| In recent years,with the rapid development of the Internet and the increasing number of users on the Internet,a large number of texts containing emotional information are emerging.These texts have high value and can be used to control public opinions,recommend commodities,and predict events.Therefore,text sentiment classification has become one of the important issues that researchers pay attention to.Some researchers propose to combine word2 vec and convolutional neural network(convolutional neural network,CNN)for text sentiment classification tasks.Since the texts in the network have a trend of becoming shorter in length and obvious in emotion,this paper improves the above method to improve the classification accuracy.The main work is as follows:Firstly,add emotional information to pre-trained word vectors.According to the fact that the distance of word2 vec vectors can be calculated,the higher the cosine value between two words is,the possibility that they have the same sentiment polarity is higher.Word2 vec is used to expand the classic emotion dictionary.Then word vectors and the emotional vectors are merged to produce the global word vectors.Then,add local information to the global word vectors.TFIDF(term frequency-inverse document frequency)is used to adjust the word vectors from word2 vec.Confidence is used to improve the emotional vectors.Finally,through a two-channel convolutional neural network,a set of word vectors is fine tuned to provide more local information but leave the other word vectors unchanged.More features can be learned from the two groups of word vectors.The feature extraction is carried out by using multiple convolution filters in different regions of word vectors.To further extract the features,the method of max pooling is applied,and the softmax function is used to classify the texts.The above proposed method is verified through theano which is a deep learning framework.The experimental results show that it has better classification effect. |
PDF Full Text Request