SBIR Phase II: Authoring Assistance via Contextual Semantic Labeling

这个小企业创新研究（SBIR）第二阶段项目的更广泛的影响/商业潜力来自于从“暗数据”中提取有意义、有用、具体的信息。暗数据是公司产生和接收的无数文档，其中包含未使用的信息-通常是因为它们的格式计算机无法理解。这些文档中的许多甚至不包含可访问的文本：只有文本的图片。文字处理文档和电子邮件确实有文本，但没有关于文本是什么的信息。计算机可以很容易地分辨出“10/05/2022”是一个日期，但是需要知道这是一个特定协议的开始或结束日期（或其他日期）才能让它有用。该项目使用了一系列人工智能（AI）技术，这些技术在人们编写新文档或从旧文档中提取数据时可以在真实的时间内工作。人工智能从示例中快速学习，在类似的文档中找到模式，并使用这种学习来节省用户在不同的上下文中一次又一次地搜索项目。这节省了大量繁琐的工作，减少了错误。提取的信息有助于企业理解、分析和制定业务决策。这个小型企业创新研究（SBIR）第二阶段项目从长的自然语言文档中识别和提取有用的信息项，特别是合同和协议。该技术比典型的提取方法更具体地识别项目;例如，不仅作为人，组织或地名，而且每个角色都扮演着什么角色。同样，地址、日期、金额和其他数据项只有在您知道它们的用途时才有用。这是推进自然语言理解的一个有价值的焦点。该团队联合收割机并扩展了机器学习技术，如少量学习，微调和语义解析，以实现这些更强大，更“语义”的结果。该解决方案使公司能够从他们已经收集但尚未自动化或利用的大量信息中产生价值。该奖项反映了NSF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而被认为值得支持。

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase II project comes from extracting meaningful, useful, specific information from “dark data.” Dark data are the countless documents companies produce and receive, which contain unused information – usually because they are in formats that computers do not understand. Many of these documents do not even contain accessible text: only pictures of text. Word-processing documents and emails do have text, but no information about what the text is. Computers can easily tell that “10/05/2022” is a date, but knowing it is the date a particular agreement starts or ends (or something else) is needed to make it useful. This project uses a range of artificial intelligence (AI) techniques that work in real time while people are writing new documents or extracting data from old documents. The AI learns quickly from examples, finds patterns across similar documents, and uses that learning to save the user from having to search for items again and again in varying contexts. This saves a lot of tedious work and reduces errors. The extracted information helps companies understand, analyze, and make business decisions.This Small Business Innovation Research (SBIR) Phase II project identifies and extracts useful information items from long natural language documents, especially contracts and agreements. The technology identifies items much more specifically than typical extraction methods; for example, not only as person, organization, or place names, but as to what role each plays. Likewise, addresses, dates, money amounts, and other data items only become useful when you know what they’re for. This is a valuable focus for advancing Natural Language Understanding. The team combine and extend Machine Learning technologies such as few-shot learning, fine-tuning, and semantic parsing to achieve these stronger, more “semantic” results. This solution allows companies to generate value from huge troves of information they already collect but cannot yet automate or leverage.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.