Detecting hate speech using deep learning on smartphones

Chestojanov, Atanas (author), Pejović, Veljko (mentor)

Hate speech detection is usually performed on remote servers. That requires the text that is being checked to leave the user’s device. This thesis presents a hate speech detection model that runs on an Android phone, and gives each user the opportunity to personalize and adapt it to a specific sub group of offensive language that they are the most sensitive to. All of that, without any data leaving the device. For that purpose, a MobileBERT encoder is first trained on the Dynahate dataset to give us a base model. The accuracy of the base model reaches 0.805 on the test split, with a ROC-AUC of 0.84, with the majority class baseline sitting at 0.550. After dynamic range quantization, the classifier’s size sits at 26 MB and classifies a sentence in 64 ms on the test device. The model is split into a frozen encoder and a classification head with 32,962 parameters. The on-device personalization trains only the head from the user’s own feedback. Training runs were done on about 220 sentences of feedback (sentence + label), for ten epochs at batch size 16. That takes under a second of processor time in total, and it does not increase memory use. The one step that does cost something is computing the sentence embeddings. Embedding a set of about 220 sentences takes roughly 5.3 s of processor time, and that is repeated every time a feedback set is loaded. The measurements show that feedback containing no examples of the original task erases that task. For that reason every training set also contains 100 sentences from the original dataset, which prevents the loss entirely.

• Language: English
• Keywords: hate speech detection, deep learning, on-device training, personalization, MobileBERT
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-187805
• Publication date in RUL: 14.09.2026

Secondary language

• Language: Slovenian
• Title: Zaznavanje sovražnega govora z uporabo globokega učenja na pametnih telefonih

Abstract

Zaznavanje sovražnega govora običajno poteka na oddaljenih strežnikih. To zahteva, da preverjano besedilo zapusti uporabnikovo napravo. V diplomskem delu predstavljamo model za zaznavanje sovražnega govora, ki teče na telefonu z operacijskim sistemom Android in vsakemu uporabniku omogoča, da ga personalizira in prilagodi tisti podskupini žaljivega jezika, na katero je najbolj občutljiv. Vse to, ne da bi kakršni koli podatki zapustili napravo. V ta namen najprej naučimo kodirnik MobileBERT na podatkovni zbirki Dynahate in tako dobimo osnovni model. Klasifikacijska točnost osnovnega modela na testni množici doseže 0,805, ROC-AUC pa 0,84, pri čemer je izhodiščna vrednost večinskega razreda 0,550. Po kvantizaciji z dinamičnim območjem klasifikator zasede 26 MB in poved razvrsti v 64 ms na preizkusni napravi. Model je razdeljen na zamrznjen kodirnik in klasifikacijsko glavo z 32.962 parametri. Personalizacija na napravi uči samo glavo, in sicer iz uporabnikovih lastnih povratnih informacij. Učenje je potekalo na približno 220 povedih povratnih informacij (poved in oznaka), deset epoh pri velikosti paketa 16. To skupaj porabi manj kot sekundo procesorskega časa in ne poveča porabe pomnilnika. Edini korak, ki nekaj stane, je izračun vektorskih vložitev povedi. Izračun vložitev za približno 220 povedi porabi približno 5,3 sekunde procesorskega časa in se ponovi vsakič, ko naložimo množico povratnih informacij. Meritve pokažejo, da povratne informacije brez primerov izvorne naloge to nalogo izbrišejo. Zato vsaka učna množica vsebuje tudi 100 povedi iz izvorne podatkovne zbirke, kar izgubo popolnoma prepreči.

• Keywords: sovražni govor, globoko učenje, učenje na napravi, personalizacija, MobileBERT