SBIR Phase I: ScenarioNet: Customizable Deep Information Extraction

这个来自nafware公司的小企业创新研究第一阶段项目解决了对更好的信息搜索方法的需求。计算语言学（事件语义）的最新进展通过实现场景网的开发，为美国知识工作者提供了提高生产力的机会，场景网可以自动学习从文本中提取深度事件信息。当前信息提取（IE）系统的一个主要缺陷是训练样本必须由人类专家精心注释。在ScenarioNet中，示例被自动提取、排序，并在快速的用户反馈周期中保留或丢弃，极大地简化了针对新领域或修订领域的系统训练。为了实现更深层次的事件信息提取，ScenarioNet集成了一个统计完整解析器、事件模型、事件构建器（带子事件合并）、关系模型（从单个短语或句子中提取多个关系）、场景模型和场景构建器（带事件合并算法）。对于深度IE，事件本身必须放在相关事件的上下文中：场景网的分层事件和场景模型包括因果关系、时间关系和结构关系的表示。事件合并算法利用共引用解决技术，包括话语级和跨文档共引用，来识别事件的多个元素以及事件如何组合成场景。场景网简化了跨域的可移植性，支持深度事件、关系和场景信息提取。一个健壮的、易于定制的事件、关系和场景信息提取系统在国防、情报、保险（审查应用程序和索赔）、医疗保健、金融服务、法律服务、商业情报收集、各级政府（审查应用程序和报告）和工程（跟上新的发展）等行业中具有强大的商业潜力。nafware提供的技术，场景网，通过减少定制工作（跨域可移植性）和将更深层次的语义整合到其面向对象的提取模型和模板中，消除了IE广泛商业化的障碍。

This Small Business Innovation Research Phase I project from Naftware, Inc. addresses the need for better information search methods. Recent advances in computational linguistics (event semantics) present opportunities for major productivity enhancements for American knowledge workers by enabling the development of ScenarioNet, which automatically learns to extract deep event information from text. A major deficiency in current information extraction (IE) systems is that training examples must be painstakingly annotated by human experts. In ScenarioNet, examples are automatically extracted, ranked, and kept or discarded in rapid user-feedback cycles, dramatically streamlining training the system for new or revised domains. To enable a 'deeper' level of event information extraction, ScenarioNet incorporates a statistical full parser, event models, event builder (with sub-event merging), relationship models (extracting multiple relationships from a single phrase or sentence), scenario models, and scenario builder (with event merging algorithms). For deep IE, events themselves must be placed in context of related events: ScenarioNet's hierarchical event and scenario models include representations of causal, temporal, and structural relationships. Event merging algorithms utilize coreference resolution techniques, including discourse-level and cross-document coreference, to recognize multiple elements of events and how events combine into scenarios. ScenarioNet eases cross-domain portability and enables deep event, relationship, and scenario information extraction.A robust, easily customizable event, relationship, and scenario information extraction system has strong commercial potential in such industries as defense, intelligence, insurance (review of applications and claims), healthcare, financial services, legal services, business intelligence gathering, all levels of government (review of applications and reports) and engineering (to keep up with new developments). Naftware's proffered technology, ScenarioNet, removes the barriers to wide commercialization of IE by cutting the customization effort (cross-domain portability) and by incorporating deeper semantics into its object-oriented extraction models and templates.