Medjezikovni prenos napovednih modelov za sovražni govor

PEČOVNIK, ŽAN (author), Robnik Šikonja, Marko (mentor), Ljubešić, Nikola (co-mentor)

Z razvojem družbenih omrežij je narasla pogostost sovražnega govora v upo- rabniških vsebinah. Osredotočili se bomo na dve trenutno najbolj aktualni temi, LGBT in migrante. Za napovedovanje sovražnega govora bomo upo- rabili nevronsko mrežo BERT in naredili primerjavo med večjezikovnim mo- delom, ki je naučen na 104 različnih jezikih ter trojezikovnim modelom, ki je naučen na slovenščini, hrvaščini in angleščini. Ugotovili smo, da trojezikovni model za približno 5% natančneje napoveduje sovražni govor na jeziku, na katerem je bil model tudi naučen. Večjezikovni model, brez ali z dodatnim učenjem, natančneje kot trojezikovni model napoveduje sovražni govor na jezikih, na katerem prvotno model ni bil naučen. To kaže na boljši medjezi- kovni prenos večjezikovnega napovednega modela.

• Language: Slovenian
• Keywords: sovražni govor, model BERT, nevronske mreže, medjezikovni prenos, strojno učenje, obdelava naravnega jezika.
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2020
• PID: 20.500.12556/RUL-116665
• COBISS.SI-ID: 18357251
• Publication date in RUL: 01.06.2020

Secondary language

• Language: English
• Title: Cross-lingual transfer of hate speech prediction models

Abstract

With the development of social networks, there has been a signicant in- crease of hate speech in user generated contents. We focus on two most discussed topics, LGBT and migrants. We use the BERT neural network for prediction of hate speech and make a comparison between the multilingual model, trained on 104 dierent languages, and a trilingual model, trained on Slovene, Croatian and English. Results show that the trilingual model is ap- proximately 5% more accurate predicting hate speech on a language that it was trained on. The multilingual model with or without additional training is more accurate on languages that it was not trained on. This indicates a better cross-lingual transfer of multilingual model.

• Keywords: hate speech, BERT model, neural networks, cross-lingual trans- fer, machine learning, natural language processing.