TL;DR
This paper investigates gender bias in Chinese word embeddings, which lack grammatical gender. By analyzing semantic similarities between words and predefined gendered terms, the researchers found that these embeddings reflect societal stereotypes, associating men with actions and negative traits, and women with appearance and positive/emotional traits. This work highlights the subtle ways language models can perpetuate harmful biases, even in languages without explicit grammatical gender markers.
What Problem Does This Paper Address?
Gender is a social construct heavily influenced by language. While many languages have grammatical gender (e.g., French, Spanish), Chinese does not. This absence of explicit grammatical markers makes it challenging to quantitatively study how gender is perceived and judged within the language itself. The authors aim to bridge this gap by using word embeddings, a powerful tool in Natural Language Processing (NLP), to uncover implicit gender biases in Chinese. They hypothesize that if word embeddings, trained on vast amounts of text, associate gender-neutral words with gendered concepts, it reflects and potentially reinforces societal stereotypes. Such biases can lead to unfair treatment and harm individuals who deviate from these mainstream perceptions, especially when these embeddings are used in downstream NLP applications.
Methods
The core of the methodology involves analyzing pre-trained Chinese word embeddings to detect implicit gender bias.
- Semantic Similarity: The researchers defined a set of "gendered words" (masculine and feminine terms in Chinese) and calculated the cosine similarity between each word embedding in their dataset and these gendered terms. This similarity score indicates how "masculine" or "feminine" a word is perceived to be by the embedding.
- Odds Ratio (OR): To quantify gender bias, they calculated an Odds Ratio (OR) for each word embedding. This metric compares a word's similarity to masculine terms versus its similarity to feminine terms. An OR close to 1 indicates masculinity, while an OR close to -1 indicates femininity.
- Linguistic Feature Analysis: The study then analyzed the relationship between these OR values and various linguistic features:
- Part-of-Speech: Categorizing words as nouns, verbs, and adjectives, with the hypothesis that verbs relate to actions and nouns/adjectives to appearance.
- Sentiment Polarity: Classifying words as positive or negative.
- Emotion Category: Using Ekman's basic emotions and Chinese cultural nuances.
- Semantic Category: Grouping words into topics like "activity," "psychology," and "state."
- Statistical Tests: Independent-samples t-tests, one-factor ANOVA, and Kruskal-Wallis tests were employed to determine the statistical significance of the observed relationships between gender bias (ORG values) and these linguistic categories.
Datasets
- Word Embeddings: The study utilized 12,251 Chinese word embeddings pre-trained on the Baidu Encyclopedia Corpus using the word2vec model (Skip-Gram with Negative Sampling). This corpus is large (4.1GB, 745M tokens) and covers a wide range of topics, aiming for objectivity.
- Gendered Words: A set of 14 masculine and 14 feminine Chinese words were selected from a related English study (Nadeem et al., 2020) to serve as anchors for measuring semantic similarity.
Key Insights and Ideas
- Implicit Bias in Gender-Neutral Languages: The study demonstrates that even in languages without grammatical gender, word embeddings can capture and reflect subtle societal gender biases. This is a crucial finding, as it suggests that bias is not solely a feature of explicit grammatical structures but is deeply embedded in language use.
- Action vs. Appearance Dichotomy: A significant finding is the strong association between verbs (representing actions) and masculinity, while nouns and adjectives (representing appearance) are more associated with femininity. This aligns with common societal stereotypes where men are judged by their deeds and women by their looks.
- Emotional Associations: The research reveals that negative emotions like disgust and anger are more strongly associated with men, while happiness and sadness are linked to women. This suggests a societal expectation for men to express negative emotions and women to be more emotionally expressive in certain ways, potentially reinforcing gender roles.
- Positive vs. Negative Sentiment: Positive sentiment words tend to be more feminine, while negative sentiment words are more masculine. This implies a societal tendency to view women more positively and men more negatively, or perhaps a tolerance for negative expressions from men.
- Content Association: "Activity" words are perceived as masculine, while "psychology" words are feminine. This reinforces the idea of men as rational actors focused on external tasks and women as more introspective or focused on emotional cognition.
Key Results
- The study successfully identified implicit gender bias in Chinese word embeddings, with masculine words relating to "war" and "power," and feminine words to "flower" and "beauty."
- Gender and Activity: Verbs (representing actions) were found to be significantly more masculine (M=0.022) than nouns (M=0.003) and adjectives (M=-0.064) (p<0.001).
- Gender and Sentiment Polarity: Positive words were more feminine (M=-0.017) than negative words (M=0.034) (p<0.001).
- Gender and Emotion Category: Disgust (M=0.030) and anger (M=0.025) were associated with masculinity, while happiness (M=-0.130) and sadness (M=-0.089) were associated with femininity.
- Gender and Content: Activity words (M=0.057) were masculine, while psychology words (M=-0.050) were feminine.
Limitations and Future Work
- Limited Linguistic Features: The study focused on four categories (part-of-speech, sentiment polarity, emotion, and semantic category). The authors acknowledge that other linguistic features could reveal further biases.
- Contextualization: The word embeddings used were not contextualized, meaning they represent a single meaning for each word. Future work could explore contextualized embeddings to see how bias manifests in different contexts.
- Binary Gender: The analysis was limited to a binary gender distinction (masculine/feminine). The authors suggest investigating more dimensions for investigation, such as in-group/inter-group, animate/inanimate, or collectivism/individualism.
- Harmful Stereotypes: The paper highlights that these biases can cause "representational harms" by erasing individual uniqueness and reinforcing stereotypes, particularly for those who do not conform to mainstream gender perceptions. Future work could focus on developing methods to mitigate these biases.
Summary updated