Research On Sentiment Analysis Model Of Movie Reviews Based On Further Pre-training And Feature Fusion

According to the text,sentiment analysis is to analyze its emotional characteristics and judge its emotional classification.Sentiment analysis has a wide range of uses in public opinion analysis,human-computer interaction,etc.Hence,sentiment analysis becomes a popular research topic in the field of natural language processing.With the rapid development of the Internet,Internet information has entered the era of big data.Countless social media platforms are flooded with online comments of all kinds.Among the online reviews,movie reviews are rich in emotion and have their own rating labels,making them a good research object in deep learning.However,movie reviews often contain fewer words,with a large number of Internet terms,and without formal grammatical structures,which puts forward higher requirements for the extraction and analysis of semantic features of the model.This paper selects short texts such as movie reviews as the research object,and conducts research on sentence-level sentiment analysis with the help of feature fusion and further pretraining.Aiming at the sentiment analysis of movie reviews,a sentiment analysis model based on further pre-training and feature fusion is proposed.The specific works include that(1)A further pre-training method for film reviews is constructed.After obtaining the sentiment dictionary based on movie reviews using the point mutual information algorithm,a mask method for sentiment words is added to the further pretrained mask language model.With the utilization of prompt learning,a variety of discrete templates are constructed manually,so that the pretraining model can be integrated with emotional knowledge;(2)A sentiment analysis method combining with emoticon vectors is designed.Through data augmentation,we use existing corpus to generate annotated movie reviews with emojis to expand the dataset.The emoticon vector generated by random initialization is spliced with the output vector of the BERT pre-training model,and input to the downstream model training to realize the method of using emoticon features to improve the effect of sentiment analysis;(3)A hybrid model combining with deep pyramid convolutional neural network and doublelayer gated recurrent network is built.Using the method of feature fusion,combining the advantages of convolutional neural network and recurrent neural network,the feature vector semantic information is richer.The sentiment analysis model is experimented on the self-built movie review dataset and the standard sentiment analysis dataset,Chn Senti Corp and NLPCC2014.The experiment results verify the effectiveness of the further pre-training and feature fusion.The further pre-trained model integrated with sentiment has a 0.5% improvement in the classification accuracy on the self-built movie review dataset.The classification accuracy of using features combined with text and Emoji improves by 0.5%.Compared with the commonly used Bi GRU deep learning model,the DPCNN-Bi GRU combination model proposed in this paper has an improvement of 1.3% on the self-built corpus,while on the standard sentiment analysis datasets Chn Senti Corp and NLPCC2014,the classification accuracy is improved by 0.7%.