Character-level NMT and language similarity

TL;DR

This paper investigates the performance of character-level neural machine translation with Transformer models across languages with varying similarities and dataset sizes, highlighting when character-level input is advantageous.

Contribution

It demonstrates the conditions under which character-level segmentation improves translation quality and confirms that fine-tuning subword models to character-level enhances results for less related languages.

Findings

Character-level models excel for similar languages.

Subword segmentation outperforms character-level for less related languages.

Fine-tuning subword models to character-level improves translation quality.

Abstract

We explore the effectiveness of character-level neural machine translation using Transformer architecture for various levels of language similarity and size of the training dataset on translation between Czech and Croatian, German, Hungarian, Slovak, and Spanish. We evaluate the models using automatic MT metrics and show that translation between similar languages benefits from character-level input segmentation, while for less related languages, character-level vanilla Transformer-base often lags behind subword-level segmentation. We confirm previous findings that it is possible to close the gap by finetuning the already trained subword-level models to character-level.