Using Distributional Models for Studying the Influence of School Textbooks in Children Bias

Abstract

School textbooks have a profound influence on shaping the thoughts of young individuals. Reducing sex, gender, and race bias in these educational materials is then an imperative social goal. Despite efforts to address bias, historical and contemporary textbooks have been found to perpetuate stereotyped associations: efficient techniques like word embeddings can be used to analyze a wide range of school textbooks to assess the presence of stereotyped biases. In this paper, we use Artificial Intelligence to test what happens if people are exposed to some textbooks or others. Our procedure can be used for in-silico testing of the persuasive power of textbooks. For this purpose, we introduce ChildDM, a model constructed from a corpus of children’s free-produced language, and its domain-adapted version, ChildDM-School, trained on a collection of school textbooks. Leveraging the Word Embedding Association Test (WEAT) framework, we investigate how biases evolve in the language of children after exposition to historical textbooks. While historical textbooks tend to avoid explicit gender-based stereotypes in the scientific and artistic domains, they still conceal biases. Specifically, women are associated with caregiving and family-related activities.

Keywords

• Distributional models
• word embedding
• gender bias
• school textbooks
• education