DeFacto: Acquiring, Curating, and Using a Bilingual Domain Aware Commonsense Knowledge Base

自动从大量非结构化文档(如Web)中提取知识并将其组织到知识库(KB)中是人工智能中的一个关键挑战。直观地说，这样的知识库应该会直接影响许多自然语言处理应用的质量，如问题回答、信息检索或文本分析。开放式信息提取，即在没有太多监督的情况下从文本中提取知识的任务(特别是不是给我开出这类信息的处方)，为这一努力带来了新的希望。*尽管一些设计良好的组件现已广泛存在，并随时可用于从文本中提取事实和关系(所谓的元组)，但在大量文本集合中挖掘信息仍然会带来许多问题。嵌入到典型的知识提取管道中的技术充满了缺点：共指解析、命名实体解析和解析错误正在崩溃，因此许多元组(如果不是绝大多数的元组)根本就是无用的。此外，大多数作品的目标是非常频繁的实体和关系，这些实体和关系排除了大量关于Web上普遍存在的领域特定文本的信息。*我们的长期目标是发展必要的专业知识来填充、管理、维护和使用知识库。由于一些关键因素，我们的建议与现有的几项倡议有所不同。首先，由于特定的域在Web上很流行，我们希望我们的技术是域感知的。其次，由于当今世界是多语言的，而且并不是所有的东西都是用英语写的，我们更希望我们的技术本质上是多语言的。最后，大多数工作都致力于开发全自动辅助人类的技术。在我们的提案中，我们感兴趣的是衡量有目的的游戏能在多大程度上让人类协助计算机。*为了成功，我们在这项提案中的目标是开发一个事实，一个多领域的，双语的知识库(法语-英语)迭代地从网上挖掘的文本中获得，并借助通过严肃的游戏从用户那里收集的反馈。

Automatically extracting knowledge from a large set of mostly unstructured documents (such as the Web) and organizing it into a knowledge base (KB) is a key challenge in artificial intelligence. Intuitively, such KBs should directly impact the quality of many NLP applications such as question answering, information retrieval or Text Analytics. Open information extraction, the task of extracting knowledge from texts without much supervision (especially not a prescription of the kind of information to mine), has brought new hope for such an endeavour. ******Despite a number of well-designed components are nowadays widespread and readily available for extracting facts and relations (so-called tuples) from texts, tapping information in large collections of texts still raises a number of issues. The technology embedded in a typical knowledge extraction pipeline is fraught with shortcomings: coreference resolution, named-entity resolution and parsing errors are collapsing so that many tuples (if not the vast majority) are simply useless. Also, most works are targeting very frequent entities and relations, which exclude a large quantity of information on domain specific texts that are pervasive over the Web. ******Our long term objective consists in developing the necessary expertise in populating, curating, maintaining and using a KB. Our proposal departs from several existing initiatives by a number of key factors. First, since specific domains are prevalent over the Web, we want our technology to be domain aware. Second, since today's world is multi-lingual and because not everything is written in English, we further want our technology to be multi-lingual in nature. Last, most works are devoted to develop fully automatic technology for assisting humans. In our proposal, we are interested in measuring how much gaming with a purpose can make humans assist the computer. ******In order to succeed, we target in this proposal the development of deFacto, a multi-domain, bilingual KB (French -- English) acquired iteratively from texts mined over the web, with the help of feedback collected from users via serious gaming.