Introduction
Machine Translation (MT) systems often inherit and amplify societal biases. A classic example is a model translating "The doctor decided to bring her phone" into a language with grammatical gender (like Spanish) and incorrectly choosing the masculine form for "doctor" due to occupational stereotypes.
The challenge in fixing this lies primarily in the data. Most training and evaluation sets are heavily unbalanced, containing far more examples of men in certain professions than women. While synthetic datasets using templates exist, they often lack the complexity and nuance of real-world language. To address this, researchers have introduced the OCCGEN toolkit—a new open-source pipeline designed to extract real-world, multilingual parallel datasets that are strictly balanced by gender across various occupations.
The OCCGEN Pipeline
The authors developed the OCCGEN pipeline to automate the extraction of balanced, high-quality data. The methodology consists of four main stages:
- Information Extraction: The system queries Wikidata to find entities (people) with their associated occupations, gender, and links to their biographies across multiple languages.
- Biography Scraping & Preprocessing: It pulls the raw text from Wikipedia biographies and cleans it by removing dates, phonetics, and other non-essential artifacts.
- Dataset Alignment: To find "parallel" sentences (sentences that carry the same meaning in different languages), the pipeline uses LASER embeddings and a margin-based criterion. For multilingual sets, it uses English as a "pivot" language to find overlapping sentences across all target languages.
- Dataset Balancing: This is the core innovation. The toolkit ensures that for every occupation (e.g., "politician" or "scientist"), the final dataset contains an equal number of sentences and entities for each gender. It prioritizes high-quality alignments to ensure the data is genuinely useful for MT evaluation.
Datasets and The "Scarcity" Reality Check
Using the toolkit, the researchers introduced two new evaluation benchmarks:
- High-Resource Multiparallel Set: Covers four distinct linguistic families: Arabic (Semitic), English (Germanic), Russian (Slavic), and Spanish (Latin).
- Low-Resource Parallel Set: Focuses on English-Swahili.
Unlike previous datasets limited to two languages or synthetic patterns, these are derived from real Wikipedia biographies and balanced specifically to isolate occupational gender bias. However, the process of balancing this data revealed a stark reality: Wikipedia has roughly three times more masculine representations than feminine ones.
After the balancing and alignment process, the number of usable entities drops drastically. For instance, in the high-resource set, the pipeline starts with 15,421 entities but yields only 286 after strict balancing. This highlights exactly how difficult it is to find high-quality, balanced, real-world data in the wild.
Key Findings and Translation Disparities
By testing state-of-the-art translation models on these newly balanced benchmarks, the researchers uncovered several critical insights:
- Persistent Masculine Bias: In high-resource languages, translation performance (measured by BLEU scores) was consistently better for masculine subsets than feminine ones, especially when English was involved.
- The English Generation Paradox: Models often show higher BLEU scores when translating into English or Spanish. The authors hypothesize that these "strong" language models might actually be too good at generating fluent text, causing them to pay less attention to the source sentence and default to the most frequent (usually masculine) gender forms.
- Morphology Matters: Languages with rich morphological information (like Arabic and Russian) showed a much wider gap in accuracy between masculine and feminine translations compared to low-inflected languages.
- Human Evaluation: Human annotators confirmed the automatic metrics. For English-to-Arabic translation, the accuracy for masculine terms was 0.78, while feminine terms plummeted to 0.41. Spanish showed the smallest gap but still favored the masculine (0.71) over the feminine (0.62).
- Low-Resource Surprise: Interestingly, for the Swahili-English use case, the feminine subset actually performed slightly better than the masculine in some models, likely because both languages are relatively low-inflected regarding gender.
Limitations and Future Work
While OCCGEN is a significant step forward, the authors note a few limitations. First is the Binary Trap: although the toolkit is customizable for any gender definition, the source data (Wikipedia/Wikidata) lacks sufficient representation for non-binary or other gender identities to create a statistically balanced set, forcing the current benchmarks to rely on binary (masculine/feminine) categories.
Additionally, the strict balancing requirements result in relatively small evaluation sets. To combat this data volume issue, the authors suggest that counterfactual data generation (automatically flipping the gender of a sentence) could be a necessary next step to augment these real-world datasets. Finally, because the datasets come from different people's biographies, the masculine and feminine subsets naturally have different vocabularies, which can slightly skew standard metrics like BLEU.
The OCCGEN toolkit and datasets are available on GitHub for researchers looking to evaluate and mitigate bias in their own models.
Summary updated