Research On Irony Recognition Models Incorporating Multi-source Information

In recent years,social media sites such as microblog and post bar have gained great popularity and attention.The development of social media sites has stimulated the public to use metaphors and creative language in public,and irony as one of them is increasingly being used on social media.From a global perspective,irony is an interesting and compelling language form,but how to recognize its inner meaning is also a difficult task.Irony recognition can be studied from two ideas:context-independent and contextdependent.Context-independent methods start with the language characteristics of ironic sentences,while context-dependent methods pay more attention to the use of contextual information.With the continuous development of the field of deep learning,traditional machine learning irony recognition methods have been gradually eliminated,especially the advancement of pre-trained models has brought more room for exploration for irony recognition.Based on the pre-trained model,this article proposes two methods for two research ideas:(1)Context-independent irony research is limited by too little information,so it can only mine information from the text itself.Previous researchers have started with parts of speech tags,language characteristics,etc.Because irony sentences are sensitive to syntactic changes,and irony sentences imply strong sentiment of users,this paper introduces dependency parsing module and sentiment analysis module to analyze each sentence and extract sentimental features,and fuses them with the context vector from the pre-trained model for irony recognition.The irony recognition method proposed in this article on the Ciron dataset improves the F1 by 2.5%compared to BERT.(2)Because the context-dependent irony recognition method can introduce more context information,theoretically it has higher accuracy than the context free irony recognition method.Reasonable transformation of the context information into the model can significantly improve the accuracy of irony recognition.Because each social media user has different usage habits of irony,this paper uses the user’s history to transform the user into a vector as a user feature.In addition to user characteristics,context information is also an important information for irony recognition.This paper will also add context information.Finally,On the balanced and unbalanced data sets of SARC,the method proposed in this paper increases the F1 of the best model by 1.5%and 2.2%,respectively,and the ablation experiment proves that the introduced context information does improve the accuracy of irony recognition.