What Might Books Be Teaching Young Children About Gender?

Extracting intersectional stereotypes from embeddings: Developing and validating the Flexible Intersectional Stereotype Extraction procedure

Social group-based identities intersect. The meaning of "woman" is modulated by adding social class as in "rich woman" or "poor woman." How does such intersectionality operate at-scale in everyday language? Which intersections dominate (are most frequent)? What qualities (positivity, competence, warmth) are ascribed to each intersection? In this study, we make it possible to address such questions by developing a stepwise procedure, Flexible Intersectional Stereotype Extraction (FISE), applied to word embeddings (GloVe; BERT) trained on billions of words of English Internet text, revealing insights into intersectional stereotypes. First, applying FISE to occupation stereotypes across intersections of gender, race, and class showed alignment with ground-truth data on occupation demographics, providing initial validation. Second, applying FISE to trait adjectives showed strong androcentrism (Men) and ethnocentrism (White) in dominating everyday English language (e.g. White + Men are associated with 59% of traits; Black + Women with 5%). Associated traits also revealed intersectional differences: advantaged intersectional groups, especially intersections involving Rich, had more common, positive, warm, competent, and dominant trait associates. Together, the empirical insights from FISE illustrate its utility for transparently and efficiently quantifying intersectional stereotypes in existing large text corpora, with potential to expand intersectionality research across unprecedented time and place. This project further sets up the infrastructure necessary to pursue new research on the emergent properties of intersectional identities.

The Emotional Content of Children's Writing: A Data-Driven Approach

Emotion is closely associated with language, but we know very little about how children express emotion in their own writing. We used a large-scale, cross-sectional, and data-driven approach to investigate emotional expression via writing in children of different ages, and whether it varies for boys and girls. We first used a lexicon-based bag-of-words approach to identify emotional content in a large corpus of stories (N>100,000) written by 7- to 13-year-old children. Generalized Additive Models were then used to model changes in sentiment across age and gender. Two other machine learning approaches (BERT and TextBlob) validated and extended these analyses, converging on the finding that positive sentiments in children's writing decrease with age. These findings echo reports from previous studies showing a decrease in mood and an increased use of negative emotion words with age. We also found that stories by girls contained more positive sentiments than stories by boys. Our study shows the utility of large-scale data-driven approaches to reveal the content and nature of children's writing. Future experimental work should build on these observations to understand the likely complex relationships between written language and emotion, and how these change over development.

Semantic Prosody: How Neutral Words With Collocational Positivity/Negativity Color Evaluative Judgments

We like people and objects more when they are described in positive than in negative terms. But even seemingly neutral words can elicit positive or negative responses. This is the case for words that predominantly occur alongside positive (or negative) words in natural language. Despite lacking positivity/negativity when evaluated in isolation, such semantically prosodic words activate the evaluative associations of their usual company, which can color judgment in unrelated domains. For example, people are more likely to infer that “endocrination” (a fictional medical outcome) is negative when it is “caused” (a word with negative semantic prosody) rather than “produced” (a synonymous word without semantic prosody). We review what is known about the influence of semantically prosodic words and highlight their importance for judgment and decision making.

Implicit Bias Reflects the Company That Words Keep

In everyday language, concepts appear alongside (i.e., collocate with) related concepts. Societal biases often emerge in these collocations; e.g., female (vs. male) names collocate with art- (vs. science-) related concepts, and African American (vs. White American) names collocate with negative (vs. positive) concepts. It is unknown whether such collocations merely reflect societal biases or contribute to them. Concepts that are themselves neutral in valence but nevertheless collocate with valenced concepts provide a unique opportunity to address this question. For example, when asked, most people evaluate the concept "cause" as neutral, but "cause" is frequently followed by negative concepts (e.g., death, pain, and trouble). We use such semantically prosodic concepts to test the influence of collocation on the emergence of implicit bias: do neutral concepts that frequently collocate with valenced concepts have corresponding implicit bias? In evaluative priming tasks, participants evaluated positive/negative nouns (Study 1) or pictures (Study 2) after seeing verb primes that were (a) strongly valenced (e.g., hate and comfort), (b) neutral in valence but collocated with valenced concepts in corpora (e.g., ease and gain), or (c) neutral in valence and not collocated with valenced concepts in corpora (e.g., reply and describe). Throughout, neutral primes with positive (negative) collocates facilitated the evaluation of positive (negative) targets much like strongly valenced primes, whereas neutral primes without valenced collocates did not. That neutral concepts with valenced collocates parallel the influence of valenced concepts suggests that their collocations in natural language may be sufficient for fostering implicit bias. Societal implications of the causal embedding hypothesis are discussed.