SBIR Phase I: Semantic Annotation of Hypertext and its Application in Event Mapping

这个小企业创新研究（SBIR）第一阶段项目的更广泛影响将是帮助增强现实，基于位置的搜索和世界实时实时地图的应用，并建立地球的数字孪生模型。事件信息是动态的和超本地的。因此，将事件的位置和时间与真实的GPS坐标和世界时联系起来是很重要的。这带来了实体消歧和解析的技术挑战，以及实时处理事件信息以反映最后一刻的更改。自然语言基础工作也有助于创建一个镜像世界，虚拟互联网与物理空间和事件之间的信息相互关联。这个小企业创新研究（SBIR）第一阶段项目将是第一个在网络规模上研究半结构化超文本神经语言模型预训练的项目（每月55 TB的压缩数据）。这将大大加快对嘈杂Web文本的理解，而迄今为止的大多数研究都是在干净的纯文本上进行的。该项目还将以半监督的方式使用文档级注释来攻击机器阅读的挑战，而研究的优势已经以完全监督的方式使用更精确的单词级注释进行。该项目的技术目标是创建一个可扩展的基础设施，允许快速迭代挖掘Web规模的数据，以及适应于通过超文本学习结构化信息的算法集合。对于大多数小企业来说，这是一个未解决的挑战，过去只有互联网巨头尝试过。由此产生的机器学习算法将能够解释半结构化的Web数据，而不是典型的结构化注释来理解超文本。该奖项反映了NSF的法定使命，并被认为值得通过使用基金会的智力价值和更广泛的影响审查标准进行评估来支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project will be to help applications of augmented reality, location-based search, and real-time live mapping of the world, and to build a digital twin of the earth. Event information is dynamic and hyperlocal. Thus it is important to ground the location and time of events to real GPS coordinates and universal time. This creates technical challenges of entity disambiguation and resolution, and real-time processing event information to reflect last-minute changes. The proposed natural language grounding work also helps create a mirrorworld with information interlinked between the virtual internet and physical spaces and events.This Small Business Innovation Research (SBIR) Phase I project will be the first to investigate neural language model pre-training over semi-structured hypertext on a web scale (55TB of compressed data monthly). This will greatly accelerate understanding of noisy web text, while the majority of research to date has been conducted on clean plain text only. This project will also attack the challenge of machine reading with document-level annotations in a semi-supervised fashion, while the predominance of study has been carried out with more precise word-level annotation in a fully supervised way. The technical goal of this project is to create a scalable infrastructure that allows quick iterations of mining web scale data, and an ensemble of algorithms that are adapted to learning structured information over hypertext. It is an unsolved challenge for most small businesses that in the past only the internet giants have attempted. The resulting machine learning algorithms will be capable of interpreting semi-structured web data, in contrast to typical structured annotation to understand hypertext.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.