| Named entity recognition is one of the basic tasks in natural language processing,and its recognition accuracy plays an important role in information retrieval,question answering system,machine translation and other downstream tasks.At present,the named entity recognition model based on deep learning is the mainstream recognition model.Compared with traditional methods,the recognition effect of this model has been greatly improved.However,the named entity recognition model based on deep learning also has shortcomings.Most studies only use a single word feature for model training,ignoring the dependency relationship between words and the importance of other semantic features to enhance the recognition ability of the model.According to the above problems,this paper will be improved from the following points:(1)The named entities are identified by neural network structures fused with dependent syntactic information.The dependency parsing based on word level is transformed into dependency parsing based on character so that the model can obtain more additional information.The graph convolutional network is added to the bidirectional long short-term memory network as an interaction layer to capture the interactive information between words.Exponential linear unit function is used as the activation function of the interaction layer,which has superior performance compared with Sigmoid and Rectified Linear Unit activation functions.(2)A named entity recognition model based on attention mechanism is proposed.By constructing a multi-feature named entity recognition model,the initial character vector,the length vector of each word and the position vector of each character in the word are connected as input extraction features of Convolutional Neural Networks.The effect of named entity recognition is improved by combining affix features with attention mechanism and part-of-speech features learned from multitasking.Compared with conditional random field,semi-Markov conditional random field is selected for decoding,which further improves the performance of the model.(3)When relying on trees to capture long distance structured information,a certain percentage of errors usually occur.To solve this problem,this paper proposes a graph convolutional neural network model guided by attention and edge tags.Firstly,the model introduces the edge-aware node joint update module and edge-aware node update module to solve the wrong dependent label information to a certain extent.Then,the attention guide graph convolutional network layer is added after these two modules,which can automatically learn how to selectively focus on relevant structured information,and the accuracy of named entity recognition is further improved. |
