The Realization Of The Tendency Analysis Method Of Network Comments Based On Deep Learning

Internet public opinion refers to the sum of different opinions held by Internet users on hot topics or certain real-time news and the comments posted online.Nowadays,whether on social platforms or news websites,a large amount of public opinion information is generated every day,but due to the looser network supervision,more and more people will arbitrarily publish their ideas on various platforms of the network,some of which are extremely Sexual language is easy to guide the wrong direction of public opinion.Therefore,an efficient text tendency analysis method is needed to process the information on the Internet and send out public opinion warnings in a timely manner.This plays a vital role in creating a good network public opinion environment and ensuring the security of my country's network public opinion environment.In order to study how to efficiently vectorize the text information.This paper optimizes the BERT model of the text language representation model,optimizes the parameters of the position embedding part of the input layer,only considers the relative position between words,to a certain extent,reduces the training parameters of the model and shortens the training time.Improve the generalization of the model.In order to more accurately classify the text information of online reviews.This paper first proposes a multi-level feature extraction model based on word level,sentence level,and context level.Through feature stitching fusion,the loss of data information is avoided to the greatest extent;then a label tree classification method based on the multilabel principle is designed.A Softmax classifier is placed at each root node,and sub-labels are obtained through the training of the training data set,and the labels are predicted by the sub-label probability to form a label tree,that is,the text sentiment classification problem is converted into a path search on the label tree problem.