乔姆斯基的语言学理论认为，只有假设人类先天便具备专门为学习语言而定制的神经组织，我们才有可能解释婴幼儿令人惊叹的语言学习能力。在婴幼儿语言习得研究中，音位范畴的习得是十分重要的研究课题。本研究采用神经计算建模的手段研究并探讨了婴幼儿习得语言过程中对语音和语义信息的加工，对相应知识的存储以及建立母语音位范畴的机制。

本研究以语言模式的二重性、知觉重组、功能重组、范畴学习、统计学习、社交学习等重要理论为理论依据，以大脑皮层中语言功能区的神经功能、各功能区之间的神经链接等重要生理构造和神经机能为生理依据，以模块化理论和联结主义理论为模型设计依据，在Kröger语言处理模型现有框架的基础上，分别提出了联结可扩展的自组织神经网络模型（I-GSOM）和基于语言模式二重性的网络模型（DI-GSOM）。针对语言习得的特点，我们为模型设计了改进的可扩展自组织网络结构及相应的学习算法，模拟了以上所列出的学习机制。

本研究共包括两个模拟实验。第一个实验是基于I-GSOM模型对标准德语音位习得的模拟。该模拟实验证明了本研究提出的改进的可扩展自组织网络的自组织能力可以帮助模型对听觉范畴和语义范畴进行归类，网络的动态扩展性可以帮助模型应对学习过程中知识的增长。虽然模型有很高的抽象性，但是已经足以描述基本的神经系统原理，并且可以模拟自组织、关联学习、自适应、神经可塑性等生理机制。第二个实验是基于 DI-GSOM 模型对汉语普通话音位习得的模拟。实验分别模拟了婴幼儿对语言与生俱来的普遍感知能力、12—18个月年龄段婴幼儿对音位范畴的知觉重组过程以及19—36个月年龄段婴幼儿对母语音位感知的完善过程。通过对语言模式二重性的模拟并引入基于语义的由高层到底层的加工机制，我们验证了知觉重组和功能重组现象。模拟实验结果表明，在习得语言过程中，婴幼儿由语音处理感知向音位处理感知的转变并不是一个由单一因素（如语音学习）所决定的过程。婴幼儿对母语音位范畴的习得以及对母语音系知识的建立至少需要语义知识学习和语音知识学习的共同参与。语义加工对形成正确的母语音位范畴感知发挥着至关重要的作用。

基于语言模式的二重性理论，本研究提出的语言习得模型综合考虑了婴幼儿习得语言过程中语音和语义学习的交互作用，并对语音信息的加工进行了细化。改进的可扩展自组织网络结构及知识学习算法可以较为真实地模拟复杂的学习机制，在语言习得的算法设计上有所突破。此外，模型对声调范畴习得的建模研究具有汉语普通话的研究特色。本研究是结合语言学、心理语言学及神经语言学的跨学科研究，具有开创性，可以填补相关研究的空白。

关键词：婴幼儿语言习得；音位范畴化；语言模式的二重性；神经计算建模；可扩展的自组织网络

Abstract

Chomsky's linguistic theory states that only by supposing that human has some innate neural structure built specifically for language learning tasks,could we explain the impressive language learning abilities of infants.Phoneme acquisition is an essential topic among all studies focusing on infant language acquisition.In this study,by using neuro-computational modeling methods,we have explored and discussed the mechanisms of how infants process phonetic and semantic information,store those knowledge and build phoneme categories of their native languages.

Based on theories of Duality of Patterning,Perception Reorganization,Functional Reorganization,Categorical Learning,Statistical Learning and Social Learning,and biological structures and neural functions of language functional cortices and synaptic links,and modeling principles of Modular Theory and Connectionism,we have proposed the Interconnected Growing Self-Organizing Map（I-GSOM）model and the Duality-based Interconnected Growing Self-Organizing Map（DI-GSOM）model,on the foundation of Kröger's language processing model.According to the characteristics of language acquisition task,we have also developed the improved growing self-organizing map structure and corresponding learning algorithms,and further simulated those learning mechanisms mentioned above.

In this study,we have conducted two simulation experiments.In the first experiment,we simulated the phoneme acquisition process of Standard German based on the I-GSOM model.Modeling results prove that the improved growing self-organizing map we proposed has the self-organizing ability to cluster auditory and semantic categories,and its network structure can help the model to handle the knowledge growth during the learning process.Although the model is highly abstract,it is plausible enough to describe basic neural principles,and model biological mechanisms such as self-organizing,associative learning,self-adaptation and neural plasticity.In the second experiment,we simulated the phoneme acquisition process of Standard Chinese based on the DI-GSOM model.We have respectively simulated infants' innate universal language perceiving ability,the perception reorganization process of infants between 12 and 18 months old,and the refining process of native phonemic perception of infants between 19 and 36 months old.By modeling the Duality of Patterning and the semanticbased top-down process,we have verified the phenomenon of perception reorganization and functional reorganization.The modeling results show that the transformation of infants' perception pattern from phonetic perception to phonemic perception is not a single-factordetermined process.Infant's acquisition of native language phoneme categories and native language phonology needs,at least,the cooperation of semantic learning and phonetic learning.Therefore,semantic processing is essential for acquiring correct phonemic perception ability of infants' native language.

Based on the Duality of Patterning theory,the two models we proposed in this study have considered the interactive effects of phonetic learning and semantic learning during infants' language acquisition process,and further detailed the modeling of phonetic processing.The improved growing self-organizing map structure and corresponding learning algorithms are capable of simulating complex learning mechanisms.Therefore,this study is an important breakthrough on language acquisition algorithm design.Moreover,the modeling of tonal acquisition shows the characteristic of Standard Chinese study.As a cross-filed study that covers fields including linguistics,psycholinguistics and neurolinguistics,our work is innovative and fills the gaps of related research areas.

Key Words: Infants' language acquisition,phoneme categorization,duality of patterning,neuro-computational modeling,growing self-organizing map