TL;DR
This paper investigates whether Large Language Models (LLMs) perpetuate a Western-centric literary canon. By analyzing the output of several LLMs, the authors found that they tend to favor male, White, and American/British authors, mirroring traditional canons and potentially limiting exposure to diverse literary voices.
What Problem Does This Paper Address?
The formation and maintenance of literary canons have long been debated, often reflecting the biases of dominant cultural groups. With the rise of LLMs as tools for discovering and recommending literature, it's crucial to understand if these models are inadvertently reinforcing existing biases or if they can offer a more inclusive view. This paper addresses the question of whether LLMs, through their text generation capabilities, are constructing their own "literary canons" that mirror or even amplify the historical biases found in traditional canons, particularly concerning gender, race/ethnicity, and nationality.
Methods
The study employed a two-pronged approach:
- Book Title Generation: LLMs were prompted to generate lists of book titles and authors using various descriptive terms like "fiction," "classic," "literary canon," "Western literary canon," and "global literary canon." This was done to see how different prompts influenced the generated literary landscape.
- Author Demographic Generation: LLMs were prompted to provide demographic information (gender, race/ethnicity, nationality) for authors. This aimed to evaluate the models' accuracy in recalling and presenting this information.
The researchers then performed manual annotations to create a ground truth dataset for author demographics. They also analyzed the frequency of generated authors and titles, comparing these distributions to established literary canons and demographic baselines.
Datasets
- LitAuthorDemoDB: A newly introduced open-source dataset containing 1,345 authors and 2,238 corresponding titles, with demographic information (gender, race/ethnicity, nationality) for each author. This dataset was manually curated by the researchers and is intended to be community-updated.
- Ground Truth for Demographic Generation: The LitAuthorDemoDB dataset served as the ground truth for evaluating the LLMs' accuracy in generating author demographic information.
Key Insights and Ideas
- LLMs Default to a Western Canon: Regardless of the prompt (even "global literary canon"), LLMs consistently prioritized authors and works associated with the Western literary tradition. This suggests that their training data is heavily skewed towards this perspective.
- Perpetuation of Gender Bias: While LLMs showed more gender diversity than historical canons (e.g., Bennett's or Bloom's), they still fell short of reflecting global gender distributions, with male authors being mentioned more frequently.
- Reinforcement of Racial and Ethnic Homogeneity: LLMs overwhelmingly favored White authors, mirroring the biases found in traditional canons and the publishing industry.
- Nationality Skew: The generated authors were predominantly from the United States and the United Kingdom, with a significant underrepresentation of authors from other parts of the world, even when prompted for a "global" canon.
- Model Accuracy in Demographics: LLMs generally demonstrated good accuracy in generating author demographics, with some variations between models. However, the smaller Llama 3.1 8B model showed more struggles.
Key Results
- Author/Title Frequency: The study generated 30,302 author and title pairs, resulting in 1,347 unique authors and 2,238 unique titles. The frequency of mentions was highly skewed, with Shakespeare being the most mentioned author.
- Top Generated Title: "Pride and Prejudice" by Jane Austen was the top generated title across all four models tested.
- Gender Distribution: Female authors represented 41% of those generated by LLMs, a significant improvement over historical canons but still below global parity.
- Race/Ethnicity Distribution: White authors constituted 78% of the dataset, with Asian, Black, and Latino authors making up smaller percentages.
- Nationality Distribution: American authors accounted for about 52% of the generated authors, followed by British authors at 20%.
- Demographic Generation Accuracy:
- Gender: Accuracy ranged from 0.91 to 0.99 across models.
- Race/Ethnicity: Recall was generally high, with values ranging from 0.86 to 1.00 depending on the category and model.
- Nationality: Recall was also high, ranging from 0.87 to 0.98.
Limitations and Future Work
- Prompt Language and Design: The study was limited to English-language prompts, potentially privileging English-speaking perspectives. Future work should include prompts in other languages and test models developed in different regions.
- Model-Specific System Prompts: The researchers did not use model-specific system prompts, which could have influenced output.
- Post-processing and Annotation: A more robust post-processing system for demographic evaluation and a larger scale of inter-annotator agreement for ground truth labels are suggested.
- Copyright Data: The study acknowledges that models were likely trained on copyrighted data, raising ethical considerations.
- Environmental Cost: Running LLM experiments incurs significant environmental costs.
- Future Work: The authors suggest exploring LLMs in different languages, varying sampling parameters, investigating genre literature diversity, including other demographic information (like LGBTQIA+ status), and evaluating LLMs on more complex tasks like question-answering about diverse authors. They also encourage contributions to their LitAuthorDemoDB to help retrain LLMs for a more representative canon.
Summary updated