Otkrivanje govora mržnje zasnovano na modelima strojnog učenja

master's thesis

• Title (english): Hate speech detection based on machine learning models
• Author: Eva Šmuc
• Mentor: Goran Delač (mentor)
• Committee member: Goran Delač (predsjednik povjerenstva)
• Committee member: Marin Šilić (član povjerenstva)
• Committee member: Klemo Vladimir (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-09-14, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Electrical Engineering

Abstract

U ovom radu, odnos govora mržnje i sličnih zadataka obrade prirodnog jezika
istražen je kroz tehniku prijenosa učenja. Skup podataka ETHOS odabran je
kao ciljni skup podataka za klasifikaciju. Nekoliko je algoritama trenirano za
srodne zadatke (npr. toksični govor) na većim skupovima podataka, a dobiveni
modeli su iskorišteni za dobivanje predikcija za ciljni skup podataka. Predikcije
su korištene kao jedine ili dodatne značajke u eksperimentima koji slijede.
Eksperimenti se provode korištenjem više algoritama (Logistic Regression, SVM,
RidgeClassifier, Decision Tree, Random Forest, AdaBoost, GradBoost, Bagging)
i višestrukih prikaza teksta, uključujući Tf-Idf, Bert embeddinge i Bert embeddinge
u kombinaciji sa spomenutim dodatnim značajkama. Modeli temeljeni na
transformerima, BERT i DistilBERT koriste se za fino podešavanje na skupu podataka
ETHOS. Svi dobiveni modeli su evaluirani i njihove metrike su međusobno
uspoređene i uspoređene s rezultatima koje su dobili autori skupa podataka ETHOS.
Kako bi se istražili preostali problemi, koristi se model-agnostička metoda
LIME za dobivanje objašnjenja za pojedine pogrešne predikcije.

Abstract (english)

In this work, the relationship between hate speech and similar tasks is explored
through application of transfer learning techniques. A hate speech dataset called
ETHOS is chosen as the target dataset. Multiple algorithms are trained for similar
tasks (e.g. toxic language detection) on larger datasets, and the obtained
models are used for getting predictions for the target dataset. These predictions
are used as sole or additional features in the experiments that follow. Experiments
are carried out using multiple algorithms (Logistic Regression, SVM, RidgeClassifier,
Decision Tree, Random Forest, AdaBoost, GradBoost, Bagging) and
multiple text representations, including Tf-Idf, Bert embeddings and Bert embeddings
combined with mentioned additional features. Transformer-based models
BERT and DistilBERT are introduced and fine-tuned on ETHOS dataset. All
the obtained models are evaluated and their performance metrics is compared
mutually and compared to results obtained by the authors of the ETHOS dataset.
In order to explore the remaining underlying issues, model-agnostic method
LIME is used to get explanations for incorrect predictions.

• Keywords: prijenos učenja; detekcija govora mržnje; duboko učenje; ekstrakcija značajki
• Keywords (english): transfer learning; hate speech detection; deep learning; feature extraction
• Language: english
• URN:NBN: urn:nbn:hr:168:534922
• Study programme: Title: Information and Communication Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije (magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-15 11:20:20