Optimal Processing Times in Reading: A Formal Model and Empirical Investigation-猫眼课题宝

DOI：

发表时间：

2008

期刊：

影响因子：

通讯作者：

中科院分区：

其他

文献类型：

作者：

Nathaniel J. Smith;R. Levy

阅读的最佳处理时间：形式模型与实证研究（njs@pobox.com）认知科学系，9500 Gilman Drive #515拉霍亚，CA 92093-0515美国罗杰·利维（rlevy@ling.ucsd.edu）语言学系，9500吉尔曼驱动器#108拉霍亚，CA 92093-0108美国抽象噪声样本，（贝叶斯最优）平均噪声并提取信号;如果我们是最佳感知鉴别器，那么我们必须收集的样本数量（以及我们必须等待多长时间）取决于信号、噪音和我们先前信念的形式。(For这种方法在单词识别上下文中的一个例子，参见Norris，2006。）然而，感觉系统似乎不太可能是所有反应延迟的罪魁祸首;每次我们启动计算机时，都会提醒我们，计算需要时间。此外，虽然这些模型似乎对视觉感知是合理的，但目前还不清楚它们如何适用于例如听觉。我们可以合理地控制我们看一个场景的时间，但很少控制我们听一个话语的时间。然而，使用单一视觉呈现的单词作为刺激的任务发现反应时间的系统变化-事实上，这些变化与相应的视觉呈现范式中观察到的变化相似（Goldinger，1996）。这就是为什么我们相信语言是一个富有成果的领域，在其中研究替代方法来建模处理时间作为一个最佳的行为。虽然语言可以以书面形式呈现，这对于实验来说非常方便，但我们接触的大部分是口语，口语形式在进化和发展中都具有首要地位。在某种程度上，那么，感官采样的方法表达在最佳感知歧视的框架是不可信的口语理解，这些方法不太可能提供完整的故事一般的语言处理。另一方面，语言处理是高度实践和非常有效的，这表明某种其他类型的优化方法仍然有价值。在本文中，我们提出了一个新的最佳反应时间的模型，我们认为，在最佳准备的框架，可能更适合的领域，如语言处理，我们不能总是控制时间的感觉暴露。该模型的动机是一个公认的事实，即语言理解中的处理时间是概率敏感的：在给定的上下文中，更可预测的单词也会读得更快（例如，埃利希和雷纳，1981）。这在直觉上是明智的--当然，我们宁愿它反过来！- 但它还没有充分的理论化。我们的模型将这一结果解释为在成本函数下的最优行为，该成本函数权衡了准备成本与处理时间;人们想支持-众所周知，人类可以对他们预期的事件比意外事件更快地做出反应，但我们仍然对原因缺乏理解。存在基于最佳感知辨别得出主观概率与响应时间之间的关系的模型，但是这些模型依赖于响应者对环境的感知采样的控制能力，使得它们对于某些领域（例如听觉语言处理）是有问题的，其中概率与响应时间之间仍然存在明显的依赖性。我们提出了一个新的模型，推导出概率和反应时间之间的关系，作为最佳制备的结果。该模型在非常一般的条件下是有效的，仅要求优化结果在输入刺激粒度的尺度上是不变的。该模型作出了强有力的预测，响应时间应与刺激的负条件对数概率线性缩放。我们目前的证据，这一预测在现有的数据库的眼动分析，在阅读自然主义文本。保留字：最佳行为;语言;响应时间模型令人惊讶;句子理解;眼球运动;阅读引言使用物理设备执行计算需要时间，人类大脑就是这样的设备。虽然显而易见，但鉴于最近对最佳行为的理性模型的兴趣激增，这一点值得重新审视（Chater等人，2006; Todorov，2004）。这些模型为行为的许多方面提供了优雅的解释，但处理时间为这种方法提供了特别的挑战。一般来说，人类对任何给定类别的不同刺激的反应速度都不同，反应时间是实验认知心理学的一个重要组成部分。从最优性的角度来看，困难在于不清楚为什么花费在执行计算上的时间应该大于物理最小值。然而，从理论的角度来看，响应时间似乎是最优方法的完美选择，因为它们与进化适应性非常相关。我们都是实时器官，必须在各种各样的情况下做出快速和正确的反应。那么，为什么我们仍然比我们可以慢这么多？在贝叶斯最优框架内解决这些问题的主要方法是将反应时间和其他此类延迟归因于感觉系统。这个论点是，我们需要关于世界的准确信息来采取行动，但我们的感觉系统是嘈杂的。因此，为了获得准确的信息，我们必须等待并收集多个

Optimal Processing Times in Reading: a Formal Model and Empirical Investigation Nathaniel J. Smith (njs@pobox.com) Department of Cognitive Science, 9500 Gilman Drive #515 La Jolla, CA 92093-0515 USA Roger Levy (rlevy@ling.ucsd.edu) Department of Linguistics, 9500 Gilman Drive #108 La Jolla, CA 92093-0108 USA Abstract noisy samples from which to (Bayes-optimally) average out the noise and extract the signal; if we are optimal perceptual discriminators then the number of samples we must gather (and thus how long we must wait) depends on the form of the signal, of the noise, and of our prior beliefs. (For one exam- ple of this approach in the context of word recognition, see Norris, 2006.) It seems unlikely, however, that the sensory system is to blame for all response delays; we are reminded every time we start up our computers that computation qua computation takes time. Furthermore, while such models seem plausible for visual perception, it is unclear how they might apply to, for instance, audition. We have reasonable control over how long we look at a scene, but very little control over how long we listen to an utterance. Yet, tasks using single auditorily presented words as stimuli find systematic variation in reaction time — and, in fact, these variations are similar to those observed in cor- responding visual presentation paradigms (Goldinger, 1996). This is one reason that we believe language to be a fruitful area in which to investigate alternative approaches to model- ing processing time as an optimal behavior. While language can be presented in the written modality, which is very con- venient for experimentation, the bulk of our exposure is to spoken language, and the spoken modality has primacy both evolutionarily and developmentally. To the extent, then, that sensory sampling approaches couched in the framework of optimal perceptual discrimination are implausible for spoken language comprehension, these approaches are unlikely to provide the full story for general language processing. On the other hand, language processing is highly practiced and very efficient, which suggests that some other kind of optimality approach would still be valuable. In this paper, we present a new model of optimal response time couched in a framework of optimal preparation that we believe may be more appropriate to domains such lan- guage processing in which we cannot always control time of sensory exposure. This model is motivated by the well- established fact that processing times in language compre- hension are probability-sensitive: in a given context, words which are more predictable are also read more quickly (e.g. Ehrlich & Rayner, 1981). This is intuitively sensible — cer- tainly we would prefer it to the reverse! — but it is, as yet, inadequately theorized. Our model explains this result as op- timal behavior under a cost function which trades off prepa- ration costs versus processing time; one would like to pro- It is widely known that humans can respond to events they ex- pect more quickly than to unexpected events, but we still have a poor understanding of why. Models exist that derive a relation between subjective probability and response time on the basis of optimal perceptual discrimination, but these models rely on the ability of the responder control over perceptual sampling of the environment, rendering them problematic for some do- mains, such as auditory language processing, in which there are nevertheless clear dependencies between probability and response time. We present a new model deriving the relation- ship between probability and reaction time as a consequence of optimal preparation. This model is valid under very gen- eral conditions, requiring only that the results of optimization are invariant across scale of input stimulus granularity. The model makes the strong prediction that response times should scale linearly with the negative conditional log-probability of the stimulus. We present evidence for this prediction in an analysis of an existing database of eye movements in the read- ing of naturalistic texts. Keywords: Optimal behavior; Language; Response time mod- eling; Surprisal; Sentence comprehension; Eye movements; Reading Introduction It takes time to perform computation using a physical device, and the human brain is such a device. While obvious, this point is worth revisiting in light of the recent surge of interest in rational models of optimal behavior (Chater et al., 2006; Todorov, 2004). Such models have provided elegant explana- tions for many aspects of behavior, but processing time pro- vides a particular challenge for this approach. In general, hu- mans respond to different stimuli within any given class with different speeds, and response times are a large part of the stock and trade of experimental cognitive psychology. From an optimality perspective, the difficulty is that it is unclear why the time to spend on performing a computation should ever be larger than the physical minimum. Yet, from a theo- retical point of view, response times seem like the perfect can- didate for an optimality approach, because they are so clearly relevant to evolutionary fitness. We are all real-time organ- isms who must react quickly and correctly in a wide variety of circumstances. So why are we still so much slower than we could be? The primary approach to these problems deployed within the Bayes-optimal framework has been to ascribe reaction times and other such delays to the sensory system. The argu- ment is that we require accurate information about the world to act, but our sensory system is noisy. Therefore, to ac- quire accurate information, we must wait and gather multiple

Optimal Processing Times in Reading: A Formal Model and Empirical Investigation

文献摘要

被引文献

相似文献