Sarcasm Detection In Social Media

The rapid development of social media,such as Twitter,Reddit,and Weibo,has played an important role in people's daily lives.Internet users like and are good at using irony rhetoric in social media to vent their emotions.The rich use of this rhetorical method brings obvious difficulties to the natural language processing task,which will seriously affect the detection accuracy of tasks such as text sentiment analysis and viewpoint mining in social media.Therefore,the automatic sarcasm detection technology for social media has important explore and research significance.The specific forms of sarcasm rhetoric are diverse.According to the cause of sarcasm,sarcasm can be divided into three categories,namely,emotional contradictory sarcasm,scene sarcasm and other sarcasm.The first type of sarcasm accounts for about 70%.In view of the automatic sarcasm detection in social media,this thesis conducts in-depth research on the sarcasm detection based on content and context.The content-based sarcasm detection can also be called context-independent sarcasm detection,that is,regardless of the context information of the recognition target sentence,it is only determined from the target sentence itself to determine whether it is an sarcastic expression.According to the fact that the before and after emotional contradictory sarcasm is the majority,this paper proposes two contextindependent detection models,the "word pairs clash" model and the "semi-sentence pairs clash" model.The former uses the word pairs attention mechanism for the contradictory words in the sentence.The method calculates the attention score of any two words in the sentence to obtain the attention score matrix,and further obtains the sentence representation,which includes the degree of contradictory information of any two words.forceing the model to pay special attention to the contradictory words in the sentence under the supervision signal.The latter is aimed at the contradiction between the first half sentence and the second half of the sentence.Using the siamese neural network to model the "semi-sentence pairs clash",the model can find the inconsistency of the two sentences,so that it can be inferred whether the target sentence uses sarcasm rhetoric.It is also possible to add an attention mechanism based on the siamese neural network so that the model highlights the keywords in the two sentences.It can be seen,to some extent,the "semi-sentence pairs clash" model is an improvement of the "word pairs clash" model,that is,there is a single word level extended to multiple sentences(half sentences).The "semi-sentence pairs clash" model has achieved better results in the sarcasm detection task.The context-based sarcasm detection algorithm is based on an empirical theory that the sarcasm rhetoric is a contextual expression.The context includes the narrative of the person's tone,expression,body movements,the state of mind at the time,the context of the text in the long text,the forwarding,replying,and commenting of the text.In principle,the use of contextual information can greatly increase the upper limit of the accuracy of sarcasm detection.The contextual sarcasm detection algorithm proposed in this paper takes the SARC corpus as the research object,and adopts the text content driving and context-driven hybrid modeling method to detect the sarcasm.The context information used has user information and subject information.Experiments show that the decision algorithm after adding context can correctly judge some sarcasm comments,although they have no sarcasm tendency,so the recognition accuracy is greatly improved compared with the pure content-driven algorithm.