With the rapid development of the Internet,massive data,which contains enor-mous value,has been produced,and over 80%of them is text.It is essential to find a way to deal with these text data automatically.Named entity recognition is a valuable technology to identify key entities from the text.However,there exist some defects in current researches of named entity recognition,including the weakness of the input fea-tures,the unreason of the label sequence.This paper has conducted a series of research on these issues,and the main contributions are as follows:1.Since the input features can't represent the characteristics of the input data well,a named entity recognition method based on topic model is designed.The method trains topic model to obtain the word-topic distribution which assists the named entity recognition training.On one hand,it can solve the problem of lacking global features.On the other hand,as a result of introducing auxiliary features,the model can be initialized better,thereby reducing the early training time.The experimental results show that the improved model can achieve better performance.2.To solve the unreasonable label sequence,this paper proposes a named entity recog-nition method based on the convolutional neural network.This method takes advan-tages of the local connection and weight sharing in convolutional neural network to obtain local features,which further learns the inheritance relationship of the output labels in the model.Obtaining local features makes the proposed model more rea-sonable,and experiments show that the performance of the propsed model achieves further improvements. |