20.500.12556/RUL-167343 Odkrivanje pristranosti in razpristranjevanje v jezikovnem modelu SloBERTa Detecting and removing biases in the SloBERT model Veliki jezikovni modeli so omogočili velik napredek pri obdelavi naravnega jezika, vendar so lahko pristrani, predvsem zaradi učnih podatkov. Pristranost lahko vodi do diskriminatornih odločitev v zaposlovanju, v pravosodju in v drugih aplikacijah. V magistrskem delu raziskujemo spolno pristranost slovenskega maskirnega jezikovnega modela SloBERTa. Predstavimo in uporabimo šest metod za analizo pristranosti ter prilagodimo angleško metodo razpristranjevanja Kaneko in Bollegala (2021) za slovenski model SloBERTa. Zmogljivost modela pred in po razpristranjevanju ovrednotimo z nalogami iz zbirke SuperGLUE ter z nalogo klasifikacije sentimenta z množico Sentinews. Rezultati testa stavčne asociacije potrjujejo pristranost v prid moškim v odnosu do poklicev v modelu SloBERTa, medtem ko ostalih pet metod pristranosti test ne zazna. Po razpristranjevanju pristranosti ne zazna več nobena metoda, vendar razpristranjeni modeli dosegajo slabše rezultate pri nalogah iz zbirke SuperGLUE in pri nalogi klasifikacije sentimenta. Large language models have significantly advanced the area of natural language processing, but they inherit biases present in the training data. The learned biases can lead to discriminatory decisions in employment, justice, and other applications. This work explores gender bias in the Slovene masked large language model SloBERTa. We present and apply six methods for bias analysis and adapt the English debiasing method Kaneko and Bollegala (2021) for the Slovenian SloBERTa model. The model's performance before and after debiasing is evaluated using tasks from the SuperGLUE benchmark and sentiment classification task with the Sentinews dataset. The Sentence Association Test results confirm a male bias in relation to professions in the original SloBERTa model, while the other five bias detection methods do not identify such bias. After debiasing, none of the methods detect bias. However, the debiased models show a decline in performance on selected tasks from the SuperGLUE benchmark and the sentiment classification task. veliki jezikovni modeli pristranost razpristranjevanje SloBERTa spolna pristranost Large language models bias debiasing SloBERTa gender bias true false false M. Narat Slovenski jezik Angleški jezik Magistrsko delo/naloga 2025-02-16 08:30:15 2025-02-16 08:30:20 2025-02-17 09:39:41 0000-00-00 00:00:00 2025 0 Ljubljana 0 Ljubljana 68 str. 0000-00-00 Zaloznikova Objavljeno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 81(043.2) 183024 226284547 38884.pdf 38884.pdf 1 0C7C23B9FEDD262E5732BA9676D4469D 07cd72018aaeb653454e35dbf34c03b816c42e2af97a9cc3e81e9ff492d67358 fc862cbc-ec36-11ef-b232-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=199700 Pedagoška fakulteta 0 0 0