Training a Bilingual Language Model by Mapping Tokens onto a Shared Character Space
Aviad Rom, Kfir Bar
TL;DR
The paper addresses cross-lingual transfer between Arabic and Hebrew by unifying scripts through transliteration, training a bilingual encoder (HeArBERT) on transliterated Arabic and Hebrew data. It demonstrates that transliteration facilitates better cross-lingual MT, with the strongest results when HeArBERT initializes the Hebrew decoder and when paired with GigaBERT, achieving competitive BLEU scores despite substantially less pre-training data. The work provides a data-efficient path for related languages and offers a transliteration-based mechanism to align token representations across scripts. It also discusses limitations of a deterministic transliteration and outlines directions for scaling data and extending transliteration into decoder architectures.
Abstract
We train a bilingual Arabic-Hebrew language model using a transliterated version of Arabic texts in Hebrew, to ensure both languages are represented in the same script. Given the morphological, structural similarities, and the extensive number of cognates shared among Arabic and Hebrew, we assess the performance of a language model that employs a unified script for both languages, on machine translation which requires cross-lingual knowledge. The results are promising: our model outperforms a contrasting model which keeps the Arabic texts in the Arabic script, demonstrating the efficacy of the transliteration step. Despite being trained on a dataset approximately 60% smaller than that of other existing language models, our model appears to deliver comparable performance in machine translation across both translation directions.