The present invention is to provide a method of improving performance compared to the existing system by solving the problem that the conventional general word embedding vector does not express the characteristics of each language analyzer well by combining the correct answer label distribution vector for each language analyzer. (A) Recognizing a Korean individual name using Bidirectional LSTM CRF as a learning model, (B) Representing a word input using at least one of a pre-learned word embedding vector, a part-of-speech embedding vector, and a syllable-based word embedding vector (C) combining the vector of syllable units forming words using LSTM as the extended word representation, combining the word unit vector and the correct answer label distribution vector for each analyzer, and (D) distribution Converting into a vector using a softmax function, an activation function, into a model. And it characterized in that formed.
展开▼
