III: Small: Collaborative Research: Scalable Schema-Based Event Extraction

当前语言理解算法的主要瓶颈之一是缺乏关于世界如何运作的常识性知识。当我们通过语言进行交流时，我们隐含地假设读者会使用这些常识性知识并做出必要的推断。另一方面，计算机无法获得这种共享的共同知识，因此通常无法很好地理解文本，无法执行诸如问答之类的重要任务。该项目将研究如何学习一种关于事件场景的常识性知识：一系列事件（动作）和所涉及的实体类型。例如，一个逮捕场景通常涉及一个犯罪事件和一个逮捕事件，其中有一个逮捕代理人（比如警察）、一名嫌疑人，可能还有一名犯罪受害者。需要明确地告诉语言理解算法来查找这些特定类型的事件和实体。这种方法不能适用于许多可能的现实世界事件场景。该项目将开发机器学习算法，自动获取这类知识，涵盖大型文本集合中的广泛领域。这种算法可以构成各种各样的辅助技术的基础，使公众能够获得信息。例如，从历史文档中生成模式，以帮助学生有针对性地学习历史事件，或者从流媒体新闻来源中提取当前世界事件中涉及的事件和参与者。更一般地说，访问文档的事件结构将实现更好的问题回答功能，适当地嵌入到搜索引擎中，可以使公众更了解情况。该项目将追求三个中心研究重点，以学习常识性事件模式。第一个重点是开发新的概率算法，用于诱导代表现实世界场景的事件模式（例如，嫌疑人被警察逮捕，向法官认罪，后来被定罪）。第二个重点是开发新的模型，从文本中提取这些学习到的模式的实例（例如，John is The Suspect）。这个项目的独特之处在于，它将这些任务形式化为单独的任务，从而可以在传统的关系提取之外对知识学习进行更深入的研究。最后，第三个重点是规范事件图式研究的潜在评估框架。由于这一研究领域的年轻性质，正式的评估和分析在以前的工作中是不一致的。该项目将通过众包产生最大和最多样化的事件模式集，从而实现对未来模型的一致和清晰的评估。

One of the major bottlenecks in current language understanding algorithms is the lack of commonsense knowledge about how the world works. When we communicate through language, we implicitly assume that the readers will use this common sense knowledge and make the necessary inferences. Computers, on the other hand, do not have access to this shared common knowledge, and as a result are often unable to understand text well enough to perform important tasks such as question answering. This project will study methods to learn one type of common sense knowledge about event scenarios: the series of events (actions) and the types of entities involved. For example, an arrest scenario typically involves a crime event, and an arrest event, with an arresting agent (say police), a suspect, and possibly a victim of the crime. Language understanding algorithms need to be explicitly told to look for these specific types of events and entities. This approach does not scale to the many possible real world event scenarios. This project will develop machine learning algorithms that automatically acquire this type of knowledge covering a broad range of domains in large text collections. Such algorithms can form the basis of a wide variety of assistive technology that enables public access to information. Examples include the generation of schemas from historical documents to assist students in targeted learning about historical events, or extraction of events and actors involved in current world events from streaming news sources. More generally, access to the event structure of documents will enable better question answering capabilities that, embedded appropriately into search engines, can lead to a more informed public.The project will pursue three central research thrusts to learning commonsense event schemas. The first thrust develops new probabilistic algorithms for inducing event schemas that represent real-world scenarios (e.g., a Suspect is arrested by Police, pleads to a Judge, and is later convicted). The second thrust will develop new models that extract instances of these learned schemas from text (e.g., John is the Suspect). This project is unique to previous work by formalizing these as separate tasks, and thus enabling deeper research into knowledge learning apart from traditional relation extraction. Finally, the third thrust will standardize potential evaluation frameworks for event schema research. Due to the young nature of this research area, formal evaluation and analysis is inconsistent across previous work. This project will produce the largest and most diverse set of event schemas through crowd-sourcing, enabling consistent and clear evaluation of future models.