| 摘要 |
The present invention relates to a Korean dependency parsing method using deep learning. According to the present invention, the Korean dependency parsing method includes the steps of: initializing the state of a document containing Korean; classifying states of initialized documents based on deep learning; and performing the Korean dependency parsing based on the classified states of the documents. The Korean dependency parsing using the deep learning is performed by containing an artificial natural network of an input layer, a projection layer, a hidden layer, and an output layer. The input layer is configured through a one-hot scheme that relevant words or features are expressed as 1, and remaining words or features are expressed as 0 among words and features extracted based on a preset state of the document. The projection layer is to reduce a dimension in the word expression of a natural network language model (NNLM) in relation to the input layer through pre-training. The hidden layer is to perform non-linear activation with respect to the projection layer. The output layer is to find out a probability for each state. According to Korean dependency parsing method of the present invention, the parsing is performed by using the deep learning representing excellent performance, so that feature design and tuning time can be reduced as compared to existing methods. |
