Research On Sentiment Analysis Of News Comments Based On Deep Learning

There is a wealth of background knowledge reflected in news comments,which also reveal the subjective preferences of the commenters.Moreover,news comments are constrained by various social and cultural factors such as ideology and values embedded in news reports,making it significant to mine the sentiment orientation in comments for individuals and society.Traditional sentiment analysis methods are based on sentiment lexicons and machine learning,but sentiment lexicon-based approaches require creating sentiment dictionaries that may not cover all domains and text types,while machine learning-based methods often rely on manually extracted features.In contrast,deep learning approaches can automatically learn features from large amounts of data and have scalability and flexibility.Therefore,this paper investigates the application of deep learning in sentiment analysis for news comments,mainly focusing on two aspects:coarse-grained sentiment analysis at the sentence level and fine-grained sentiment analysis at the aspect level.The main contributions of this study are:(1)Based on sentence-level coarse-grained sentiment analysis,considering that different news content has a certain degree of influence on users’ comment sentiment tendencies,this article proposes a model that uses news features to modify comment features.Different levels of news and comment features are obtained through keyword extraction,sequence-to-sequence-based summary generation,etc.Attention mechanism is used to modify the comment features,and the modified comment vector is used for sentiment analysis.Experimental results show that the modified comment features have significant improvements in accuracy,、recall、and F1 value.(2)For aspect-level fine-grained sentiment analysis,this article proposes a sentiment analysis model based on a dual-graph neural network.Effectively characterizing the grammatical relationships between long and short distances through dependency syntax trees and constituent syntax trees.The dependency and constituent syntax trees are modeled through edge-graph convolutional neural networks and graph attention neural networks.Different sources of information are fused through feature fusion,and then news features are used for modification to generate more rich text feature vectors,which are then used to identify aspect categories.(3)After determining the aspect category of the comment text,in order to improve the accuracy of mixed sentiment(both positive and negative sentiment),two graph neural networks are used to process the text,capturing positive and negative sentiment respectively.Then,the attention mechanism is applied to fuse these two feature vectors,and the resulting vector is input into the output layer to output the final sentiment category result.The experiment shows that the model achieves an accuracy of 84.6%and 70.5%for overall sentiment and mixed sentiment,respectively,which is better than other models.