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

TL;DR

This paper investigates how multilingual language models encode and shift cultural values during fine-tuning, revealing the influence of new data sources and languages on these cultural representations.

Contribution

It is the first study to analyze how fine-tuning affects cultural value encoding and cross-lingual influence in multilingual language models.

Findings

Cultural values in MLMs can shift during fine-tuning.

Fine-tuning data sources influence the direction of value shifts.

Cross-lingual influence affects cultural value encoding.

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.