Affective Computing For Multi-characteristic Social Media Texts

With the rapid development of Internet technology,social media has penetrated people’s life more and more as a daily tool.It has become an important platform for people to participate in political,economic,and cultural life.A large number of users communicate and interact through social media platforms,gradually accumulating a large amount of users’ social media data.These social media data,containing the user’s stance and attitude,have distinct emotional characteristics and provide rich data resources for the emotional analysis of network social media data.The comprehensive use of natural language processing,deep learning,and other technologies to identify,understand,and apply the rich emotional information contained in social media data by intelligent means is increasingly recognized by industry and academia.Furthermore,these attempts can actively promote the research of text affective computing and provide a good foundation for the realization of emotional intelligence on social media platforms.It is worth noting that due to the numerous platforms and diverse scenes of social media,the different text have different data characteristics,such as the uncertain semantics and emotion of social media short text,the complicated semantics and emotion of social media long text,and the dynamic interactivity of emotion in dialogue text.These diverse characteristics of social media text bring great challenges to its text emotion classification,text emotion understanding,and dialogue text emotion recognition.Therefore,this paper will take multi-feature social media texts as the main research object to carry out the emotion calculation research of social media texts,focusing on three aspects of social media emotion classification of short text,emotional cause detection of long text,emotional recognition of interactive text.The main work and contributions of this paper are summarized as follows:Firstly,aiming at the uncertainty of semantics and emotion caused by the lack of information contents in social media short texts,this paper proposes a method named visual context enhancement and multi-grained semantics expression for social text emotion recognition.Traditional studies about social media short text emotion recognition mainly focus on text semantic modeling to realize text emotion classification.In fact,in addition to text information,social media data also contains multi-modal information such as images,audio,video,and so on.These rich multi-modal data have strong contextual correlations with social media texts,which can provide a good contextual complement in the semantic representation of social media short texts.Inspired by this,our emotion recognition method of social media short text takes good use of the text content and associated visual image information to enhance the semantic representation ability of social media short text.By supplement and integration of visual context information and multi-granularity semantic information,the semantic representation efficiency of short texts can be improved,and the emotion recognition accuracy of short texts can be promoted.A large number of experimental results on the real multimodal dataset show that this method has a good effect on short text emotion classification.On this basis,taking video-sharing websites as the specific application scenario,this paper conducts emotion recognition on social media short texts in the specific field of video user-generated comments and then applies it to video emotion analysis.Experiments on the real dataset show that this method can identify the emotion information of video user comments effectively and realize the task of video emotion analysis at a low cost.Secondly,given the difficulty in understanding semantics and emotion caused by the long length and complex information content in social media long text,this paper considers the fact that people tend to rely on narrative information to assist the reading of the long and complex text,and proposes a method named causal Narrative comprehension model for emotion cause extraction.Through the collaborative use of the module of causal correlation with narrative perception and the module of emotional attention with result perception,the method models the causal information contained in the emotional causal text and integrates the attention mechanism to better understand the complex causal semantics and emotional causal correlation of the causal text.In this way,the method improves the performance of the emotion cause extraction(ECE)task.Finally,experimental results on common Chinese and English reference data sets for ECE tasks validate the effectiveness of text emotion cause identification methods based on causal narrative understanding and demonstrate the potential of narrative information in long text understanding.Thirdly,based on the above studie about sentence emotion identification and long text emotion understanding,this paper also conducts the study about dialogue emotion.In view of the dynamic evolution of semantics and emotions caused by the interactive characteristics of social media dialogue texts,this paper is inspired by affect attribution theory and considers that the generation and dynamic evolution of dialogue emotions are derived from various emotional stimuli in the process of dialogue interaction and individuals’ cognitive responses to them.Furthermore,we proposed a novel approach named conversational emotion recognition based on global context enhanced and local multi-attribution fusion.Firstly,this paper models global context to enhance the global context of dialogue sentences to ensure that the semantics of each sentence under the semantic evolution of dialogue do not deviate from the global context information.Then,from the perspective of attribution understanding,this paper deeply analyzes the factors that influence the dynamic evolution of emotion in dialogue sentences,namely,adjacency information attribution,speaker information attribution,and local context attribution.As we know,the adjacent context of dialogue has more deterministic semantic echo than the distant context,in order to improve the efficiency of attribution understanding,this paper conducts sequential semantic modeling and fusion for each attribution element of the sentence based on the narrative information of the adjacent context window of each sentence.The output is the final emotion result state of multiattribution of each dialogue sentence after narrative understanding.Base on the final emotion result state,the emotion of each dialogue sentence can be predicted by the classifier.Experimental results on several public datasets show that the proposed method can effectively model dialogue emotion and improve the emotion recognition performance of dialogue text.