Sentiment Analysis Of Short Annotated Video Fusing Text Information

With the increasing role of social media in people's lives,more and more people tend to express their opinions through GIF videos on social platforms.However,due to the complexity of video information,the existing video emotion analysis methods can not effectively extract the Spatiotemporal features,and the analysis of video emotion based on only one mode can not fully express the emotion.In view of the above problems,this thesis has done the following work:(1)To solve the problem that GIF video semantic information is complex and can't be extracted spatiotemporal features effectively,the thesis proposes a video emotion analysis method(RFCM)based on 3D residual convolution neural network(Res Net3D)and an improved convolution long-term memory network(FConv LSTM).This method first uses Res Net3 D to obtain the short-term local spatiotemporal emotional features of the video,and then uses FConv LSTM to model it for a long time to obtain two-dimensional long-term spatiotemporal emotional features.Then,based on the two-dimensional spatiotemporal emotional features,the deep features are learned again using convolutional neural networks.Then get the emotion classification of GIF video through Softmax classifier.Finally,the experimental results on the T-GIF data set,GSO-2016 data set and the self-made D-GIFGIF data set show that this method can effectively improve the effect of video sentiment classification.(2)Aiming at the problem that the annotation of GIF video is short text,which contains no semantic information and lacks emotional features,this thesis proposes a text emotion classification method that combines word embedding vectors and word emotion vectors.This method first obtains word embedding vectors with rich semantics through the Word2 Vec model.Then select the word elements that have great influence on the text sentiment to construct the word sentiment feature vector.In order to enrich the emotional features of short text,the word embedding vector and the word emotion vector are fused.Finally,the convolutional neural network is used to model the fusion information,and the sentiment classification of the text is obtained.The experimental results show that this method can effectively improve the classification effect of the text sentiment.(3)Aiming at the problem of insufficient expression of emotion in a single modality,this paper proposes a method of fusing video and text information to perform sentiment analysis on annotated GIF short videos.The method first maps the sentiment probabilities obtained from the video and text to the sentiment intensity value of the sample.Then,in order to enrich the feature expression of emotions,a weighted fusion model combining the sentiment scores of video and text is designed,and the optimal fusion weight is obtained by the circular search method,and the short annotated video sentiment analysis results of the fused text information are obtained.Finally,a comparative experiment was conducted on T-GIF,GSO-2016 and the self-made D-GIFGIF sentiment data set.The experimental results show that,compared with the current emotion analysis methods for short videos,the short annotated video emotion analysis method proposed in this thesis,which integrates text information,has better classification effect.