On compositional generalization of transformer-based neural machine translation

Published: 01 Jan 2024, Last Modified: 08 Feb 2025Inf. Fusion 2024EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We create a benchmark to study compositional generalization of translation models.•Novel compounds pose a challenge on Transformer models including pretrained ones.•Composition exposure during pretraining influences generalization performance.
Loading