Sarcasm Detection Based On Online Social Media

People often use social media to post their own opinions to express their moods,and many comments have sarcasm tendency.The emotions on the surface of the comments are often contrary to the real emotions.Therefore,correct identification of the sarcasm features of the comments has become an important task for sarcasm detection.Sarcasm detection aims to determine whether the comment contains a tendency to be sarcastic or not.Since it requires deep semantic information to determine the comment whether have sarcasm tendency or not,it is a challenging task.From the perspective of research,sarcasm detection can be divided into two perspectives:target comment and joint contextual information.From the perspective of research methods,the research has been biased towards deep learning methods in recent years,but research methods such as rule-based approaches and statistics-based approaches have also been used in the past.The methods that scholars often use in the above-mentioned research perspectives,such as Convolutional Neural Networks(CNN),Recurrent Neural Networks(RNN),etc.,they use regular data structures that are easy to process.However,in reality,the structure of data is often irregular,such as social networks,chemical molecules,etc.Therefore,in data extraction,the data with irregular structure solved as data with regular structure,will cause information loss,which is necessary for deep semantic information in sarcasm detection.In this thesis we have researched this problem and finds a new type of neural network-Graph Convolutional Neural Networks(GCN).Its data structure can be networked and irregular,which is convenient for extracting the global features of the comment.The research has the following two aspects:1.From the perspective of comment in sarcasm detection,in order to solve the data structure problems of CNN,RNN and other models,as well as ignore the comment features which include the global features of discontinuous and long-distance semantics,we have proposed two model based on Text GCN:TGbL model and TGsabL model,Long Short-term Memory Neural Networks(LSTM)and Self-Attention Mechanism have added to the above two models in order to extract the timing information of the comments and weight distribution information which highlight the contrast sentiment.The experimental results of the two models are better than the benchmark models.2.From the perspective of context in sarcasm detection,the context information we are concerned is the author feature of the comment,we propose models based on two methods of constructing author features:author feature based on GCN model and naive two-dimensional author vector method model,besides the vector representation of the comment is learned by the first stage model.The author feature constructed by GCN can learn the global characteristics of each author,and the author feature constructed by the naive two-dimensional vector method can express the author's sarcasm tendency.The experimental results of the above two models are better than the benchmark models,and play their respective advantages.