Knowledge Graph Generation from Web Text

该项目的总体目标是通过大规模网络文本语料库的自动阅读来创建知识图。这意味着从文本中提取信息并将其组织在一个结构化资源中，这使得它更容易用于计算。这种资源的主要用例是问答。我们的方法在以下因素上的新奇：（1）我们的方法将允许使用多种语言的源文本构建这样的图，并且还可以用多种语言对其进行查询。(2)此外，与现有的知识图相比，自然语言查询和我们的图之间的接口将是直接的和没有问题的，因为图的表示语言与自然语言话语的语义表示相同。传统的知识表示、语义解析语言是单独开发的，这意味着将自然语言查询映射到知识数据库的任务是具有挑战性的。(3)最后，我们的图将利用自动挖掘的关于事件之间的蕴涵的信息，以补充从文本中提取的信息，这些信息没有提到，但暗示了。该项目提出的问题是，上述方面是否会为下游应用程序提供好处。其中最明显的是问题回答，我们将其用作我们的评估任务。该项目涉及设计一种合适的表示语言，开发知识图生成算法，将其与蕴涵资源集成，测试可用的问答数据集，并可能开发一个专门用于需要访问和集成多个文档信息的问题的数据集。

The overall goal of the project is Knowledge Graph creation through the automated reading of large scale web text corpora. This means extracting information from text and organizing it in one structured resource, which makes it more accessible for computational use. The main use case of such a resource is question answering. The novelty of our approach on the following factors: (1) our method will allow for construction of such a graph using source text in multiple languages and will also make it possible to query it in multiple languages. (2) Moreover, in contrast to existing Knowledge Graphs, the interface between natural language queries and our graph will be direct and unproblematic because the representation language of the graph is the same as semantic representation of natural language utterances. Traditionally knowledge representation semantic parsing languages were developed separately which means that the task of mapping natural language queries onto knowledge databases is challenging. (3) Finally, our graph will make use of automatically mined information about entailments between events in order to supplement information extracted from text with the facts that were not mentioned but that are implied.The question the project asks is whether the above listed aspects will provide benefits to downstream applications. The most obvious of those is question answering, which we are using as our evaluation task. The project involves designing a suitable representation language, developing a Knowledge Graph generation algorithm, integrating it with entailment resources, testing on available question answering datasets and potentially developing a dataset specifically for questions requiring access and integration of information from multiple documents.