Research On Short-texts Sentiment Classification Based On Convolutional Neural Network

Text sentiment classification is an important topic in the field of natural language research.At present,the research of text sentiment classification is mostly based on emotional dictionary and machine learning method.However,when they complete the task of sentiment classification,they separate the relationship between words in the text,ignore the meaning of words in the context,and cannot capture the deep semantic information of the text,and their accuracy of sentiment analysis for the expression of irregular online text needs to be improved.In recent years,research based on deep learning methods has deepened,and various scholars have made many breakthroughs in the field of natural language research by using deep learning techniques.Therefore,this paper proposes to use the deep learning method to judge the emotional categories of short texts on the network platform.The main research work is as follows:(1)Aiming at the problem of fewer Chinese datasets with emotional category labeling,this paper uses a crawler program to crawl 25,000 comment texts from the official website of Jingdong Mall.Meanwhile,it also uses Word2vec tool to train the semantic vector of each word on the basis of Chinese public corpus to solve the problem of how to digitize the text data in the sentiment classification experiment.(2)Aiming at the problem of text representation,this paper constructs the text data of classical traditional machine learning model experiments by using the vector space model representing feature weight with tf-idf and the word vector model trained with word2vec respectively.Through the analysis of the experimental results,it is found that word vector trained with word2vec considering the context can enable the model to learn the emotional information from the text and improve its classification performance.However,the degree of improvement is limited,and it is impossible to break through 90%.This shows that the shallow machine learning method is not enough for complex and deep knowledge of semantic information,and it also explains the necessity of using deep learning method to realize the sentiment classification of network short text.(3)Aiming at the shortcomings of traditional machine learning model and classical convolutional neural network model in text sentiment classification task,this paper proposes a two-channel convolutional neural network model SFD-CNN that integrates emotional features.In the vectorization of experimental text data,the model not only considers the semantic information of word features,but also incorporates its corresponding emotional attributes,so as to obtain more emotional information.Meanwhile,the model also extracts text features from different aspects with the mechanism of two channels.(4)Setting up several comparative experiments to compare the two-channel convolutional neural network model SFD-CNN which integrates emotional features with the convolutional neural network model SF-CNN which integrates emotional features,the two-channel convolutional neural network model D-CNN and the SVM model which has the best effect in the traditional machine learning methods.The experimental results show that the SFD-CNN model has the best effect,with an accuracy of 92.94%,which is 2.19%higher than the original CNN model,and confirm that the research work of this paper has certain significance and contribution to the research of short text emotion classification.