Research On Sentiment Analysis Algorithm For Video Scene

With the development of sensor devices and networks,social media has developed rapidly and has become the main platform for users to express their opinions and emotions.Sentiment analysis can analyze social media data from an emotional perspective,understand users' views on social current events and products,and analyze the trend of social public opinions,which has become a current hot research topic.As a new type of social media data,video data contain not only the emotional content related to the object,but also the scene environment information related to the conveying of emotions which has become one of the mainstream ways for the public to express emotions.Therefore,this thesis focuses on the problem of video sentiment analysis,tracking the related research of existing video sentiment analysis,and researches on issues such as insufficient extraction of visual information from video frame sequences and emotional confusion caused by the diversified expression of video emotions.The specific work is as follows:(1)Due to insufficient information of the collected video data,such as lack of subtitles,audio,etc.To solve the problem of insufficient visual information extraction of video frame sequences,a single-modal sentiment analysis algorithm based on three-dimensional residual attention is proposed.First,the video is converted into a sequence of continuous video frames for emotional feature extraction.Secondly,a three-dimensional residual convolutional neural network is used to introduce a time dimension to ensure the temporal coherence between emotional features.Finally,on the basis of the residual network,an attention mechanism is added to highlight emotion-related features.The comparative simulation experiments are carried out on the benchmark dataset and self-built dataset.The experiments show that the algorithm in this paper can effectively improve the accuracy of sentiment analysis.(2)General video data contain visual information and audio information,both of which carry emotional expression.Due to the diversification of emotional expression,the problem of emotional confusion in video data is caused,which leads to a decrease in the accuracy of video emotional analysis.In response to the above problems,the extraction and fusion of the two types of information when both visual information and audio information exist are studied,and a multi-modal sentiment analysis algorithm based on decision fusion is proposed.For audio features,a feature extraction algorithm based on audio atlas is proposed;for visual features,the sentiment analysis algorithm with three-dimensional residual attention proposed above is used to extract visual features,and then a cross-voting decision fusion mechanism is proposed to fuse audio and visual features to achieve the complementation of audio and visual features to improve the accuracy of emotion classification.The comparative simulation experiments are conducted on the benchmark dataset and the self-built dataset.The experiments show that the algorithm in this paper is better than other algorithms in the accuracy of sentiment analysis.