Research On The Sentiment Tendency Of Internet Public Opinion Based On Deep Learning

The main research content of this article is the sentiment analysis of Internet information based on deep learning.In the current society,the rapid dissemination of public opinion information will have a certain impact on the current social stability.Therefore,it is beneficial to study the sentiment orientation in the Internet.The urgent state of the social situation is controlled.Therefore,this paper mainly makes some researches on the four aspects of the extraction of public opinion information,the representation method of this article,the feature extraction and the sentiment orientation calculation.The above mentioned,there are following contents:(1)This article collects website links through the use of understanding websites,user agents,site maps,crawl delays,and various crawling strategies.The use of regular expressions,Beautiful Soup and lxml these three methods for data capture,you can finally get the lyric classification text.(2)For the word distributed representation technique,this paper uses CBOW and CWE two word embedding models for comparison experiments.These two models belong to the vector space model which can convert words into continuous value vector expressions,at the same time can be from a large number of unlabeled unsupervised learning of word vectors and solving matrix sparsity problems in plain text data.(3)For the feature extraction technology,two improved language models are proposed in this paper.They are a language model based on bidirectional modified long-short memory neural network and a language model based on deep simplified gated unit.Among them,the long-short memory neural network structure based on two-way variant long-short memory neural network adopts three variants: peephole memory structure,dynamic cortical memory structure and double-door coupling structure.The traditional long-term memory neural network structure can be better optimized,and the long-range dependence problem can be solved by combining the two-way neural network structure.A language model based on a deep simplified gating unit can simplify the parameters while using the memory structure to store the context information.In order to deepen the training of the structure in this article,a deep loop network structure is used to ensure better parameter values while extracting better feature values.(4)For the calculation of the final sentimentality,it is transformed into the problem of feature vector classification.In this paper,we use the classification method based on reverse independent self-encoder and the classification method based on stack self-coding neural network in the selection of classifiers.Experimental results show: In the five comparison experiments,the best classification effect can be achieved based on the bidirectional double-gated memory language model and the deep simplified gated single language model.The CWE model takes advantage of the combination of internal characters and external contexts,and can be easily integrated into a word-embedding model.A bidirectional cyclic network can use past and future data from an input in the sequential data,so CWE can achieve more than CBOW.A good word vector expresses the effect.The three types of long and short-term memory structures can be obtained through experiments.The dual-gated structure can achieve better feature extraction effect while achieving structural optimization.The bidirectional cyclic network can solve the problem of long-term dependence and is superior to the traditional single-cycle network structure.