融领域。在这一领域工作良好的机器阅读理解方法的发展将是对这一研究领域以及金融行业的有益和创新的贡献。

Neural network models have pushed the boundaries of NLP research in the last few years, improving state-of-the-art performances on a variety of tasks, including natural language understanding and machine reading comprehension. One of the main challenges in Natural Language Processing is dealing with the richness and variety of expressions in human languages. The goal of our industrial partner (Ark Paradigm) is to develop a method for automatically verifying the completeness of the corporate financial reports. Due to errors and omissions, the cost of manually or algorithmically verifying information continues to become more burdensome as legislative requirements increase each year, making the process more complex, time consuming and costly. With an objective of improving and building upon their existing product, the company seeks to break prevailing trade-offs between speed, cost, and quality under current approaches to machine reading comprehension.In this proposal, we approach the problem of natural language understanding as a machine reading comprehension problem. Reading comprehension in Machine Learning involves training a model that would be capable of processing a document, making inferences based on the given document and external knowledge, and answering complex questions with respect to the document's content. The proposed project is expected to make novel contributions to the research on machine reading comprehension, specifically in the financial domain. The financial domain offers unique challenges, therefore the methods that work on the existing reading comprehension datasets may not work as well on the financial domain. The development of machine reading comprehension methods that work well in this domain will be a useful and innovative contribution to this research field, as well as the finance industry.