Prepoznavanje nasilja na društvenim mrežama

master's thesis

• Title (english): Detection of Cyberbullying on Social Networks
• Author: Ana Vukasović
• Mentor: Ivica Botički (mentor)
• Committee member: Ivica Botički (član povjerenstva)
• Committee member: Tomislav Jagušt (član povjerenstva)
• Committee member: Danijel Mlinarić (predsjednik povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-14, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Kontinuiranim rastom društvenih mreža, ljudi imaju sve veću virtualnu povezanost, ali i nažalost sve veću platformu za online nasilje. Kako bi se takvo ponašanje spriječilo, počinju se koristiti sustavi za automatsko detektiranje nasilnog govora na društvenim mrežama. Ovaj rad opisuje nekoliko takvih sustava temeljenih na strojnom učenju za objave na engleskom jeziku na Twitter društvenoj mreži. Modeli određuju radi li se o nasilju te o kojem obliku nasilja. Osim evaluacije isprobanih modela,... More detaljno su opisani i korišteni skupovi podataka. Model s najboljim performansama temelji se na ugađanju predtreniranog BERT modela, pomoću kojeg se iz teksta dobije kontekstualizirana gusta vektorska reprezentacija, te dodatnih značajki napravljenih predobradom teksta koje opisuju prisutnosti određenih znakova u tekstu (interpunkcijski znakovi, hashtagovi, itd.) koji zajedno ulaze u višeslojni perceptron. Opisani model postigao je 89% točnosti na testnom skupu. Less

Abstract (english)

With the continuous growth of social networks, people are more and more virtually connected, but this also, unfortunately, increases the platform for online violence - cyberbullying. To prevent such behavior, systems for automatic detection of cyberbullying on social networks are starting to be used. This graduate thesis describes several such new machine learning-based systems for posts in English on the Twitter social network. The models determine whether it is cyberbullying and what form... More of cyberbullying. In addition to the evaluation of the tested models, used datasets are described in detail. The best-performing model is based on tuning a pre-trained BERT model, which yields a contextualized dense vector representation from the text, and additional features made by text preprocessing that describe the presence of certain characters in the text (punctuation, hashtags, etc.) are put at the input of a multilayered perceptron. The described model achieved 89% accuracy on the test set. Less

• Keywords: nasilje; društvene mreže; obrada prirodnog jezika; NLP; strojno učenje; BERT; višeslojni perceptron; guste vektorske reprezentacije; naivni Bayesov klasifikator
• Keywords (english): cyberbullying; social networks; natural language processing; NLP; ML; machine learning; BERT; multilayer perceptron; embeddings; naive bayes classificator
• Language: croatian
• URN:NBN: urn:nbn:hr:168:674063
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 10:16:21