Towards Wide-Coverage Semantic Interpretation
Towards Wide-Coverage Semantic Interpretation
复制标题
走向广泛覆盖的语义解释
DOI：
--
复制
发表时间：
2005
期刊：
影响因子：
--
通讯作者：
Johan Bos
中科院分区：
文献类型：
--
作者：
Johan Bos
广泛的覆盖面和强大的NLP技术似乎总是与浅层分析密切相关。几年前确实是这样，但是最先进的随机方法已经取得了相当大的进步，现在有成熟的解析器可以实现高覆盖率并产生准确的语法分析。看来我们终于到达了NLP的一个阶段，我们可以在更大的范围内应用众所周知的形式语义和计算语义技术，并从广泛覆盖的解析器中获得详细的语义分析。[BCS04]证明了这一想法的概念，报纸文本的覆盖率超过95%。在本文中，我们讨论了这项工作的进一步发展，为句子或小文本生成语义表示，展示了我们如何计算推理所需的背景知识，并使用最先进的定理证明器和模型构建器进行推理。我们将使用的语义表示语言是一阶语言，认为给定当前自动演绎的状态，任何具有更强表达能力的语言（如二阶或高阶逻辑）都不能有效地用于执行推理任务。然而，对于一阶逻辑有非常复杂的推理工具，我们将在我们的工作中使用。尽管在形式语义中有使用高阶逻辑的传统，但一级逻辑能够涵盖（可能令人惊讶的）大量有趣的自然语言现象。我们将采用的语言是在话语表示理论（DRT）中发展起来的，在一阶逻辑的翻译下关闭，并在第2节中描述。选择DRT的动机是其令人印象深刻的语言现象的理论覆盖[KR93， VdS92]。当然，接下来，我们需要一种适合计算语义的语法形式（即能够产生细粒度语法分析的语法形式）。
Wide-coverage and robust NLP techniques always seemed to go hand in hand with shallow analyses. This was certainly true a couple of years ago, but the state-of-the-art in stochastic approaches has advanced considerably and nowadays there are sophisticated parsers available achieving high coverage and producing accurate syntactic analyses. It seems we have finally reached a stage in NLP where we can apply well known techniques of formal and computational semantics to a larger scale, and get a detailed semantic analysis from a wide-coverage parser. A proof of concept of this idea was demonstrated in [BCS04], with a coverage of over 95% on newspaper texts. In this paper we discuss the further developments in this work, generating semantic representations for sentences or small texts, showing how we can calculate background knowledge required for reasoning, and performing inferences using state-of-the-art theorem provers and model builders. The semantic representation language that we will use is a first-order language, arguing that given the current state of automated deduction, any language with more expressive power (such as second or higher-order logic) cannot be used efficiently to perform inference tasks. There are however highly sophisticated inference tools for first-order logic available which we will use in our work. Despite the tradition in formal semantics to use higher-order logics, firstorder logic is able to cover a (perhaps surprisingly) large variety of interesting natural language phenomena. The language we are going to adopt is developed in Discourse Representation Theory (DRT), closed under a translation to first-order logic, and is described in Section 2. The choice for DRT is motivated by its impressive theoretical coverage of linguistic phenomena [KR93, VdS92]. Next, of course, we need a grammar formalism suitable for computational semantics (i.e. one that is able to produce fine-grained syntactic analyses).