Learning covert phonological interaction : an analysis of the problem posed by the interaction of stress and epenthesis
Learning covert phonological interaction : an analysis of the problem posed by the interaction of stress and epenthesis
复制标题
学习隐性语音交互：重音和附加词交互带来的问题分析
DOI：
--
复制
发表时间：
2002
期刊：
影响因子：
--
通讯作者：
B. Tesar
中科院分区：
文献类型：
--
作者：
B. Tesar
本报告的目的是通过分析隐性语音互动所带来的问题，为优选论中的形式学习理论做出贡献。基于音节结构和韵律重音的标准理论，我们分析了一个类型学系统，在这个系统中，有规律的插入过程以不透明的方式与韵律重音相互作用。这种相互作用的性质，这意味着几个复杂的学习决策，被证明支持以下结论相关的一般机制的语言学习：1。在词汇习得过程中，表层语音结构直接融入词汇（有时称为“身份图”），导致学习者致力于过度生成的语法。2.词汇表征（LR）的某些方面的收购可能会发生在没有形态音位交替;收购的LR是不相同的表面形式是必要的，以解决过度生成的问题。3.过度生成的问题可能会出现在学习约束的相互作用，是不同于那些存在于忠诚的约束，站在一个特殊/一般的关系。1. 1.1子集问题对词汇习得的影响形式学习理论中一个常见的问题是子集问题（Angluin，1980）和（Baker，1979）。语言学理论提供了逻辑结构丰富的类型学：例如，某些语法产生的结构库存是同一类型空间中其他语法库存的真子集。由于这些库存之间的关系，在学习中可用的语言输入可能是一致的一个以上的语法。这使得学习者有可能选择一种语法，这种语法具有比必要的更大的生成能力，同时与可用的数据保持一致。此外，学习是在积极证据的基础上进行的，因此，如果学习者确实假设了一种过度生成的语法，那么再多的额外证据也不会与这种错误的语法相矛盾。因此，子集问题的方法提供了在给定可用数据的情况下计算最严格语法的方法。早期学习OT模型中子集问题的标准分析（即，在形态分析之前）涉及对标记约束支配忠实约束的语法放置归纳偏见。这种偏好可以用不同的方式来尝试，或者作为“初始状态”的条件，（（Gnanadesikan，to appear），（海耶斯，出现），（Levelt，1995年），（Smolensky，1996）），或作为一个持续的原则，坚持整个语法学习（普林斯和特萨尔，出现），但 * 这份报告大大受益于与格雷厄姆霍伍德，I-Ju桑德拉赖，西谷浩一，和艾伦普林斯的谈话。它得到了NSF BCS-0083101和授予罗格斯认知科学中心的NIH NRSA培训补助金（NIH 1-T32-MH-19975-05）的部分支持。如果有任何错误仍然存在，尽管有这些帮助，我们独自负责。
The purpose of this report is to contribute to formal learning theory in Optimality Theory by providing an analysis of the problem posed by covert phonological interaction. Building on standard theories of syllable structure and metrical stress, we analyze a typological system in which regular processes of epenthesis interact in non-transparent ways with metrical stress. The nature of this interaction, which implies several complex learning decisions, is shown to support the following conclusions relevant to general mechanisms of language learning: 1. A process of lexical acquisition in which surface phonological structure is directly incorporated into the lexicon (sometimes called ‘the identity map’) leads to a state in which the learner is committed to a grammar that over-generates. 2. Acquisition of some aspects of a lexical representation (LR) may take place in the absence of morpho-phonemic alternations; acquisition of LRs that are not identical to the surface form is necessary to solve the over-generation problem. 3. Over-generation problems may arise in learning from constraint interactions that are distinct from those that exist among faithfulness constraints that stand in a special/general relation. 1. Learning overt versus covert phonological interaction 1.1 Implications of the Subset Problem for lexical acquisition A familiar problem in formal learning theory is the Subset Problem ((Angluin, 1980) and (Baker, 1979)). Linguistic theory provides typologies that are rich in logical structure: for example, the structural inventories produced by some grammars are a proper subset of the inventories of other grammars in the same typological space. Because of these relations between inventories, the linguistic input available in learning may be consistent with more than one grammar. This makes it possible for a learner to select a grammar that has greater generative capacity than necessary, while remaining consistent with the available data. Moreover, learning takes place on the basis of positive evidence, so if a learner does postulate a grammar that over-generates, no amount of additional evidence will be inconsistent with this incorrect grammar. Approaches to the Subset Problem therefore provide methods for computing the most restrictive grammar given the available data. The standard analysis of the Subset Problem in OT models of early learning (i.e., before morphological analysis) involves placing an inductive bias for grammars in which markedness constraints dominate faithfulness constraints. This preference can be attempted in different ways, either as a condition on the ‘initial state’ ((Gnanadesikan, to appear), (Hayes, to appear), (Levelt, 1995), (Smolensky, 1996)), or as a durative principle that persists throughout grammar learning (Prince and Tesar, to appear), but the * This report has benefited greatly from conversations with Graham Horwood, I-Ju Sandra Lai, Koichi Nishitani, and Alan Prince. It is supported in part by NSF BCS-0083101 and a NIH NRSA training grant awarded to the Rutgers Center for Cognitive Science (NIH 1-T32-MH-19975-05). If any errors remain, despite this help, we alone are responsible.