Recognition And Classification Of Offensive Language In Social Media

With the development of Internet-related technologies,more and more social media platforms have emerged in people's eyes.While enjoying the many conveniences brought by social media platforms,people are also deeply disturbed by the bad information in them,including Offensive Language.Offensive language in social media refers to offensive speech that is not accepted by the world,such as cyberbullying,sexism,racial discrimination,etc.Many social media platforms will use the form of blocked keywords to restrict the offensive language,but simple rule matching is difficult to effectively and accurately judge the existence of the offensive text,and more and more scholars have begun to try to use natural language processing related knowledge to solve this task.The content of this article is to identify and classify the offensive language in social media.We mainly conducted the following research:(1)Recognition of offensive language in social media.This task is a binary classification task that refers to a given text to determine whether there is offensive behavior in the text.For this task,we propose a BCNN model,which uses BERT to extract context-sensitive text representations,and obtains n-grams features in the text through a convolutional neural network model.Using this model,we have effectively recognized offensive behavior in social media.(2)Identification and classification of attack objects in social media.This task refers to the given offensive text,identify whether there is an attack object,and classify the attack object.Based on the BCNN model,we add attention mechanisms in different ways to capture the attack object information in the text through attention.At the same time,we propose two network structures to better combine the offensive information extracted through the CNN network and the attack object information obtained through the attention mechanism.(3)Create and incorporate external knowledge.First,we use the contextsensitive model to carry out the word offensive evaluation task,that is,for a certain word in a given text,evaluate the aggressive information of the word in the current context.Then we use the model to evaluate each word in the data set of the first two tasks,and then integrate the evaluation information of each word into the model of the first two tasks.It is proved through experiments that the fusion of external knowledge can effectively improve the efficiency of the model when a simple model or a data set with few samples is used.