汉语句子格角色分配的一种神经网络方法

提出了一个基于分布式表征的计算模型，通过并行分布加工方式完成六类汉语句子的格角色分配任务。模型是一个四层的前传网络，包括输入层（词的分布式表征层），两个隐层，输出层（格角色层）；其中第一隐层的一部分反馈到输入层。模型采用误差反传算法，通过提供学习样本和目标输出，不断调整三个权值矩阵，使得网络稳定时能得到正确的结果。经过训练后的网络具有一定的稳定性和鲁棒性。还对这种方法与传统的符号处理方法作了比较和分析。

ASSIGNING CASE ROLE OF CHINESE SENTENCES WITH A NEURAL NETWORK

According to PDP theory, we tried to use a calculating model based on distributed representation to complete the task of case role assignment of Chinese sentences by parallel processing. There were six types of sentences.The model was a four-layer forward neural network: In input layer, we used distributed representaton to collect syntactic,semantic information(of the word in Chinese sentence) and context information,and there were two hidden layers, and output layer(case role layer). One part of the first hidden layer was feedback to the input layer. Error back propagation learning algorithm Was used to adjust three weight matrices sequently according to learning samples and target output in order to get correct answers When the network was stable, After training, the network was somewhat robust. In addition, the neural network method and the traditional symbol processing method used in natural language understanding was compared and analyzed.