Design And Implementation Of User Comment Sentiment Analysis System Based On Deep Learning

With the continuous improvement and progress of mobile Internet technology,various application platforms with rich functions are springing up like bamboo shoots after a spring rain.While enjoying convenient services,people can express their views on all kinds of things through the Internet.With the rapid growth of the number of Internet users,the scale of data containing users' subjective information on the Internet is also exploding.These subjective information is of great reference significance for individuals to make consumption decisions and businesses to improve their marketing strategies.It is of great research value to use sentiment analysis technology to analyze these massive data and mine users' views from them.In view of this,various methods of sentiment analysis are summarized and compared in this thesis,and each algorithm process in sentiment analysis task is deeply studied.The traditional method is combined with deep learning algorithm to improve the two links of text vectorization representation and feature extraction,and finally a more effective sentiment analysis method is proposed.In the part of text vectorization,this thesis improves the Word2 vec model.Word2 vec model generates word vectors based on context information,but ignores the internal composition of words themselves,and may lead to two words with opposite emotional polarity having relatively similar vector representations due to similar contextual information,thus weakening the semantic relevance that word vectors can represent.Therefore,this thesis designs an accurate word embedding model that combines the internal character semantics of words with traditional emotional dictionaries.Based on the skip-gram model,the AWE model is improved from the following three aspects.Firstly,when training the word vector,the position-based vector representation method is adopted to generate the character enhanced word vector by comprehensively considering the context information of the target word and the semantic information of its component characters.Secondly,taking advantage of the strong ability of traditional sentiment dictionaries for word-level text recognition and classification,the words with similar meanings to the target words are screened out to assist the generation of subsequent precise word vectors.Thirdly,by reducing the distance between the target word and the word vectors of each approximate word,the previously generated character-enhanced word vectors are adjusted to generate the accurate vector representation of the target word.Finally,the experiment proves that the word vector generated by the AWE model can improve the classification performance of the downstream model to some extent.In the link of text feature extraction,this thesis designed a sentiment analysis model based on the Attention Mechanism,Bi GRU network and multi-channel CNN,which is referred to as BMC model.Among them,the Attention Mechanism enables the model to pay attention to the words in the text that have a great impact on sentiment polarity classification.Bi GRU network is able to extract contextual semantic information of long text.Multi-channel CNN extracts different local features by setting convolution kernel of different window sizes,so as to provide richer feature information for subsequent sentiment classification tasks.BMC model uses different networks to extract the corresponding text features and fuse them in the feature fusion layer,and then input them into the Softmax classifier for classification prediction,so as to complete the task of sentiment analysis.In the end,this thesis proved the effectiveness of the processing method combining the Attention Mechanism,Bi GRU network and CNN through experiments.Finally,this thesis designs and implements a sentiment analysis system for user reviews,and applies the sentiment analysis technology to the catering industry to achieve automatic classification of restaurant reviews.