Research On Sentiment Analysis Using Neural Networks

Sentiment analysis is one of the key technologies for realizing human-computer dialogue and emotional interaction,and has been widely used in fields such as answering robots,public opinion guidance,and commodity review analysis.However,due to the complexity of data,the long-term dependence of sentiment,and the heterogeneity between different modal data,existing sentiment analysis methods suffer from some deficiencies,such as insufficient feature extraction,missing of key original information,and lack of correlation modeling between different modal features.To address these problems,sentiment analysis methods for text and image-text are investigated by employing neural networks,respectively.In the aspect of text sentiment analysis,a text sentiment analysis model based on attention mechanisms and multi-channel neural networks is studied to address the problems of insufficient feature extraction and missing of key original information.Firstly,a multi-channel feature extraction module is constructed by combining dilated convolutional neural networks with different dilated rates,bi-directional long short-term memory networks and self-attention mechanisms to obtain multi-scale high-level context information.Secondly,in order to reduce the deviation and omission of features,an original feature extraction channel is designed to capture the original context features.Then,a global attention mechanism is used to highlight the sentiment-related features in the global features.Finally,some experiments are conducted on real datasets(NLPCC2017-ECGC dataset and Chn Senti Corp-Htl-unba-10000 dataset),and the corresponding results show that the proposed model respectively achieves 97.9% and93.8% classification accuracy on these two datasets,which outperforms the existing models.In terms of image-text sentiment analysis,in view of the problem of insufficient correlation modeling between different modal features,considering the complementarity,difference,and consistency between different modal features,an image-text sentiment analysis model based on interactive fusion neural networks is proposed.It mainly consists of four parts: text feature extraction module,visual feature extraction module,interactive fusion module,and sentiment classification module.Specifically,the text feature extraction module mainly includes convolutional neural networks,bi-directional long short-term memory networks and attention mechanisms,which is applied to extract text features from image-text data;the visual feature extraction module is used to obtain visual features from image-text data via deep convolutional neural networks;the interactive fusion module is responsible for the correlation modeling between different modalities;the sentiment classification module is utilized to discriminate the sentiment tendency.In addition,the experimental analysis on real datasets(Getty Images dataset,Twitter dataset,and Flickr dataset)demonstrates that the proposed model respectively achieves 95.6%,97.6%,88.7% classification accuracy,which is also better than the existing models.