ニューラル・ネットワーク・モデルによる日本語文節の識別の試み

概要

論文の詳細を見る
Recognizing that clauses in Japanese tend to end with a relatively limited number of letters, this study attempted to recognize ends of clauses in written Japanese by a neural network model with back propagation. Focusing on ten letters which are most likely to be used as a last letter in Japanese clauses, this study set up a neural network model to learn to tell whether or not an encountered letter (target letter) was used as an ending letter of a clause. Although only a partial tests were attempted due to the relatively small sizes of the data sets, the obtained results were quite encouraging. Based on the findings, this study suggests four points of further improvement: (1) an enlargement of volume of data set, (2) an improvement of data generation method, (3) an examination of cases which were not properly recognized, and (4) a possible use of a certain type of dictionary as well as an inclusion of grammatical rules.
放送大学の論文