The Echoes of Multilinguality: Tracing Cultural Value Shifts during LM Fine-tuning

TL;DR

The paper tackles how multilingual LMs encode cultural values and how fine-tuning across languages and data sources shifts these encodings. Using World Values Survey probes and TRAK-based training data attribution, it shows that while pretrained LMs hold cultural information, alignment to human data is limited; fine-tuning can move cultural profiles in language- and dataset-dependent directions, with multilingual fine-tuning generally improving human alignment and preserving cross-language distinctions. The work highlights that the magnitude of value shifts is more tied to data size than to domain or language alone, and that the semantic content of fine-tuning data may not fully explain shifts. Overall, the findings inform practical approaches for culturally-aware multilingual NLP and emphasize the need for language- and data-specific adaptation for value alignment.

Abstract

Texts written in different languages reflect different culturally-dependent beliefs of their writers. Thus, we expect multilingual LMs (MLMs), that are jointly trained on a concatenation of text in multiple languages, to encode different cultural values for each language. Yet, as the 'multilinguality' of these LMs is driven by cross-lingual sharing, we also have reason to belief that cultural values bleed over from one language into another. This limits the use of MLMs in practice, as apart from being proficient in generating text in multiple languages, creating language technology that can serve a community also requires the output of LMs to be sensitive to their biases (Naous et al., 2023). Yet, little is known about how cultural values emerge and evolve in MLMs (Hershcovich et al., 2022a). We are the first to study how languages can exert influence on the cultural values encoded for different test languages, by studying how such values are revised during fine-tuning. Focusing on the fine-tuning stage allows us to study the interplay between value shifts when exposed to new linguistic experience from different data sources and languages. Lastly, we use a training data attribution method to find patterns in the fine-tuning examples, and the languages that they come from, that tend to instigate value shifts.

Figures (14)

• Figure 1: An example of our evaluation setup. We analyze the effect that fine-tuning on different data sources in a fine-tuning language $A$ (Farsi) has on the cultural values that are encoded for a test language $B$ (German). Translation of the example: 'It is _ to do interesting work.' (options: 'important' or 'unimportant').
• Figure 2: The percentage of predictions that remain unchanged after fine-tuning mT5-small.
• Figure 3: Pearson correlation between the average percentage of unchanged values across fine-tuning languages for each data source pair per WVS category.
• Figure 4: Starting from the cultural profiles extracted from pretrained mT5-small, the image depicts in which direction each test language changes depending on the language selected for fine-tuning on PBC (left) and Tanzil (right). The cultural profiles are projected down to 2-dimensions using PCA bro2014principal.
• Figure 5: Spearman correlation between the similarity matrices of the cultural profiles computed from the ground truth data, and pretrained and fine-tuned models.