摘要
针对藏文字校对问题,提出一种不使用藏字字典的联合二层BiLSTM模型和CNN模型展开学习的藏字校对模型。模型通过研究藏文字构字语法、字母训练,得到藏文字母的向量表示,对字母向量进行BiGRU学习,所得的特征向量用三个卷积核进行CNN和全连接运算,最后用最小化交叉熵来优化模型查检藏文字的正确性。为了验证方法的实际表现,建立一共73155个藏文字实验语料,其中正样本占55.1%,负样本占44.9%。实验表明,该方法对藏文字对错识别率的F值达94.06%。
To solve the problem of Tibetan words proofreading,a Tibetan character proofreading model is proposed,which combines the two-layer BiLSTM model and CNN model without using Tibetan dictionary.The model obtains the vector representation of Tibetan letters by studying the grammar and letter training of Tibetan characters,learns the letter vectors by BiGRU,and performs CNN and full join operations on the obtained feature vectors with three convolution kernels and the model is eventually optimized to check the correctness of Tibetan characters by minimizing cross entropy.In order to verify the actual performance of the method,a total of 73,155 Tibetan experimental corpora are established,of which positive samples account for 55.1%and negative samples account for 44.9%.Experiments show that the F value of the recognition rate of Tibetan characters is 94.06%.
作者
安见才让
Anjian-cairang(The Computer College of Qinghai Nationalities University,Xining 810007,China)
出处
《微处理机》
2021年第2期41-44,共4页
Microprocessors
基金
国家自然科学基金项目(61862054)
青海省应用基础研究项目(2019-ZJ-7066)资助。
关键词
藏文字
神经网络模型
校对
Tibetan words
Neural network model
Proofreading
