Sentiment Analysis Of Chinese Reviews Based On Multi-feature Fusion And LSTM Neural Network

With the rapid development of mobile Internet,online shopping has become a part of people's daily life.There are a lot of product reviews on e-commerce;Mining the emotional tendency of these reviews not only can provide information of goods for seller,but also make for consumer to make objective judgments on goods and make decision about purchase.For a large amount of reviews,it's time consuming to mining emotional tendency of reviews by browsing,so how to use the technology of artificial intelligence to analyze the sentiment of product reviews automatically is an important and meaningful topic.The research of sentiment analysis mainly based on template,machine learning and deep neural network.With the development of big data and the diversification of language forms,deep neural network has become the hot technology in the field of nature language processing,it's also well used in sentiment analysis,this paper mainly studies sentiment analysis based on deep neural network.The mainly works of this paper are as follows:(1)To solve the problems of high dimensions and semantic unrelated about text representation,the word embedding mechanism is used.First the neural network language model is used to train a large number of reviews,and then reviews can be represented as a low-dimensional vector by the form of distributed representation.This text representation method also contains semantic information,which is beneficial to the neural network model to better understand the text.(2)Considering the particularity of sentiment analysis task,a new method of text representation method,multi-feature word vector,is presented,which is an optimization of the representation method of distributed representation.Considering the influence of the words' emotional elements on the emotional polarity of the text,we capture the words which contains emotional elements by construction the emotional feature dictionary,then represent these words by the method of construting the emotional feature word vector.Finally fuse the emotional feature word vector and the word vector expressed by distributed representation to form the multi-feature word vector.The representation of the multi-feature word vector can capture the emotional information of the text,which is more sui Tab.for the sentiment analysis.(3)Essentially sentiment analysis is a classification task,the calculation of feature weight is an important step in the classification task.In this paper we use the feature weight algorithm to distribute weight to the multi-feature word vector,thus highlighting important words for classification.The text represented method of using word vector contains multi-features based on weight distributing can enriches the expression of text semantics and it is more suiTab.for the neural network model to capture deep-level feature and sentiment analysis for the text to use the word vector contains multi-features based on weight distributing as the input of nerual network.(4)In this paper,the LSTM neural network model is used to capture the deep features of the text.The text are represented by the word vector contains multi-features based on weight distributing,and use these as the input of the LSTM neural network model,then use LSTM neural network to study the sequence characteristics and context dependency.Through the contrastive experiment with the traditional LSTM neural network,the validity of the improved method proposed in this paper can be verified.In the above four work,this paper takes full account of the characteristics of the sentiment analysis task,introduce prior knowledge,such as emotional dictionary resource and the weight information,into neural network model.On this basis,the multi-feature word vector based on weight distributing is presented,which can capture more sui Tab.features for the task of sentiment analysis,and depend on the LSTM neural network model,richer combination of features can be capture,so as to improve the understanding of the sentiment analysis model and the accuracy of the sentiment analysis.