著者

浅川 伸一

出版者

東京女子大学比較文化研究所

雑誌

東京女子大学比較文化研究所紀要 (ISSN:05638186)

巻号頁・発行日

vol.75, pp.1-18, 2014-01-01

---

A neural network model to read aloud proposed. Although two succeeded models\r were proposed so far, some problems still remained unsolved. The problems are the way\r of implementation about the“lookup table”in the dual route cascaded model, and the\r existence of“division of labor”in the triangle model. The model proposed here was\r intended to integrate both models in order to give a solution to these problems. The model\r is consisted of two local perceptrons to deal with information of orthography and\r semantics, and a gating perceptron to adjust outputs of local perceptrons. Introducing a\r Gaussian function and its interpretation, this model can describe contribution of semantics\r clearly. This model can also explain the grapheme-to-phoneme-correspondence rule to\r read regular words, and the way of reading irregular words. According to this model,\r there is no difference between the dual route cascaded model and the triangle model. The\r outward difference between these models can be absorbed in a variance parameter. The\r variance parameter will be adjusted or learned through training many times. Therefore,\r this model can formulate for the problems in which both models could not describe. This\r model can be regarded as an extended and generalised version of previous models which\r are absorbed in this model.