Meaning Representation in Natural Language Categorization-猫眼课题宝

自然语言分类中的含义表示

DOI：

发表时间：

2010

期刊：

影响因子：

通讯作者：

中科院分区：

文献类型：

作者：

Trevor Fountain;Mirella Lapata

自然语言范畴化中的意义表示特雷弗喷泉（t.喷泉@ sms.ed.ac.uk）和米雷拉拉帕塔（mlap@inf.ed.ac.uk）爱丁堡大学信息学院10 Crichton Street，Edinburgh EH 8 9AB，UK摘要近年来提出了大量的范畴化形式模型。其中许多都是在人工类别或感知刺激上进行测试的。在本文中，我们专注于自然语言概念的范畴化模型，并具体解决如何表示这些概念的问题。许多语义认知的心理学理论认为，概念是由人类通常引出的特征定义的。规范化研究产生了关于意义表征的详细知识，但是它们是小规模的（针对几百个词获得特征），并且对于自然语言分类的一般模型的使用有限。作为一种替代方法，我们研究是否可以用从大型文本集合中提取的简单共现统计数据来定量表示类别意义。基于特征的分类模型与基于数据驱动表示的模型的实验比较表明，后者代表了通常使用的特征规范的可行替代方案。相当多的心理学研究表明，人们对他们遇到的新奇物体的推理是通过识别这些物体所属的类别，并从他们过去与该类别其他成员的经验中推断出来的。这种分类的任务，或者说把物体分成有意义的类别，是齿轮科学领域的一个经典问题，其研究历史可以追溯到阿里斯-托特。这并不奇怪，因为对类别进行推理的能力是许多其他任务的核心，包括感知，学习和语言使用。关于人类如何对物体进行分类，存在着许多理论。这些理论本身往往属于三个思想流派之一。在古典主义（或亚里士多德）的观点中，范畴是由一系列“必要和充分”的特征定义的。例如，BACHELOR这个概念的定义特征可能是男性、单身和成年人。不幸的是，这种方法无法解释类别的大多数普通用法，因为许多现实世界的对象具有某种模糊的定义，并且不完全适合定义良好的类别（Smith和Medin，1981）。原型理论（Rosch，1973）提出了这一思想的另一种表述，其中范畴是由一个理想化的原型成员定义的，该成员拥有对范畴至关重要的特征。如果对象表现出足够的这些特征，则它们被认为是该类别的成员;例如，FRUIT的特征可能包括包含种子，生长在地面上，并且是可食用的。粗略地说，原型理论与经典理论的不同之处在于，范畴的成员不需要拥有原型中规定的所有特征。虽然原型理论提供了一个上级和可行的替代经典理论，它已受到挑战的范例方法（Medin和Schaffer，1978）。在这个视图中，类别不是由单个表示定义的，而是由以前遇到的成员的列表定义的。代替维护列出水果的典型特征的FRUIT的单个原型，范例模型简单地存储它已经暴露于的水果的那些实例（例如，苹果、橘子、梨）。如果新对象与存储在内存中的一个或多个FRUIT实例足够相似，则将其分组到类别中。在过去的实验研究中，许多实验工作已经在实验室研究中测试了基于原型和样例的理论的预测，涉及分类和类别学习。这些实验倾向于使用感知刺激和人工分类（例如，例如100000或0111111的数字序列串）。类似地，许多建模工作都集中在如何表示类别和刺激的问题上（Griffiths等人，2007 a; Sanborn等人，#20006;，如何更好地进行分类。后者在原型模型和范例模型中都起着重要的作用，因为对新对象的正确概括取决于正确识别错误遇到的项目。本文主要研究自然语言概念的范畴化问题.与大量使用知觉刺激或人工分类的研究相比，令人惊讶的是，关于成年人如何学习或使用自然语言类别的研究很少。一些值得注意的例外是Heit和Barsalou（1996），他们试图在自然语言概念的背景下实验性地测试范例模型，Storms等人（2000）在许多自然分类任务上评估范例和原型模型之间的性能差异，Voorspoels等人（2008）为自然语言概念的典型性评级建模。这项工作的一个共同假设是，分类中涉及的概念的含义可以由一组特征（也称为属性或属性）表示。事实上，特征表征在语义认知和知识组织的心理学理论中发挥了核心作用，并且已经进行了许多研究来引出特征的详细知识。在一个典型的过程中，参与者被给予一系列对象名称，对于每个对象，他们被要求说出他们能想到的所有属性，这些属性是对象的特征。尽管特征规范通常被解释为语义表征结构的有用代理，但仍存在一些困难

Meaning Representation in Natural Language Categorization Trevor Fountain (t.fountain@sms.ed.ac.uk) and Mirella Lapata (mlap@inf.ed.ac.uk) School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB, UK Abstract A large number of formal models of categorization have been proposed in recent years. Many of these are tested on artificial categories or perceptual stimuli. In this paper we focus on cat- egorization models for natural language concepts and specif- ically address the question of how these may be represented. Many psychological theories of semantic cognition assume that concepts are defined by features which are commonly elicited from humans. Norming studies yield detailed knowl- edge about meaning representations, however they are small- scale (features are obtained for a few hundred words), and ad- mittedly of limited use for a general model of natural language categorization. As an alternative we investigate whether cate- gory meanings may be represented quantitatively in terms of simple co-occurrence statistics extracted from large text col- lections. Experimental comparisons of feature-based catego- rization models against models based on data-driven represen- tations indicate that the latter represent a viable alternative to the feature norms typically used. Introduction Considerable psychological research has shown that people reason about novel objects they encounter by identifying the category to which these objects belong and extrapolating from their past experiences with other members of that cat- egory. This task of categorization, or grouping objects into meaningful categories, is a classic problem in the field of cog- nitive science, one with a history of study dating back to Aris- totle. This is hardly surprising, as the ability to reason about categories is central to a multitude of other tasks, including perception, learning, and the use of language. Numerous theories exist as to how humans categorize ob- jects. These theories themselves tend to belong to one of three schools of thought. In the classical (or Aristotelian) view cat- egories are defined by a list of “necessary and sufficient” features. For example, the defining features for the concept BACHELOR might be male, single, and adult. Unfortunately, this approach is unable to account for most ordinary usage of categories, as many real-world objects have a somewhat fuzzy definition and don’t fit neatly into well-defined cate- gories (Smith and Medin, 1981). Prototype theory (Rosch, 1973) presents an alternative for- mulation of this idea, in which categories are defined by an idealized prototypical member possessing the features which are critical to the category. Objects are deemed to be members of the category if they exhibit enough of these features; for example, the characteristic features of FRUIT might include contains seeds, grows above ground, and is edible. Roughly speaking, prototype theory differs from the classical theory in that members of the category are not required to possess all of the features specified in the prototype. Although prototype theory provides a superior and work- able alternative to the classical theory it has been challenged by the exemplar approach (Medin and Schaffer, 1978). In this view, categories are defined not by a single representation but rather by a list of previously encountered members. Instead of maintaining a single prototype for FRUIT that lists the fea- tures typical of fruits, an exemplar model simply stores those instances of fruit to which it has been exposed (e.g., apples, oranges, pears). A new object is grouped into the category if it is sufficiently similar to one or more of the FRUIT instances stored in memory. In the past much experimental work has tested the predic- tions of prototype- and exemplar-based theories in laboratory studies involving categorization and category learning. These experiments tend to use perceptual stimuli and artificial cat- egories (e.g., strings of digit sequences such as 100000 or 0111111). Analogously, much modeling work has focused on the questions of how categories and stimuli can be rep- resented (Griffiths et al., 2007a; Sanborn et al., 2006) and how best to formalize similarity. The latter plays an impor- tant role in both prototype and exemplar models as correct generalization to new objects depends on identifying previ- ously encountered items correctly. In this paper we focus on the less studied problem of cat- egorization of natural language concepts. In contrast to the numerous studies using perceptual stimuli or artificial cate- gories, there is surprisingly little work on how natural lan- guage categories are learned or used by adult speakers. A few notable exceptions are Heit and Barsalou (1996) who attempt to experimentally test an exemplar model within the context of natural language concepts, Storms et al. (2000) who eval- uate the differences in performance between exemplar and prototype models on a number of natural categorization tasks, and Voorspoels et al. (2008) who model typicality ratings for natural language concepts. A common assumption underly- ing this work is that the meaning of the concepts involved in categorization can be represented by a set of features (also referred to as properties or attributes). Indeed, featural representations have played a central role in psychological theories of semantic cognition and knowl- edge organization and many studies have been conducted to elicit detailed knowledge of features. In a typical procedure, participants are given a series of object names and for each object they are asked to name all the properties they can think of that are characteristic of the object. Although fea- ture norms are often interpreted as a useful proxy of the struc- ture of semantic representations, a number of difficulties arise

Meaning Representation in Natural Language Categorization

自然语言分类中的含义表示

文献摘要

被引文献

相似文献