Detecting gender bias in Arabic text through word embeddings

Area of Science:

• Computational Linguistics
• Sociolinguistics
• Gender Studies

Background:

• Historical focus on Western women's rights and economic status leaves other regions, like the Middle East, understudied.
• A shortage of gender-based economic statistics in the Arab world necessitates alternative methods for assessing gender disparities.
• Sociocognitive theory posits that language reflects social norms, making textual analysis crucial for understanding gender biases.

Purpose of the Study:

• To examine gender-based biases in occupational representation within Arabic textual corpora.
• To adapt and apply Word Embedding Association Test (WEAT) and Direct Bias quantification tests for the Arabic language.
• To provide evidence of persistent gender stereotypes in the Arab world, particularly in the absence of comprehensive census data.

Main Methods:

• Adaptation of WEAT and Direct Bias quantification tests for Arabic language.
• Analysis of diverse Arabic text datasets, including Lebanese news archives, Arabic Wikipedia, and electronic newspapers from UAE, Egypt, and Morocco.
• Utilizing word embedding models to detect and quantify gender-associated biases in occupational terms.

Main Results:

• WEAT tests consistently linked career, science, and intellectual pursuit terms with men across all datasets.
• Family and art-related terms were predominantly associated with women in the analyzed texts.
• Direct Bias analysis revealed a consistent bias associating females with roles like nurse, house cleaner, maid, and secretary, with some datasets also linking them to researcher, doctor, and professor.

Conclusions:

• Arabic text datasets reflect significant gender stereotypes concerning occupations.
• These findings provide empirical evidence of enduring gender biases in the Arab world, despite limited statistical data.
• The study underscores the importance of critical discourse analysis in understanding and addressing gender inequality reflected in language.