The origins of word order universals: Evidence from corpus statistics and silent gesture
The origins of word order universals: Evidence from corpus statistics and silent gesture
复制标题
词序共性的起源：来自语料库统计和沉默手势的证据
DOI：
--
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
M. Schouwstra
中科院分区：
文献类型：
--
作者：
S. Kirby;J. Culbertson;M. Schouwstra
为什么有些名词短语顺序的模式（例如，N-Adj-Num-Dem，'房子大五这些'）比其他更常见（例如，N-Dem-Num-Adj，“这五个大房子”）？一个有趣的可能性是，这种分布的出现是为了响应一种普遍的认知偏见，这种偏见倾向于概念结构和线性顺序之间的透明关系-同构偏见。在概念结构中，Adj最接近N，然后是Dem，然后是Dem（图1A）。可以从这种嵌套结构中读出的线性顺序是同构的，这些顺序是最常见的跨语言顺序。先前的实验工作已经在传统的人工语言学习（卡伯特森& Adger，2014）和无声手势实验（卡伯特森等人，2016年）。例如，卡伯特森等人（2016）要求非手语英语使用者仅使用手势传达简单的图片。物品是4或5个（N）物体（N）的组，在近端或远端位置（Dem）有斑点或条纹（Adj）。参与者自发地即兴同构的手势，并不反映他们的母语。这些结果表明，同构的偏见是存在的，但他们离开开放的起源，这种偏见。特别是，他们不能告诉我们词序同构的概念结构的起源。在这里，我们表明，概念结构的名词短语是学习通过观察简单的统计对象在世界上。直观地说，我们的建议是：属性（~Adj）比数字（~ Dem）更内在于对象，位置或话语状态（~Dem）通常不是对象的内在特征（参见。Rijkhoff 2002）。更准确地说，我们使用逐点互信息（图1B）来量化这一概念，这是一种衡量元素对之间关联强度的指标。使用语言语料库作为世界的代理，我们可以测量平均pmi 212
Why are some patterns of noun phrase order (e.g., N-Adj-Num-Dem, ‘houses big five these’) much more common than others (e.g., N-Dem-Num-Adj, ‘houses these five big’)? An intriguing possibility is that this distribution emerges in response to a general cognitive bias favouring a transparent relationship between conceptual structure and linear order – an isomorphism bias. In conceptual structure, Adj is closest to N, then Num, then Dem (Fig. 1A). Linear orders that can be read off this nested structure are isomorphic, and these orders are the most common cross-linguistically. Previous experimental work has found an isomorphism in both traditional artificial language learning (Culbertson & Adger, 2014), and silent gesture experiments (Culbertson et al., 2016). For example, Culbertson et al. (2016) asked non-signing English speakers to communicate simple pictures using only gesture. Items were groups of 4 or 5 (Num) objects (N), either spotted or striped (Adj), in a proximal or distal location (Dem). Participants spontaneously improvised isomorphic gestures that did not reflect their native language. These results suggest that an isomorphism bias is present, however they leave open the origin of this bias. In particular, they cannot tell us the origins of the conceptual structure that word order is isomorphic to. Here we show that the conceptual structure of the noun phrase is learnable by observing simple statistics about objects in the world. Intuitively, our proposal is this: properties (~Adj) are more inherent to objects than numerosities (~Num), and location or discourse status (~Dem) is generally not an inherent feature of objects (cf. Rijkhoff 2002). More precisely, we quantify this notion using point-wise mutual information (Fig. 1B), a measure of the strength of association between pairs of elements. Using linguistic corpora as a proxy for the world, we can measure average pmi 212