SBIR Phase I: Automatic Data Series Extraction from a Text Corpus

这项小企业创新研究（SBIR）第一阶段项目的广泛影响将是提高市场效率，促进金融服务和邻近行业的竞争。目前，这些部门由大公司主导，它们有资源创造和利用相对于小公司的不对称信息优势。造成这种不对称的一个关键原因是，公司信息——构建高质量、详细的财务模型的基础——可能过于昂贵，不是因为它不可获得，而是因为它以非标准化的方式报告，因此难以提取和实施。要在整个行业系统地进行这样的工作，每年需要花费数千个工时来手工筛选数百万份文件，这是只有世界上最大的公司才能承担的成本。自动化提取这类信息并使其广泛提供和容易获得，有助于为小公司创造公平的竞争环境，同时以较低的总成本提高大公司决策的速度和质量。这个小企业创新研究（SBIR）第一阶段项目旨在开发一个机器学习平台，用于从财务文件集合中自动提取数据。该平台将利用自然语言处理模型架构的最新进展，但仍然面临以下挑战：(a)即使文档文本量和语义变化不断增长，也要实现和保持高水平的准确性；(b)以经济有效的方式生成足够的标记训练数据。这个项目通过一个基于最新元学习技术的持续模型训练的新框架来解决这些双重挑战。这种监督学习的方法可以大大加快模型的改进，同时降低训练成本。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will be improved market efficiency and greater competition in financial services and adjacent industries. Currently, these sectors are dominated by large firms that have the resources to create and exploit asymmetric information advantages over smaller firms. A key reason for this asymmetry is that company information – the basis for building high-quality, detailed financial models – can be prohibitively expensive to surface, not because it is unavailable, but because it is reported in non-standardized ways and therefore difficult to extract and make actionable. To do so systematically and industry-wide requires thousands of man-hours per year of manually sifting through millions of documents, a cost that only the largest firms in the world can bear. Automating the extraction of such information and making it both widely available and easily accessible helps to level the playing field for small firms while simultaneously improving the speed and quality of decision-making – at a lower total cost – for large firms.This Small Business Innovation Research (SBIR) Phase I project aims to develop a machine learning platform for automatically extracting data from a collection of financial documents. The platform will take advantage of recent advances in natural language processing model architectures, but nonetheless faces the challenges of (a) achieving and maintaining a high level of accuracy, even as document text volume, and thus semantic variation, grows; and (b) generating sufficient labeled training data in a cost-effective way. This project addresses these dual challenges with a novel framework for continuous model training based on recent meta-learning techniques. Such an approach to supervised learning can substantially accelerate model improvement and simultaneously drive down training costs.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.