Abstract: Gender bias in word embeddings gradually becomes a vivid research field in recent years. Most studies in this field aim at measurement and debiasing methods with English as the target language. This paper investigates gender bias in static word embeddings from a unique perspective, Chinese adjectives. By training word representations with different models, the gender bias behind the vectors of adjectives is assessed. Through a comparison between the produced results and a human-scored data set, we demonstrate how gender bias encoded in word embeddings differentiates from people's attitudes.
Introduction
While algorithmic gender bias research has predominantly focused on English—often examining occupational links like "doctor" vs. "nurse"—gender stereotypes are also deeply woven into the adjectives used to describe individuals. In Chinese, a logographic language, this bias is further complicated by the semantic weight of individual characters. This paper investigates how these biases manifest in static word embeddings and how they compare to human social attitudes.
Methodology: Capturing Character-Level Nuance
The study compares two primary embedding techniques: the standard Continuous Bag-of-Words (CBOW) model, which treats words as atomic units, and Character-enhanced Word Embeddings (CWE). CWE models leverage the internal structure of Chinese words, integrating the embeddings of component characters (Hanzi) with the word-level vector.
To quantify bias, the researchers utilized a projection method, defining a "gender direction" vector by subtracting the vector for "she" (她) from "he" (他). Adjectives were then projected onto this vector to generate a bias score. These scores were validated against the AGSS (Adjectives List with Gendered Skewness and Sentiment) dataset, a human benchmark where 466 adjectives were rated on a gendered scale from 1 (female-related) to 5 (male-related).
Data and Training
The models were trained on a 1.54GB corpus of online Chinese news from Sogou Labs, totaling approximately 375 million tokens. Before bias testing, the embeddings were evaluated on 17,813 word analogy tasks to ensure high semantic quality and functional performance.
Key Findings: The "Male Neutrality" Gap
The research revealed a significant discrepancy in how AI models and humans perceive gendered adjectives. While the CWE model showed a higher overall correlation with human attitudes—confirming that semantic information hidden within characters (like the "woman" radical 女) carries significant weight—the alignment was not uniform across genders.
For adjectives humans associated with women (AGSS < 3), there was a strong correlation (~0.67) between model projections and human scores. However, for adjectives associated with men (AGSS > 3), the correlation vanished, dropping to a statistically insignificant 0.036.
Discussion: Language Evolution vs. Social Stagnation
The authors suggest this "Male Neutrality Gap" stems from a divergence between evolving media language and static social stereotypes. In the news corpus used for training, adjectives traditionally reserved for men are increasingly used in neutral, professional contexts. Conversely, human respondents may still rely on traditional stereotypes when scoring these same words.
Additionally, the study noted a phenomenon of "people hypocrisy": human scores were tightly clustered compared to the wider distribution of model scores. This suggests that individuals may subconsciously moderate their reported biases when asked directly, even if their actual language use—captured in the news data—remains skewed.
Conclusion
This study highlights that gender bias in Chinese word embeddings is not a simple mirror of human prejudice. While character-level information improves model alignment with human perception, the gap in male-associated adjectives suggests that language models may capture a more "neutralized" version of masculine traits found in modern reporting, contrasting with the more rigid stereotypes held by human observers.
Summary updated