Winogender 데이터 세트를 활용한 한국어 언어모델의 성 편향성 분석

초록

This paper investigates the impact of gender bias inoccupation nouns and adjectives in Korean language models on the genderreferred to by pronouns, utilizing the Winogender dataset. The experiment wasconducted in three ways: Measuring surprisal scores in Korean and English usingencoder models, and conducting experiments with a decoder model, namely isChatGPT-4, to test its responses. The encoder models showed that, regardlessof the gender bias differences in occupation nouns and adjectives, the malepronoun was more naturally used in sentences than the female pronoun. Onthe other hand, the decoder model detected gender bias especially in sentencescontaining adjectives. This result identifies the influence of gender imbalancein training data and the functional differences between the language generationmodel and the language comprehension model. This study suggests to constructa unique dataset that reflects the characteristics of Korean, in order to moreeffectively analyze gender bias in Korean language models.

키워드