Based On BERT-BiLSTM-CNN Multi-feature Fusion Research And Application In Public Opinion Analysis Of Three-child Policy

In order to effectively slow down the aging process,The State Council introduced a three-child policy to encourage births,but the expected results are unsatisfactory.In order to clarify the will and attitude of social groups towards having a third child,emotion analysis technology is used to excavate the emotional value of public opinion texts of the three-child policy on Weibo platform,assist government functional departments to form correct public opinion guidance and improve corresponding supporting measures,and help enterprises to excavate target customers and explore new markets.Therefore,this paper constructs multi-feature fusion technology based on deep learning model to realize emotion analysis of public opinion text of three-child policy.The main research work is as follows:(1)As traditional short text sentiment analysis ignores the important emotional value conveyed by non-literal elements,this paper forms a multi-feature fusion matrix by vector conversion of frequently used Emoji and popular phrase abbreviations in Chinese context,and conducts experiments on deep neural network model BiLSTM-CNN and its benchmark model.The experimental results show that,The effect of emotion analysis is better than that of traditional experiment.(2)Considering the sparse feature of the short text,the sentence meaning is not sufficient,resulting in low accuracy of sentiment analysis.In this paper,a hybrid neural network BERT-BiLSTM-CNN model is constructed.The pretraining model BERT is dynamically encoded to fully extract the meaning of statements.BiLSTMCNN models the text,and the model extracts important text information from the local and the whole dimensions.Compared with the traditional neural network models CNN,BiLSTM and BERT fine-tuning,the classification performance is better.The final experimental results of BERT-BiLSTM-CNN model show that the accuracy rate,accuracy rate and recall rate respectively reach 88.2%,88.9% and87.7%.In order to further explore the topic distribution of public opinion on the three-child policy,a BERT-LDA model is constructed to carry out text topic mining.The input text is preliminary-processed and converted by BERT,and the matrix vector similarity is calculated.The suitable similarity threshold is selected to filter out similar statements,so that the data input to the LDA model is of higher quality.To improve the ability of LDA topic modeling.Through the theme modeling of different emotional tendencies,the influential factors of positive and negative emotions on the three-child policy are analyzed,and reasonable suggestions are given.(3)Build a multi-feature integrated emotion analysis and theme mining system to realize public opinion analysis and emotion mining of the three-child policy.The core functions include data preprocessing,model training,theme visualization and other modules to visually display the theme distribution of different emotional tendencies.The research in this paper provides convenient analysis tools for the effective implementation of the three-child policy.