TL;DR
This paper uses word embeddings to analyze gender bias in Dutch newspapers from 1950-1990, revealing shifts in how men and women were represented over time and across different ideological leanings. By comparing word associations, the study offers a computational method to trace historical language change and uncover subtle societal biases.
What Problem Does This Paper Address?
In an era of increasing concerns about filter bubbles and media polarization, this research asks a fundamental question: did historical news media also exhibit biases, and how did these biases evolve? Specifically, the paper investigates gender bias in Dutch newspapers between 1950 and 1990, a period marked by significant societal changes like industrialization, modernization, and the decline of "pillarization" (societal stratification along religious and ideological lines). The authors aim to understand if and how newspapers, as key conduits of public discourse, reflected and potentially reinforced societal views on gender, and whether these biases differed based on the newspaper's ideological stance.
Methods
The core methodology of this paper relies on word embeddings, a technique that represents words as numerical vectors in a high-dimensional space, where words with similar meanings are located closer to each other. The authors trained word embedding models (specifically, using the Gensim implementation of Word2Vec) for each of six Dutch national newspapers, with each model representing a decade from 1950 to 1990.
To measure gender bias, they constructed two vectors representing the "male" and "female" dimensions of gender by averaging vectors of words like "man," "his," and "father" for the male dimension, and "woman," "her," and "mother" for the female dimension, along with popular Dutch first names. The gender bias for any given word was then calculated as the difference in its distance to the male and female gender vectors. A positive difference indicated a bias towards women, and a negative difference indicated a bias towards men.
They applied this bias measurement to three sets of target words:
- Job titles: Extracted from the HISCO dataset.
- Emotion words: Selected from the Cornetto lexicon based on polarity and confidence scores.
- LIWC categories: Utilizing Dutch translations of psychological and cognitive state categories from the LIWC2001 dictionary.
Finally, Bayesian linear regression was used to analyze how these bias values changed over time for each newspaper and across different categories.
Datasets
The study utilized a dataset of articles from six Dutch national newspapers:
- NRC Handelsblad (NRC) (liberal)
- Het Vrije Volk (VV) (social-democratic)
- Parool (liberal)
- Telegraaf (neutral/conservative)
- Trouw (Protestant)
- Volkskrant (VK) (Catholic)
These newspapers were chosen for their diverse ideological backgrounds. The paper also incorporated external lexicons:
- HISCO (Historical International Classification of Occupations) for job titles.
- Cornetto lexicon for emotion words.
- LIWC2001 (Linguistic Inquiry and Word Count) for psychological and cognitive states.
- Nederlandse Voornamenbank (Dutch first names database) for gendered name vectors.
Key Insights and Ideas
- Word embeddings as a historical lens: The paper demonstrates the power of word embeddings not just for analyzing contemporary language but also for tracing historical shifts in societal attitudes and biases embedded within language.
- Divergent gender bias trajectories: While some societal trends might suggest increasing gender equality, the analysis reveals that gender bias in newspapers did not uniformly move in one direction. Different newspapers and different thematic categories showed distinct patterns of change.
- Ideological divides matter: The study highlights that the ideological orientation of a newspaper significantly influenced its gender bias. Social-democratic and religious newspapers tended to show a clearer shift in bias towards women, while liberal/conservative newspapers leaned more towards men.
- Nuances in "work" vs. "jobs": Interestingly, while the association with specific "job titles" showed a slight move towards women, the broader concept of "working" gravitated towards men, suggesting a subtle but important distinction in how gender was framed.
- Thematic variations in bias: Bias did not manifest uniformly across all topics. Categories like "Sexuality," "TV," and "Metaphysical Issues" showed a move towards women, while "Money" and "Grooming" moved towards men, even if the overall trend for some categories was weak.
Key Results
- Overall stability with underlying shifts: Combined linear models across all newspapers showed minimal overall shifts in bias, but this masked significant divergences within individual newspapers.
- Social-democratic/religious vs. liberal/conservative: Newspapers with social-democratic (Vrije Volk) and religious (Volkskrant, Trouw) backgrounds demonstrated the clearest shifts in bias towards women. Liberal/conservative newspapers (Telegraaf, NRC, Parool) generally shifted bias towards men.
- Sexuality and money: A notable finding was that across almost all newspapers, the category "sexuality" showed a move towards women, while the category "money" consistently moved towards men.
- Job titles vs. Occupation: The association with job titles showed a slight move towards women (0.05), whereas words from the LIWC category "Occupation" moved marginally towards men (-0.05).
- Vocabulary growth: The size of the vocabulary in the embedding models approximately doubled for some newspapers between 1950 and 1990, indicating a growth in the language used.
Limitations and Future Work
- Limitations of distributional semantics: The authors acknowledge that distributional semantics, which relies on word co-occurrence, may not capture all aspects of gender expression, particularly those related to syntax (e.g., active vs. passive voice).
- OCR quality: The quality of the digitized text varied due to imperfections in the original material and OCR limitations, leading to the exclusion of words not found in a Dutch dictionary.
- Future work: The paper suggests future research could involve fine-tuning more advanced contextualized embedding models like ELMO and BERT to incorporate greater contextual information and potentially capture more nuanced aspects of gender bias. Further detailed analysis of individual trends within each decade and identifying specific words driving divergent behavior are also recommended.
Summary updated