Research On Sentiment Analysis Based On Pre-LN Transformer

Sentiment analysis refers to the process of recognizing and analyzing sentiment tendencies of subjective texts with emotional colors.A large number of online reviews are generated all the time on the Internet,and they mainly exist in the form of single modal text and static multimodal data(text and pictures).Sentiment analysis of users' online reviews is helpful for businesses to formulate marketing strategies,which has a certain commercial value.This paper investigates the research status of sentiment analysis,and conducts research on single modal text sentiment analysis and static multimodal sentiment analysis using the deep learning method.The contents are as follows:For single modal text sentiment analysis,this paper proposes a neural network model based on Pre-LN Transformer to solve the problems of weak parallel computing ability and insufficient remote dependence in existing models.Most of the sentiment classification models can effectively identify the extreme polarity,but can't distinguish the intermediate polarity clearly.This paper reconstructs the VADER lexicon and integrates the sentiment vectors extracted from the emotion lexicon into the model,which further improves the ability to identify detailed sentiment.Finally,this paper performs sentiment classification tasks on two publicly available datasets.Compared with other models,experimental results show that our model achieves state-of-art performance on Amazon and Yelp datasets.Based on the research of single modal text sentiment analysis,this paper proposes a static multimodal sentiment classification model based on Pre-LN Transformer.The model considers the feature learning of subspace within the modal and the interaction between modals,which combines the multi-headed attention mechanism in Pre-LN Transformer with visual aspect attention to improve the sentiment classification of online reviews.Finally,the effectiveness and feasibility of the proposed method are verified by comparative experiment and internal ablation experiment.This paper also explores the influence of attention space dimension on model performance,which improves the accuracy of model classification while ensuring computational efficiency.