Categorizing Grammar: Differential Effects of Preceding and Succeeding Contextual Cues
Categorizing Grammar: Differential Effects of Preceding and Succeeding Contextual Cues
复制标题
语法分类：前后语境线索的不同影响
DOI：
--
复制
发表时间：
2005
期刊：
影响因子：
--
通讯作者：
Akifumi Toksumi
中科院分区：
文献类型：
--
作者：
H. Murai;Akifumi Toksumi
分类语法：前后语境线索的差异效应米歇尔·C·圣克莱尔(msc110@york.ac.uk)约克大学心理学系，YO10 5DD，UK Padraic Monaghan(pjm21@york.ac.uk)约克大学心理学系，YO10 5DD，UK摘要分布信息已被证明与语音信息相结合，有助于将单词分类为语法类别。关于哪种类型的分布信息在类别学习中最有用，特别是二元语法是否足以用于类别习得，一直存在争议。本文提供了两个实验来测试人们对用于分类的二元语法信息的敏感性。句子由类别标记出现在类别词之前(实验1)或之后(实验2)的词组成。判决书的陈述有声有色。在两个实验中都获得了分类信息，但之前和之后的分布线索以不同的方式促进了学习。此外，在两种情况下，语音信息都有助于低频词的学习。儿童如何在他们的语言中习得语法范畴？特别是，有哪些潜在的信息来源可以使这一进程得以发生？在人工语法学习实验中，我们提出了两个实验，探索儿童环境中语境信息和语音信息之间的相互作用，以及这些来源在多大程度上能够推动学习。Mintz(2002,2003)建议儿童根据儿童语言环境中频繁出现的“框架”来习得语法类别。这样的框架由紧接在目标词之前和之后的共现词来定义，即，形式为aXb的短语，其中X是类别词，并且a和b在文本中频繁地共现。在以儿童为导向的英语演讲中，“The X is…”就是一个例子其中X可以是几个名词中的一个。Mintz(2002)进行了一项人工语言研究，将频繁框架作为类别的标记，并发现了基于这种结构进行分类的证据。为了进行这项工作，Mintz(2003)对来自Childes(MacWhinney，2000)的儿童定向言语的小样本进行了语料库分析。使用这些语料库中最频繁的45个帧，Mintz(2003)发现基于这些帧的类别分组的准确率非常高。然而，完整性非常低，尽管显著高于随机基线。Monaghan和Christian(2004)提出，这种低完备性表明频繁的帧具有非常低的覆盖率，因此这不是用于分类的信息来源。相反，他们认为二元语法信息可以提供关于类别的更丰富的线索。AXb帧中的三元语法信息与AX(例如，X)和XB(例如，X是)二元语法合并，并且在Mintz(2002)中的学习可能仅由二元语法级别的信息驱动。Monaghan和Christian(2004)重复了Mintz(2003)的语料库分析，发现aXb框架的准确率很高，完备性很低，而二元语法信息的准确率略低，但完备性要高得多：对69.9%的单词进行分类，而使用aXb框架的话只有14.3%。此外，Monaghan和Christian(2004)在学习语法范畴的神经网络模型中发现，当aXb框架被分解为ax和xb时，学习会增加。然而，二元语法信息可能对类别学习有用的发现并不等同于证明它是可用的。然而，Vian和Coulson(1988)构建了一种人工语言，其中二元语法信息被用于类别学习。句子的形式是AABB，其中a和b是高频标记词，A和B是类别词集合。他们发现，新奇的AABB句子比那些违反二元语法信息的句子更受欢迎：*ABBA。在这项研究的延伸中，Monaghan，Chater和Christian(2005)发现了来自同一类别的词被归类在一起的直接证据。这两个研究都探索了AX双词信息的学习，但到目前为止还没有证据表明类别可以从xB双词中学习。此外，这些人工语言研究是在句子的视觉呈现下进行的。当呈现方式改变时，分类表现可能会有所不同。在一项人工语言学习实验中，Onnis、Christian ansen、Chater和Gomez(2003)发现，视觉呈现的刺激与听觉呈现的刺激有类似的趋势，尽管前者的影响有所减弱。分布信息并不是决定语法分类的唯一线索；语音信息也与语法和性别类别的学习有关(Braine等人，1990；Brooks，Braine，Catalano，Brody，&Sudhalter，1993)。事实上，Braine(1987)声称，语音信息的支持对于类别的可学习至关重要。凯利(1992)已经展示了
Categorizing Grammar: Differential Effects of Preceding and Succeeding Contextual Cues Michelle C. St. Clair (msc110@york.ac.uk) Department of Psychology, University of York York, YO10 5DD, UK Padraic Monaghan (pjm21@york.ac.uk) Department of Psychology, University of York York, YO10 5DD, UK Abstract Distributional information has been shown to combine with phonological information in aiding categorisation of words into grammatical categories. There has been debate about the type of distributional information that is most useful in category learning, in particular whether bigrams are sufficient for category acquisition. This paper presents two experiments testing people’s sensitivity to bigram information for categorisation. Sentences were composed of words with category markers occurring either before (Experiment 1) or after (Experiment 2) the category word. Sentences were presented auditorily. Categorical information was learned in both experiments, but the preceding and succeeding distributional cues contributed to learning in different ways. Furthermore, in both cases phonological information assisted the learning of low frequency words. Introduction How do children acquire grammatical categories in their language? In particular, what sources of information are potentially available to enable this process to occur? We present two experiments exploring the interaction between contextual information and phonological information in the child’s environment, and the extent to which these sources can drive learning in artificial grammar learning experiments. Mintz (2002, 2003) has suggested that children acquire grammatical categories based on frequently occurring “frames” in the child’s language environment. Such frames are defined by co-occurring words immediately prior to and succeeding the target word, i.e., a phrase of the form aXb, where X is the category word, and a and b co-occur frequently in text. In child-directed speech in English, one example is the frame “The X is…” where X could be one of several nouns. Mintz (2002) conducted an artificial language study that incorporated frequent frames as markers for categories, and found evidence for categorisation based on this structure. To follow this work, Mintz (2003) conducted corpus analyses of small samples of child directed speech taken from CHILDES (MacWhinney, 2000). Using the 45 most frequent frames in these corpora, Mintz (2003) found that the accuracy of the category groupings based on these frames was extremely high. However, completeness was very low, though was significantly above a random baseline. Monaghan and Christiansen (2004) suggested that this low completeness indicated that frequent frames had very low coverage and that this was consequently a poor source of information for categorisation. Instead, they suggested that bigram information could provide richer cues about category. The trigram information in the aXb frames was conflated with aX (e.g., “The X”) and Xb (e.g., “X is”) bigrams, and learning in Mintz (2002) may have been driven by information at the bigram level only. Monaghan and Christiansen (2004) replicated Mintz’s (2003) corpus analyses, finding high accuracy and low completeness for aXb frames, and slightly lower accuracy but much higher completeness for bigram information: categorizing 69.9% of the words, compared to only 14.3% using the aXb frame. In addition, Monaghan and Christiansen (2004), in neural network models trained to learn grammatical categories, found that when the aXb frame was decomposed as aX and Xb then learning increased. Yet the finding that bigram information is potentially useful for category learning is not the same as demonstrating that it is useable. However, Valian and Coulson (1988) constructed an artificial language where bigram information was exploited for category learning. Sentences were of the form aAbB, where a and b were high frequency marker words, and A and B were sets of category words. They found that novel aAbB sentences were preferred to those that violated the bigram information: *aBbA. In an extension of this study, Monaghan, Chater, and Christiansen (2005) found direct evidence for words from the same category being grouped together. Both these studies have explored the learning of aX bigram information, but as yet there is no evidence that categories can be learned from Xb bigrams. Furthermore, these artificial language studies were conducted with sentences presented visually. There may be a difference in categorisation performance when the modality of presentation is altered. In an artificial language learning experiment, Onnis, Christiansen, Chater, and Gomez (2003) found a similar trend in visually presented stimuli compared to auditory presentation, though the effect was reduced in the former. Distributional information is not the only cue determining grammatical categorisation; phonological information has also been implicated in the learning of grammatical and gender categories (Braine et al., 1990; Brooks, Braine, Catalano, Brody, & Sudhalter, 1993). Indeed Braine (1987) has claimed that support from phonological information is vital for categories to be learnable. Kelly (1992) has shown