Chinese Sentiment Analysis With Convolution Neural Network

With the rapid development of Internet-related technologies,users are free to express their opinions on the Internet platform on new self-media represented by Weibo and Taobao.The review point of view is not only a reference for users to purchase products or discuss current events,but also an important feedback for collecting public opinion information.From this we can see that the tendency to tap the mass of commentary texts and analyze the polarity of emotions can promote the needs of all parties and have important research value for social development.Sentiment analysis has also become an important research topic in the field of natural language processing.The traditional sentiment analysis methods mainly use manual annotation,which makes it easy to automatically classify texts and extract more accurate features.In recent years,machine learning,deep learning related technologies have brought new opportunities for natural language processing.Words as the most basic unit of study,the quality of word vector training is crucial to the accuracy of task completion.The traditional word vector only considers the semantic information in the word,and the convolutional neural network does not consider the structure of the sentence and the importance of the sentence.This paper focuses on the word vector training method and the Chinese emotion analysis related technology based on the convolutional neural network.Focused on the use of information word vectors in the improved new convolutional neural network sentiment classification method,and achieved certain results,summarized as follows:(1)The word2vec method learns from the context the semantic word vector does not contain the word's own sentiment information,and the traditional convolutional neural network model does not take into account the structure of the sentence.Aiming at the shortcomings of these two aspects,we proposed a dynamic multi-pooling convolutional neural network model based on sentiment word vector.The skip-gram model and sentiment dictionary are used to train the sentiment word vectors,and the sentences are segmented by the transition words.The multi-segmented pool retains the multiple largest eigenvalues of the sentence structure.The experimental result is analyzed according to the Precision,Recall and F1.The results show that the proposed method has better performance than the machine learning model and the traditional convolutional neural network model,and it is proven that the proposed method can effectively analysis the sentiment task.(2)By introducing product-level information into the neural network method to deal with the word vector problem,and according to the different importance of the sentence in the document,we proposed a weighted convolutional neural network(PWCNN)model combining product features.The model first trains the product information word vector based on the vector-based semantic synthesis calculation principle,and then uses a pooled weighted convolutional neural network to learn the document-level representation of the review.In order to prevent over-fitting and improve the generalization ability,the output layer adopted a dropout strategy.Experimental results show that the PWCNN model significantly improves the accuracy of sentiment classification and accelerates the speed of model training in multi-disciplinary product review datasets.