Typicality and Object Reference
Typicality and Object Reference
复制标题
典型性和对象参考
DOI：
--
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
Kees van Deemter
中科院分区：
文献类型：
--
作者：
Margaret Mitchell;Ehud Reiter;Kees van Deemter
典型和对象参考Margaret Mitchell （m.mitchell@jhu.edu）人类语言技术卓越中心，约翰霍普金斯大学巴尔的摩，马里兰州21211 USA Ehud Reiter (e.reiter@abdn.ac.uk) Kees van Deemter （k.vdeemter@abdn.ac.uk）阿伯丁大学计算科学系，苏格兰阿伯丁AB24 3FX UK摘要一个对象的典型是否会影响我们如何识别它？当我们对一个可见的物体产生最初的参考时，我们受到各种因素的影响，包括视觉上显著的东西（自下而上的影响）以及我们以前对这个物体的经验（自上而下的影响）。在本研究中，我们试图了解自上而下的典型性影响如何影响对对象的初始参考。我们使用现实世界，日常物品，并专注于形状和材料的视觉特性。我们的研究结果表明，人们倾向于选择非典型而不是典型。但是，我们对现实世界对象的理解才刚刚开始触及表面。本研究的标注语料库可为未来视觉领域的建模参考研究提供参考。关键词：指称表达式；描述;参考;vi -锡安;(a) GRE3D3场景。图1：来自GRE3D3语料库和TUNA Furniture子语料库的示例场景。参与者做出诸如红色立方体上的黄球（GRE3D3）或小扇子（TUNA）之类的环形表达。我从来没见过紫色的奶牛。我从来没有希望看到一个，但我可以这样说：我宁愿看到而不是成为一个。-格列特·伯吉斯当我们为听者识别一个物体时，我们有很多选择要提到什么。当对象对说话者和听者都可见时，有助于引导视觉注意力的属性，如颜色和大小，特别具有信息作用（Treisman & Gelade, 1980; Wolfe, 2006）。话语中突出或与说话者和受话者之前的互动相关的属性也会影响我们将要提到和描述的内容（Clark & Wilkes-Gibbs, 1986; Brennan & Clark, 1996; Clark & Krych, 2004）。我们假设，当我们为听者生成对一个对象的初始引用时，我们对同一类型对象的知识也可能影响我们提到的内容。换句话说，当我们第一次描述一个对象时，我们对一个对象类别的典型特征的理解会影响我们对修饰语的选择——形容词和更长的描述性短语。这种对典型对象类别的理解可能源于存储的对象原型（Rosch & Mervis, 1975; Rosch， C. Mervis， W. Gray, Johnson, & Braem, 1976）或在以前情况下对类似对象的心理表征（Yeh & Barsalou, 2006; Wu & Barsalou, 2009）。由于典型化，我们知道上面例子中提到的紫色奶牛是非凡的。以往的文献研究很少关注典型性的作用。这同样适用于心理语言学1 http://m-mitchell.com/corpora/typicality语料库/ (b) TUNA场景。工作（阿诺德，2008年）和计算模型的参考（Dale & Reiter， 1995年；Krahmer, van Erk, & Verleg， 2003年；Krahmer & van Deemter， 2012年）。目前的研究解决了我们认为在我们对参考的理解上存在的重大差距。我们研究典型性的作用，参考现实世界，日常物品，重点放在材料和形状属性。对象是这样呈现的，这两个属性之一将区分对象。我们测试了当一组参与者在典型形状或典型材料之间选择描述目标物体，而另一组参与者在典型形状或非典型材料之间选择描述目标物体时，两组之间是否存在显著差异。我们的研究结果表明，有一种选择非典型而不是典型的倾向。虽然这项研究的重点是形状和材料的典型特征，但我们发布了我们实验的完整语料库，并与各种视觉特性相关联，希望有助于进一步构建真实物体的参考模型。目前可用于参考可见物体的语料库，如GRE3D3语料库（Viethen & Dale, 2008）或TUNA语料库（van Deemter, Gatt, van der Sluis, & Power, 2012），都是根据对计算机屏幕上显示的简单物体图形的参考而构建的（见图1）。在这项工作中，我们试图更好地理解现实世界视觉领域中丰富的参考细节，其中许多不同的视觉属性相互作用。这开辟了以前没有研究过的参考的几个方面，并进一步提出了影响初始参考和视觉对象描述的因素的问题。我们将讨论其中的一些问题，以及它们对计算机的影响
Typicality and Object Reference Margaret Mitchell (m.mitchell@jhu.edu) Human Language Technology Center of Excellence, Johns Hopkins University Baltimore, MD 21211 USA Ehud Reiter (e.reiter@abdn.ac.uk) Kees van Deemter (k.vdeemter@abdn.ac.uk) Computing Science Department, University of Aberdeen Aberdeen, Scotland AB24 3FX UK Abstract Does the typicality of an object affect how we identify it? When we produce initial reference to a visible object, we are influenced by a variety of factors, including what is visually salient (bottom-up influences) as well as our previous experi- ences with the object (top-down influences). In this study, we seek to understand how the top-down influence of typicality affects initial reference to an object. We use real world, every- day objects, and focus on the visual properties of SHAPE and MATERIAL . Our findings suggest that there is a tendency to select the atypical over the typical. But we have only begun to scratch the surface of understanding reference to real world ob- jects. The annotated corpus from this study is made available for future research on modeling reference in visual domains. 1 Keywords: referring expressions; description; reference; vi- sion; typicality (a) A GRE3D3 scene. Figure 1: Example scenes from the GRE3D3 Corpus and the TUNA Furniture sub-corpus. Participants produce refer- ring expressions such as yellow ball on top of the red cube (GRE3D3) or small fan (TUNA). Introduction I never saw a purple cow. I never hope to see one, But I can say this anyhow: I’d rather see than be one. — Gelett Burgess When we identify an object for a hearer, we have a number of choices to make about what to mention. When the object is visible to both speaker and hearer, properties that help guide visual attention, such as color and size, are particularly infor- mative (Treisman & Gelade, 1980; Wolfe, 2006). Properties that are salient to the discourse or relevant to the speaker and hearer’s previous interactions also affect what we will men- tion and describe (Clark & Wilkes-Gibbs, 1986; Brennan & Clark, 1996; Clark & Krych, 2004). We hypothesize that when we generate initial reference to an object for a hearer, our knowledge about objects of the same type is also likely to affect what we mention. In other words, our understanding of what is typical for an object cat- egory influences the selection of modifiers – the adjectives and longer descriptive phrases – that we produce when we first describe an object. This understanding of what is typical for an object category may stem from stored object proto- types (Rosch & Mervis, 1975; Rosch, C. Mervis, W. Gray, Johnson, & Braem, 1976) or mental representations of simi- lar objects in previous situations (Yeh & Barsalou, 2006; Wu & Barsalou, 2009). Because of typicality, we know that the purple cow mentioned in the example above is remarkable. Previous work on reference has paid little attention to the role of typicality. This is equally true for psycholinguistic 1 http://m-mitchell.com/corpora/typicality corpus/ (b) A TUNA scene. work (Arnold, 2008) and for work on computational mod- els of reference (Dale & Reiter, 1995; Krahmer, van Erk, & Verleg, 2003; Krahmer & van Deemter, 2012). The present study addresses what we believe to be a significant gap in our understanding of reference. We examine the role of typicality in reference to real world, everyday objects, focusing on material and shape properties. Objects are presented so that one of these two properties will distinguish the object. We test whether there is a sig- nificant difference between groups when participants in one may choose between atypical shape or typical material to describe target objects, while participants in a second group may choose between typical shape or atypical material to de- scribe target objects. Our findings suggest that there is a ten- dency to select the atypical over the typical. Although this study focuses on shape and material typical- ity, we release the full corpus from our experiments, anno- tated with a variety of visual properties, in hopes of helping further work in constructing models of reference to real ob- jects. Current available corpora for reference to visible ob- jects, such as the GRE3D3 Corpus (Viethen & Dale, 2008) or the TUNA Corpus (van Deemter, Gatt, van der Sluis, & Power, 2012), were built from reference elicited to graphics of simple objects presented on a computer screen (see Figure 1). In this work, we seek to better understand the rich details of reference in real world visual domains, where a multitude of different visual properties interact. This opens up several aspects of reference that have not been researched before and gives rise to further questions about the factors influencing initial reference and visual object descriptions. We discuss some of these issues, and their implications for a computa-