Generating Biographies on Wikipedia: The Impact of Gender Bias on the Retrieval-Based Generation of Women Biographies

Angela Fan, Claire GardentACL 2022 · DOI

TL;DR

Researchers have developed a retrieval-augmented system capable of generating full-length, multi-section Wikipedia biographies complete with citations. While the system shows strong performance in document-level coherence, the study reveals a sobering reality: the "evidence gap" on the open web—where women are less frequently documented—directly degrades the quality of AI-generated biographies, particularly for women from marginalized regions.

What Problem Does This Paper Address?

Wikipedia is a cornerstone of global knowledge, but it suffers from a massive gender imbalance; biographies are overwhelmingly written about men. Projects like "Women in Red" aim to fix this manually, but the task is daunting.

From an AI perspective, generating these articles automatically presents three major hurdles:

  1. Information Gathering: How do you find relevant, factual evidence across the messy web?
  2. Long-form Structure: How do you maintain a coherent narrative across multiple paragraphs and sections (e.g., "Early Life" vs. "Career")?
  3. Factuality: How do you ensure the model doesn't "hallucinate" details and provides citations for its claims?

Crucially, the authors ask: If the internet itself is biased and contains less information about women, how does that "digital shadow" affect the performance of state-of-the-art generative models?

Methods

The authors propose a modular architecture that generates biographies section-by-section:

The modular architecture for generating biographies, including retrieval, generation, and citation modules.
  • Retrieval Module: Given a subject's name, occupation, and a section heading, the system queries the web (CommonCrawl). It uses a RoBERTa-based encoder to rank and select the top 40 most relevant sentences. Unlike standard RAG (Retrieval-Augmented Generation), the encoder here is fine-tuned end-to-end with the generator.
  • Generation Module: This uses BART-Large, a standard sequence-to-sequence model. To solve the "long-form" problem, they implement a caching mechanism inspired by Transformer-XL. This allows the model to "remember" hidden states from previous sections, ensuring that the "Career" section doesn't repeat information already mentioned in "Early Life."
  • Citation Module: The system tracks which source documents were used for the generated text and appends bibliographic references at the end of each section, mimicking the standard Wikipedia format.

Datasets

The study utilizes two primary datasets to train and evaluate the system:

  • WikiSum: A standard large-scale dataset of Wikipedia articles paired with web references, used here for primary training (approx. 677,000 biographies).
  • New Evaluation Dataset: The authors curated a novel dataset of 1,527 high-quality biographies of women, categorized into four groups: Women (general), Women in Science, Women in Asia, and Women in Africa.
Comparison of the WikiSum dataset and the specialized evaluation dataset focusing on biographies of women, highlighting the lower overlap of web evidence.

This specialized dataset was designed to test the model's performance in "low-evidence" scenarios. Notably, the average overlap between web hits and the ground-truth biography is significantly lower for this dataset (24.9%) compared to the general WikiSum dataset (39.8%).

Key Insights and Ideas

  • The "Evidence Gap" is a Performance Ceiling: The most significant finding is that model quality is tethered to retrieval quality. For women in Asia and Africa, the lack of "raw material" on the web leads to a measurable drop in generation quality.
  • Context Matters for Retrieval: Using a "detailed query" (Name + Occupation + Section Heading) improved ROUGE-L scores by nearly 2 points compared to just using a name. This suggests that disambiguation is a critical step in the pipeline.
  • Caching vs. All-at-Once: Generating an article section-by-section with a cache is more effective than trying to generate a massive block of text at once. The cache provides the "document-level" context needed to maintain flow without hitting the token limits of models like BART.
  • Fine-tuning as a Partial Fix: The authors found that fine-tuning the model specifically on biographies with less reliable evidence actually improved performance on the marginalized categories, suggesting that models can be "taught" to handle noisy or sparse data more gracefully.

Key Results

  • Quantitative Gains: Adding the retrieval module increased ROUGE-L by 1.4 points over a BART baseline. Adding the caching mechanism added another 0.5 points.
  • Factuality: Human evaluators found that 68% of the information in generated sections was present in the reference text. However, they also found that 17% of the generated info was "hallucinated" (not in the reference or the citations).
  • The Gender/Geography Tax: The model performed worst on biographies of women from Asia and Africa, with ROUGE-L scores roughly 2.5 points lower than the WikiSum average, directly correlating with the lower volume of available web evidence.

Limitations and Future Work

  • English Centricity: The study is limited to English Wikipedia. The authors note that for subjects in non-English speaking regions, the evidence gap is likely even more severe.
  • Style and Tone Bias: Gender bias also manifests in how people are described (e.g., focusing on a woman's marital status or physical appearance). This model does not yet have a mechanism to explicitly detect or correct these stylistic biases.
  • Factuality is Not Solved: While 68% factuality is a start, it is not high enough for a "set it and forget it" autonomous system. The authors envision this as a tool to assist human editors, rather than replace them.

Summary updated

BibTeX
@inproceedings{fan2022generating,
  author = {Angela Fan and Claire Gardent},
  title = {{Generating Biographies on Wikipedia: The Impact of Gender Bias on the Retrieval-Based Generation of Women Biographies}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.586},
}
Figure