Profiliranje širitelja govora mržnje na društvenim mrežama metodama strojnog učenja

master's thesis

• Title (english): Profiling Hate Speech Spreaders on Social Networks Using Machine Learning Methods
• Author: David Dukić
• Mentor: Ana Sović Kržić (mentor)
• Committee member: Ana Sović Kržić (predsjednik povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-01, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Kako društvene mreže rastu, sve više i više njihovih korisnika širi govor mržnje kroz svoje objave. Uočeni fenomen često nastaje kao posljedica osjećaja lažne sigurnosti i anonimnosti u virtualnom okruženju. Da bi zaustavili širitelje govora mržnje, znanstvenici su počeli razvijati sustave strojnog učenja koji ih automatizirano detektiraju temeljem sadržaja njihovih objava. Ovaj diplomski rad opisuje nekoliko takvih novih sustava za profiliranje autora, učenih na korpusu engleskih Twitter objava, s krajnjim ciljem predviđanja širi li autor danih objava govor mržnje ili ne. Osim evaluacije isprobanih modela, detaljno je opisan i skup podataka te je nad njim provedena eksplorativna analiza. Značajke modela koji je najbolje radio nad neviđenim podacima kreirane su konkatenacijom dva tipa značajki. Prvi tip se sastojao od fino podešenih kontekstualiziranih gustih vektorskih reprezentacija izvučenih iz BERT modela koje su sumirane kroz njegovih zadnjih 12 skrivenih slojeva vezanih uz klasifikacijski token. Drugi tip su činile tri binarne varijable zvane indikatori. Binarne varijable su indicirale nalaze li se nizovi znakova rt, url ili hashtag u autorovim objavama. Klasifikator koji je postigao najveću točnost od 81% na skupu za testiranje koristeći opisane značajke bila je logistička regresija.

Abstract (english)

As social media grows, more and more users are disseminating hate speech through their posts. The noticed phenomenon often comes as a consequence of feeling a false security and anonymity in a virtual environment. To stop hate speech spreaders, researchers started developing machine learning systems that automatically detect spreaders of hate speech based on the contents of their posts. This graduate thesis describes several such novel author profiling systems which were trained on a corpus of English Twitter posts with a goal to predict if the author of the given posts spreads hate speech or not. Except for the evaluation of tried out models, a detailed description of the data set and its exploratory analysis was given. The features of the model that worked best on unseen data were crafted by concatenating two types of features. The first type were fine-tuned BERT contextualized embeddings summed over the last 12 hidden states corresponding to the classification token. The second type was made of the three binary variables called indicators. Binary variables were indicating whether strings rt, url or hashtag were present in the author's tweet posts, respectively. The classifier that achieved the highest accuracy score of 81% on the test set using described features was logistic regression.

• Keywords: BERT; društvene mreže; fino podešavanje; govor mržnje; guste vektorske reprezentacije; indikatori; logistička regresija; obrada prirodnog jezika; profiliranje autora
• Keywords (english): BERT; social networks; fine-tuning; hate speech; embeddings; indicators; logistic regression; natural language processing; author profiling
• Language: croatian
• URN:NBN: urn:nbn:hr:168:724095
• 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: 2023-01-13 12:09:53