TL;DR
This paper tackles the critical challenge of evaluating biases in language models across diverse languages and cultural contexts, an area significantly lacking resources beyond English. They introduce a multilingual extension of the CrowS-Pairs dataset, enabling more robust and comparable bias assessments for models operating in a globalized digital landscape.
What Problem Does This Paper Address?
The field of Natural Language Processing (NLP) has made strides in identifying and mitigating biases within language models. However, this research has largely been confined to English and Western cultural contexts. This creates a significant gap, as the vast majority of the world's population interacts with technology in languages other than English, and cultural nuances heavily influence how stereotypes manifest and are perceived. Existing bias evaluation datasets are often limited in scope, focusing on a narrow range of biases (like gender) and lacking multilingual support. This paper aims to bridge this gap by providing a resource that allows for the evaluation of biases across multiple languages and cultural settings, moving beyond the limitations of English-centric research.
Methods
The core of this work involves extending the existing CrowS-Pairs dataset, a benchmark for measuring social biases in masked language models (MLMs). The authors translated and adapted the original English dataset into seven additional languages: Arabic (Maghreb), Catalan, German, Spanish (Argentina), Italian, Maltese, and simplified Chinese. This translation process was meticulous, involving native speakers with linguistics and NLP backgrounds who worked in pairs to ensure accuracy and cultural appropriateness. They also revised existing English and French datasets to address issues like "non-minimal pairs" (sentences that differ by more than one token, making bias assessment less precise). The resulting multilingual dataset was then used to evaluate 16 monolingual MLMs and two multilingual models (mBERT and XLM-ROBERTa).
Datasets
The primary dataset used and extended in this paper is CrowS-Pairs. The authors created a new multilingual version, Multilingual CrowS-Pairs, by translating the original English dataset into seven new languages, adding to the existing English and French versions. This new dataset comprises 11,139 sentence pairs covering nine types of biases across seven cultural contexts. They also mention revising 275 pairs from the original English CrowS-Pairs dataset and 252 pairs from the French dataset to address issues like non-minimal pairs.
Key Insights and Ideas
- Cultural Context is Crucial: Stereotypes are not universal. What might be a clear stereotype in one culture could be nonsensical or even offensive in another, highlighting the need for culturally specific datasets.
- Translation is More Than Word-for-Word: Adapting stereotypes across languages requires more than direct translation; it involves understanding cultural references, idioms, and societal norms to create equivalent meaning. Some stereotypes, like those referencing specific cultural groups (e.g., KKK), proved untranslatable.
- Language Models Show Consistent Bias: Across all tested languages and models, language models generally exhibited a preference for sentences expressing stereotypes, underscoring the pervasive nature of bias in NLP.
- Monolingual Models Can Outperform Multilingual Ones: Surprisingly, monolingual models sometimes showed higher bias scores than their multilingual counterparts, suggesting that specialized models might be more susceptible to biases within their specific linguistic and cultural training data.
- The Challenge of Comparability: Creating a resource that spans multiple languages and cultures reveals the inherent difficulty in achieving true comparability in bias evaluation. Linguistic structures and cultural specificities make direct cross-lingual comparisons challenging.
Key Results
The paper reports that their newly created Multilingual CrowS-Pairs dataset contains 11,139 sentence pairs. Experiments using this dataset revealed that language models in all tested languages favored stereotypical sentences across most bias categories. While specific quantitative results for each model are presented in Table 3, a general trend observed is that models with a RoBERTa architecture tended to exhibit higher bias scores compared to BERT models. The authors also note that the overall metric score for monolingual models was often higher than that of multilingual models for the same language, though exceptions exist.
Limitations and Future Work
The authors acknowledge that their resource, while expanded, is still limited to the specific cultural contexts of the seven languages addressed. They also note that the dataset is primarily intended for masked language models, and its application to autoregressive models like LLMs requires further investigation. Future work could involve broader application of the resource for more extensive bias evaluation experiments and analyses across a wider range of languages and cultural settings. The authors also suggest that further research is needed to adapt bias evaluation frameworks for newer LLM architectures.
Summary updated