Abstract
Multilingual Language Models (MLMs) do not merely learn linguistic structures; they internalize the cultural values embedded within their training corpora. This paper investigates how fine-tuning an MLM in one language inadvertently shifts its cultural profile across others. By probing models with the World Values Survey (WVS), the researchers demonstrate that cross-lingual sharing causes cultural "bleed-over," where the values of a fine-tuning dataset in one language alter the model's responses in another. The study reveals that these shifts are often driven by internal model architecture and structural data patterns rather than semantic meaning alone.
Introduction
When a single model is trained on over a hundred languages, those languages do not exist in isolation. MLMs rely on cross-lingual sharing to function, which creates a pathway for cultural beliefs and biases to migrate between linguistic representations. While the encoding of cultural values in pre-trained models is well-documented, the evolution of these values during the task-specific fine-tuning stage remains poorly understood. This research seeks to map these "echoes" of multilinguality to understand how fine-tuning steers a model’s cultural compass across its entire linguistic repertoire.
Methods and Experimental Framework
The researchers employed a multi-step framework to quantify cultural shifts:
- Cultural Probing: The team used "cloze-style" (fill-in-the-blank) prompts derived from the World Values Survey (WVS). For example, a model might be asked to complete: "It is [MASK] to do interesting work" with options like important or unimportant.
- Fine-tuning Strategies: They compared monolingual fine-tuning (training on one language) against multilingual fine-tuning (training on several languages simultaneously) to observe differences in value preservation.
- Quantification: Cultural profiles were constructed as N-dimensional vectors representing the probability of the model selecting specific values. These were then compared to human survey data using Spearman correlation.
- Training Data Attribution (TDA): Using the TRAK method, the researchers traced specific shifts back to individual training examples to identify which sentences "instigated" changes in the model's values.
Datasets and Languages
- World Values Survey (WVS): Wave 7 (2017-2020) provided 237 probes across 13 categories, including Social Values, Religious Values, and Corruption.
- Fine-tuning Sources: To test domain bias, the study used Flores-101 (Wikipedia/neutral), Parallel Bible Corpus & Tanzil (Religious/didactic), and All-The-News (News/political).
- Languages: Models were fine-tuned on four languages (Farsi, Korean, Hindi, Russian) and tested across 13 diverse languages, including Romanian, Urdu, and German.
Key Insights
The "Bleed-Over" Effect
Fine-tuning in a single language significantly impacts others, though the magnitude of the shift depends heavily on the test language. Farsi, for instance, appeared particularly "brittle," with its encoded values changing more easily than other languages regardless of the training data used.
Directionality and Structure
Surprisingly, the direction of a cultural shift is determined more by the language being tested than the language used for fine-tuning. Most fine-tuning setups pushed the model's various languages in similar directions in the cultural vector space. Furthermore, TDA analysis showed that the model often relies on the same training examples to trigger shifts across different test languages, suggesting that structural patterns in the data may be more influential than semantic meaning.
Multilingual Fine-tuning vs. Diversity
While monolingual fine-tuning tends to collapse the cultural distances between languages, making them more homogenous, multilingual fine-tuning better preserves the distinct cultural nuances found in real-world human data.
Key Results
- Initial Misalignment: Pre-trained MLMs like mT5 initially show very poor correlation with actual human survey data from the regions where those languages are spoken.
- Significant Volatility: Fine-tuning on as few as 2,000 sentences caused 7% to 43% of the model's cultural predictions to change.
- Domain Sensitivity: Religious texts (Bible/Quran) generally induced larger value shifts than neutral Wikipedia or news content.
- Accidental Alignment: Fine-tuning often increased the model's correlation with human data, suggesting that the "blank slate" of a pre-trained model is further from human values than a model steered by specific, even if biased, data.
Limitations and Conclusion
The study acknowledges that mapping a single language to a single country's values is a simplification, as most languages serve diverse subcultures. Additionally, the research was constrained by the availability of multi-parallel datasets. Future work is required to develop "culturally-aware" fine-tuning techniques that allow models to adapt to specific communities without eroding the values of others. This research highlights the necessity of considering cultural shifts as a primary factor in the safety and alignment of global AI systems.
Summary updated