Research On Multi-Modal Cyberbullying Detection Based On Deep Learning

Cyberbullying refers to malicious,offensive,or insulting behavior against individuals or groups on the Internet.This behavior includes,but is not limited to,the use of words or images to harass,threaten,insult,and spread rumors,causing mental harm to the victim,and even causing property damage and personal image damage.With the widespread use of social media on the Internet,cyberbullying has become a global social problem.In recent years,there has been a gradual increase in the use of voice for cyberbullying,which causes greater harm to the bullied than using text or images.Therefore,developing a multimodal fusion network model that integrates text and speech features to detect cyberbullying content has extremely important application value.This model can effectively analyze and identify the characteristics of voice and text in cyberbullying,and better protect the rights and interests of victims.Considering that cyberbullying security comes from multimodal forms of information sources,this article aims to conduct research from two aspects: natural language processing algorithms and multimodal fusion networks,in order to explore effective methods for cyberbullying detection.Specifically,the research work of this article is as follows:(1)A text detection model based on hierarchical attention network(BHF)is proposed.Using the pre training model BERT to output the basic features of the text,key information is extracted from both word level and sentence level dimensions through a hierarchical attention mechanism,further improving the feature extraction ability of the model and obtaining deeper semantic features;At the same time,a fusion mechanism is introduced to adjust the feature distribution of the model output,fuse basic features and deep features,and obtain a semantic representation that integrates three dimensions of words,sentences,and full text,enhancing the learning ability of the model.(2)A multimodal fusion cyberbullying detection model based on spatial representation(MFNSF)is proposed.The model constructs shared space and specific space,respectively mining shared and specific features between different modes,and constrains the mapping direction of the space through shared loss and specific loss.In addition,the improved attention network is used to obtain modal fusion features and input them as final features into the cyberbullying model for detection and classification.Comparing the MFNSF model with mainstream multimodal fusion models,the experimental results show that the MFNSF model outperforms other methods in three different multimodal data sets,CMCAD,COLD,and CMU-MOSI.