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

Texts written in different languages reflect different culturally-dependentbeliefs of their writers. Thus, we expect multilingual LMs (MLMs), that arejointly trained on a concatenation of text in multiple languages, to encodedifferent cultural values for each language. Yet, as the 'multilinguality' ofthese LMs is driven by cross-lingual sharing, we also have reason to beliefthat cultural values bleed over from one language into another. This limits theuse of MLMs in practice, as apart from being proficient in generating text inmultiple languages, creating language technology that can serve a communityalso 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 exertinfluence on the cultural values encoded for different test languages, bystudying how such values are revised during fine-tuning. Focusing on thefine-tuning stage allows us to study the interplay between value shifts whenexposed to new linguistic experience from different data sources and languages.Lastly, we use a training data attribution method to find patterns in thefine-tuning examples, and the languages that they come from, that tend toinstigate value shifts.