Bimodal Sentiment Analysis Combining Text And Short Video

With the accelerated commercial grounding of 5G,the rapid development of microblogging,Tik Tok and other applications,the network has generated a large number of user-generated Short videos with textual content.The analysis of these textual content and short videos allows us to understand the perceptions of social events,people,products and the evolution of public opinion by a wide range of users.thus,multimodal sentiment analysis has become a very popular research topic at present.However,the data obtained directly from the web are too redundant and complicated to be directly applied in practical research,and among practical research available The dataset is still lacking,and the text and video content are independent and cannot be analyzed in a study of sentiment analysis.To address these issues,this thesis proposes the following works:(1)To address the problem of lack of data samples for short video sentiment analysis,we propose a short video sentiment based on few shot learning analysis methods.The dataset is first divided into a support set and a query set;then,the visual features are extracted from the dataset respectively;and the similarity between the query set and support set is calculated by the metric module after the features extracted from the support set and those extracted from the query set are concatenated together.finally use the classifier to predict the categories of the query set sample.(2)To address the problem that there are multiple modalities in the dataset,a sentiment analysis method for text and short video is proposed.In the textual modalities,we first use the partitioning tool to pre-process the textual modal data,and then use the word embedding tool to obtain the textual correspondence.of the word vector,the resulting word vector will be used to extract the affective features of the textual information using the LSTM network of the attention mechanism,and finally the emotional features of the textual information will be extracted by the The classifier computes the sentiment probability of text messages on three categories: positive,neutral,and negative;on the short video modality,3D residuals are used The video feature extraction method for dense networks is used to establish a video sentiment classification model and a classifier is used to determine the sentiment probability of short video messages.(3)For the problem that the data of text and short video bimodal cannot be effectively merged,we propose a decision-level weighted fusion-based sentiment classification method.After obtaining the recognition rate for multimodal emotion classification,the method designs different weights for each modality,then introduces a weighting matrix,and finally the fusion of sentimentclassification results of two modalities,text and video,to establish a bimodal decision-level weighted fusion-based sentiment recognition model.