Cognate Production Using Character-Based Neural Machine Translation Without Segmentation

Cognates are words that share a common origin or have been borrowed across languages, often exhibiting similarities in both sound and meaning. In this work, we introduce a fully character-level neural sequence-to-sequence model for cognate production that does not require any segmentation. Our model operates at the character-level to transform a source word into its corresponding cognate in the target language, thereby obviating out-of-vocabulary issues and alleviating the need for subword segmentation. We evaluated our approach on a novel dataset and found that it outperforms both statistical machine translation baselines and prior neural methods on the same training dataset, as measured by standard coverage and mean reciprocal rank metrics. These results underscore the effectiveness of character-level sequence-to-sequence architectures for cognate generation in diverse language settings, including cross-alphabetic transformations.