汉字阅读的联结主义模型

A Connectionist Model of Chinese Characters Reading

汉字与英语词汇阅读受到相同统计属性的影响,表现出阅读加工的普遍性,汉字独特的形-音对应关系又体现出阅读的语言特异性。模型1建构与英文阅读模型完全相同的计算机模型,设计汉字的字形与语音表征方案,成功模拟出汉字阅读的规则性、一致性效应及其与频率的交互作用,得到与行为实验相同的结果模式;模型2改变声旁独立成字时的字形表征,结果规则性效应消失。模拟结果一方面表明汉字与英语词汇阅读可能具有普遍的加工机制,都是对输入语料的形-音对应关系统计学习的结果;另一方面表明输入语料的不同统计属性可能是汉字阅读的语言特异性来源。

A connectionist model regards reading as statistical learning acquired from an input corpus. The regularity and consistency effects in mapping from orthography to phonology are regarded as the statistical properties of word reading. These effects originate from the impact of families at different levels of orthography-phonology mapping in English. On the other hand, in Chinese, only the consistency effect originates from the impact of phonetic families, while the regularity effect emanates from the impact of the phonetics. In this paper, we trained two connectionist models to understand the nature of the universal （consistency） and specific （regularity） effects of reading Chinese characters in a general framework. In Simulation 1, a connectionist model adopted from English was trained to understand the mapping of 4,468 Chinese characters from orthography to phonology. The naming accuracy and sum squared error （SSE） were measured to test the model＇s performance. Simulation 2 was trained to test the nature of the regularity effect by modifying the orthographic representation of phonetics, which appeared as an independent character. The model in Simulation 1 successfully captured the regularity and consistency effects and their interaction with frequency. The results revealed a pattern similar to that found in previous empirical research. Following the occlusion of orthographic similarity in Simulation 2, the disappearance of the regularity effect revealed that the nature of Chinese regularity is the result of similar orthography-phonology mapping between phonograms and their phonetics. The universal-language （consistency） and Chinese-specific （regularity） effects were simulated and understood in combination, through statistical learning acquired from orthography-phonology mapping in a general framework. It can be concluded that the mechanism of reading is universal across languages and a language-specific mechanism arises from the different statistical properties of the input corpus.