Target-dependent Sentiment Analysis Based On Deep Learning

With the continuous development of computer science and large-scale popularization of mobile devices,people exchange views and express their feelings on social networks.Huge amounts of text information are accumulated.The use of NLP techniques to mine sentimental information has received extensive attention from academia and industry.In recent years,with the development of deep learning technology,people have applied it to sentiment analysis tasks,and achieved remarkable success.However,the technology still has many shortcomings in sentiment analysis tasks.For example,deep learning can hardly extracts sentimental features from short texts and identify sentimental polarities of different targets in the same sentence.In order to solve the problems above,this thesis conducts research on deep neural network and text mining technology and designs three highly effective deep neural network sentiment analysis models through improving the model architecture and training methods:We propose a convolution neural network(CNN)model based on multiple features.The model combines the word vector information of text,sentimental information of words and position information of words to construct the input matrix of CNN,and adjust the importance of different features through controlling the corresponding weights during the training process.We propose a deep network model combined with convolution neural network and long short-term memory(LSTM)network.The model uses the local features extracted by CNN as the attention input of the LSTM network,making the model highly concerned with the target's own information and the word information closely related to the target during the training process.We propose a hierarchical LSTM network model,which combines the regional level of attention mechanism.The model used the LSTM of regional layer to minethe target's attention information,and to obtain the long-distance information of the target combining the LSTM of sentence layer.Finally,we run experiments on datasets with multiple targets and compare the experimental results with multiple classification models which have achieved breakthrough results in the related target sentiment analysis tasks.The accuracies of classification and training performance verify the effectiveness of three models proposed in this thesis in target-dependent sentiment analysis tasks.