Investigation of the viability of direct evaluation of natural language queries with respect to event based triplestores

网络包含了海量的知识；其中大部分是以自然语言(NL)文本编码的。为了方便计算机处理，已经启动了几个项目，以便从文本中提取数据，以通用格式对其进行编码，并将其存储在灵活的数据结构中，这些结构被称为二元关系存储、基于图形的数据库或三元组，其中实体和关系被赋予全球唯一的名称，称为统一资源标识符(URI)。这些灵活的数据结构统称为“语义网”。URI使世界各地的人们能够添加关于同一实体的数据。一个三元组的例子是DBpedia，它从维基百科网络知识库中提取了大约30亿个事实。 可以使用正式查询语言SPARQL(类似于关系数据库查询语言SQL)来查询语义网。SPARQL需要时间来学习，对许多人来说并不容易。 如果可以使用NL访问三元组，这将是有用的。只有很少或根本没有计算机专业知识的用户可以用他们喜欢的自然语言提问。已经建立了几个到语义网的自然语言查询接口(NLQI)；大多数将自然语言查询转换为SPARQL。由于后面解释的原因，现有的NLQI不能容纳包含任意嵌套的量化、否定和链式复杂介词短语的复杂查询，例如在查询中：谁发现了一颗用望远镜绕火星运行的月球？以及“哪个行星没有月球绕轨道运行？” 另一种方法是将NL查询视为lambda演算的表达式，并直接针对存储的数据对其求值。这类似于计算算术表达式的方式。对于自然语言查询，一些词表示从三元组检索的值，一些词表示数学函数，并且通过使用表示查询的语言的语法来解析查询来确定函数应用的顺序。 在过去的两年里，我们已经开发了一套用于语义网查询处理的英语单词的含义(语义)。作为概念验证，我们已经将我们的语义与英语解析器集成在一起，并构建了一个指向包含6000个关于行星和卫星的事实的远程三元组的原型接口。该查询处理器可以处理数以百万计的查询，例如：1938年尼科尔森或霍尔在威尔逊山用望远镜发现了围绕木星运行的哪颗真空卫星？(后面会给出更多的例子)。查询界面可在http://speechweb2.cs.uwindsor.ca/solarman2/demo_sparql.html上访问 下一步是1)调查我们的方法用于大规模三元组存储(如DBpedia)的可行性，2)调查如何帮助用户克服缺乏用于命名实体的URI的知识，以及3)使用已建立的基准查询集调查在召回率和精确度方面的性能。

The web contains a huge amount of knowledge; much encoded in natural-language (NL) text. To facilitate processing by computer, several projects have been launched to extract data from text, encode it in a common format and store it in flexible data structures, variously called binary-relational stores, graph-based databases or triplestores, in which entities and relationships are given worldwide unique names called Uniform Resource Identifiers (URIs). These flexible data structures are collectively called the “semantic web”. The URIs enable data about the same entity to be added by people across the world. An example triplestore is DBpedia with approximately 3 billion facts extracted from the Wikipedia web knowledge base. The semantic web can be queried using the formal query language SPARQL (similar to the relational database query language SQL). SPARQL requires time to learn and it is not easy for many people. It would be useful if triplestores could be accessed using NL. Users with little or no computer expertise could then ask questions in their preferred NL. Several NL query interfaces (NLQIs) to the semantic web have been built; most convert the NL query to SPARQL. For reasons that are explained later, no existing NLQI can accommodate complex queries containing arbitrarily-nested quantification, negation, and chained complex prepositional phrases, such as in the queries “who discovered a moon that orbits mars with a telescope?”, and “which planet is orbited by no moon?” An alternative approach is to treat NL queries as expressions of the lambda calculus and evaluate them directly with respect to the stored data. This is similar to the way in which arithmetic expressions are evaluated. For NL queries, some words denote values which are retrieved from the triplestores, some words denote mathematical functions, and the order of function application is determined by parsing the query using a grammar for the language in which the query was expressed. Over the last 2 years, we have developed a set of meanings (a semantics) for various words of English for use in Semantic-web query processing. As proof of concept, we have integrated our semantics with an English parser, and have built a prototype interface to a remote triplestore containing 6,000 facts about the planets and moons. The query processor can handle millions of queries such as “which vacuumous moon that orbits jupiter was discovered by nicholson or hall with a telescope in 1938 in mt_wilson?” (more examples given later). The query interface is accessible at: http://speechweb2.cs.uwindsor.ca/solarman2/demo_sparql.html The next steps are to 1) investigate the viability of our approach for use with massive triplestores, such as DBpedia, 2) to investigate how to help overcome users lack of knowledge of the URIs used for naming entities, and 3) to investigate performance with respect to recall and precision using established benchmark query sets.