Cyberbullying Detection With Multi-type Auxiliary Information

In recent years,online social platforms have gradually become an integral part of people's daily life.However,while they bring convenience to people's daily life,bullying speeches on social platforms also affect people's mental health seriously.Therefore,how to accurately and efficiently detect bullying speeches is of great significance.Cyberbullying detection methods can be mainly divided into two categories,textual information-based cyberbullying detection methods and auxiliary information-based cyberbullying detection methods.The former methods only use textual information to detect bullying speeches.However,the characteristics of "full of noisy data,lack of context" of online speeches make it difficult to extract textual features,which results in that this kind of methods cannot achieve good detection performances.To solve this problem,the latter methods use emotional information,user information,multimodal information and social platform meta-information as auxiliary information to supplement context and improve detection performances.After analyzing the fore-mentioned cyberbullying detection methods,it can be noted that there are still exists the following challenges.Challenge One:Textual feature extraction methods cannot represent well on "deliberately obfuscated words" in bullying speeches.Challenge Two: Most of the auxiliary information has limitations,which leads to the limitation of the application range of auxiliary information-based cyberbullying detection methods.Challenge Three: The improvement of detection performance is closely related to the feature extraction methods of auxiliary information,so it is necessary to represent auxiliary information effectively.This thesis analyzes and studies the cyberbullying detection methods base on the above three challenges,and the main contributions are as follows:First,this thesis gives an introduction to the background of cyberbullying detection and analyzes the related work and technologies in detail.Second,based on the analysis on a real-world dataset,two detection methods are proposed to solve the above three challenges:(1)Original Text Emotional Feature-Based Cyberbullying Detection Method: In addition to the textual features,this method uses the emotional features of the original text as auxiliary information to detect the comment texts.In particular,for challenge one,a Locality Sensitive Hashing-Based Text Representation Learning Method is proposed for textual feature extraction.This textual feature extraction method can represent the "deliberately obfuscated words" both accurately and effectively.(2)Multi-Type Auxiliary Information-Based Cyberbullying Detection Method:This method improves the Original Text Emotional Feature-Based Cyberbullying Detection Method.As for challenge two,it adds multi-type meta-information as auxiliary information to expand the application range of Original Text Emotional Feature-Based Cyberbullying Detection Method.In addition,an Attributed Heterogeneous Information Network-Based Feature Fusion Method is proposed for challenge three,which can efficiently fuse and represent multi-type meta-information.Last,experiments on real-world datasets demonstrate the effectiveness of the proposed cyberbullying detection methods.