Research And Application Of Public Opinion Monitoring Of Food Safety Comments Based On Deep Learnin

With the continuous popularization and development of Internet technology,the Internet has brought convenience to everyone,and at the same time has become one of the channels for Internet users to vent their emotions.This has also led to a variety of public opinion that spreads rapidly and is difficult to distinguish between true and false.problem.In today's rich life,food safety has always been a hot issue that attracted much attention.People expect a food safety supervision channel and platform that can help find and avoid some counterfeit and shoddy foods on the market.Therefore,the Internet has also,become the main front of food safety public opinion.Some netizens have vented inappropriate emotions on the internet platform,and at the same time triggered the emotional resonance of other netizens on the internet,it may develop into a social public event.Social panics,such as the early "heart eggs" incident,to the later"ditch oil" incident,etc.Since public opinion will affect people's lives,any improper handling may cause serious consequences.Typical problems such as a pusher on the Internet will drive the direction of public opinion.They will induce or change according to the current event development.Netizens are concerned that if there are unspeakable purposes,it may cause inestimable losses to society and the country,so if they can accurately and timely grasp the guidance of public opinion and the time point of public opinion outbreak,it will be able to restore more loss.Therefore,the research on Internet public opinion has great significance and value.There are many indicators for analyzing the development of Internet public opinion,such as clicks,reposts,and activity.Among them,online reviews are often used as a powerful reference for judging and analyzing the development of public opinion.This article mainly predicts the development of public opinion through the analysis of the subject matter of the review and the study of the sentiment tendency.Combining deep learning methods can accurately extract the emotional polarity and topic summary of each review,so that not only the main concerns of public opinion can be known The focus of the problem can also better judge the attitude of netizens to the current public opinion so that they can timely and effectively manage the possible impact of public opinion.The research contents of this paper are as follows:(1)Preprocess the original text data.Clean the data set,replace characters such as punctuation and other useless characters,remove unavailable data regulations,convert food scores to emotional polarity,clean the summary data set,and turn the data set into a word embedding vector matrix.(2)In-depth study of ELMo model and improvement of ELMo's downstream task model.This paper combines the two models of ELMo and GRU to study the classification of emotional polarity.The text vector is mainly generated through EMLo pre-training,and the vector is trained through the GRU neural network and passed through the pooling layer and activation function to predict the emotional tendency.The advantage of the ELMo model is that it uses a bidirectional LSTM in the operation,and through the superposition of the text's position vector and syntactic vector and the original word vector,a new text space representation is obtained.At the same time,the space vector of each word under different semantics is also distinguished.Based on the word vector output by the ELMo model,it is trained by the GRU neural network and finally predicts the emotional polarity.The experimental results show that compared with GRU+wod2vec,the accuracy rate of this experimental model has increased by about 14%.(3)On the premise of using the word vector of ELMo training output as the input layer,this paper improves the structure of the encoder and decoder in the Encoder-Decoder model.In the encoder,this paper borrows the ByteNet network and uses hole convolution to gradually expand the receptive field of the encoder when doing information compression,so that the encoder can better consider and compress the full-text semantics.A network structure of GRU overlay is proposed in the decoder,and an Attention mechanism is added to each GRU layer so that the decoder can better focus on the key prediction information of each layer.The experimental results prove that the proposed model improves the subject summary of the review text by 9%and 3%respectively.(4)Implementation of public opinion monitoring and analysis system.By using the two models in this paper,the models are individually called and put forward to implement the public opinion monitoring and analysis system.The system mainly includes queries of public opinion information,summary of public opinion by date,and display of public opinion trend charts.