Querying Heterogeneous Data for Non-Expert Users

使用互联网回答我们的问题几乎是一种日常做法，通常是使用谷歌或必应等搜索引擎。通常，这项任务需要浏览网页以找到答案。在过去的几年里，如果问题很简单(例如，找到关于一个著名实体的事实)，答案就会显示在所谓的信息箱中。搜索引擎通过准确地提取自然语言问题中感兴趣的实体并将其映射到其内部数据表示世界中的等效实体来显示这一事实。这种内部表示当然是专有的，不同的搜索引擎会有所不同。然而，方法是相同的：数据以结构化格式(表或通常称为知识图的形式)存储，并对其发出查询。自然语言是访问底层结构化数据的最流行的接口之一。这种方法的简单性允许非专家用户访问过多的数据集。然而，对于稍微复杂的问题，检索自然语言查询的准确答案是一项具有挑战性的任务。在这种情况下，为了检索准确的答案，用户需要用结构化查询语言编写正式查询。当然，这样的要求是不实际的，并且限制了利用底层结构化数据中可用的知识的力量来支持拥有足够技术技能与复杂数据集交互的用户。该研究计划旨在帮助非专家用户查询来自不同来源并以不同格式存储的数据集。为了实现这一目标，我们将致力于如何整合来自多个来源和格式的数据，以便为查询做好准备，如何建立一个支持自然语言界面和其他新颖查询模式的查询系统，以帮助非专家用户查询集成的数据，以及如何建立能够准确评估查询系统质量的基准。

Using the Internet to answer our questions is an almost daily practice, typically, using search engines like Google or Bing. Usually, this task involves browsing web pages to find the answer. In the past few years, if the question is simple (e.g., finding a fact about a famous entity), the answer is shown in what is called an infobox. Search engines show this fact by accurately extracting the entity of interest in the question written in natural language and map it to its equivalent entity in their internal data representation of the world. This internal representation is of course proprietary and different from one search engine to another. However, the approach is the same: The data is stored in a structured format (tables or what is typically called a knowledge graph) and queries are issued against it. Natural language is one of the most popular interfaces to access the underlying structured data. The simplicity of this approach allows non-expert users to access this plethora of datasets. However, retrieving accurate answers to natural language queries is a challenging task for even slightly more complicated questions. In this scenario, in order to retrieve accurate answers, the user needs to write a formal query in a structured query language. Of course, such a requirement is not practical and limits harnessing the power of knowledge available in the underlying structured data to power users who have sufficient technical skills to interact with the complex datasets. This research program aims at helping non-expert users to query datasets that come from heterogeneous sources and are stored in different formats. To achieve this goal, we will work on how to integrate data from multiple sources and formats for it to be ready for querying, how to build a querying system that supports natural language interfaces and other novel querying models to help a non-expert user to query the integrated data, and finally how to build benchmarks that can accurately assess the quality of the querying system.