Language Contamination Helps Explain the Cross-lingual Capabilities of English Pretrained Models

TL;DR

English pretrained language models contain significant non-English data, which unexpectedly enhances their cross-lingual transfer capabilities, challenging the notion of true monolinguality in large-scale pretraining.

Contribution

This paper reveals that non-English content in English pretraining corpora explains the models' cross-lingual abilities, highlighting the impact of data contamination.

Findings

Non-English data constitutes over 1% of training corpora.

Foreign language tokens in datasets facilitate cross-lingual transfer.

Performance correlates with the amount of in-language data seen during pretraining.

Abstract

English pretrained language models, which make up the backbone of many modern NLP systems, require huge amounts of unlabeled training data. These models are generally presented as being trained only on English text but have been found to transfer surprisingly well to other languages. We investigate this phenomenon and find that common English pretraining corpora actually contain significant amounts of non-English text: even when less than 1% of data is not English (well within the error rate of strong language classifiers), this leads to hundreds of millions of foreign language tokens in large-scale datasets. We then demonstrate that even these small percentages of non-English data facilitate cross-lingual transfer for models trained on them, with target language performance strongly correlated to the amount of in-language data seen during pretraining. In light of these findings, we argue that no model is truly monolingual when pretrained at scale, which should be considered when evaluating cross-lingual transfer.