Skip to content
Conference Open access

A French Version of the SmolSent Corpus

2026 · Proceedings of the Eleventh Conference on Machine Translation · 0 citations · 67 references

Abstract

We present a French version of the S MOL S ENT corpus, a contribution to the WMT 2026 Open Language Data Initiative (OLDI) shared task. Originally curated to maximize unique-word coverage across 863 English sentences, S MOL S ENT constitutes a valuable, token-efficient training set for the machine translation of low-resource languages. We describe our translation workflow, which relied on a custom-built interface to select and refine translation hypotheses generated by a traditional encoder-decoder model, alongside two state-of-the-art reasoning-enabled large language models (LLMs). Experimental validation based on the MetricX-24-XXL model for quality estimation, supported by bootstrap resampling significance tests, indicates that human post-editing results in significant quality improvements over all raw model outputs, even as reasoning models like Gemma-4-31B-it achieve a competitive 39.51% zero-edit rate. This French partition is not an end in itself, but a pivot resource meant to facilitate ongoing data collection efforts for under-resourced regional languages of France, such as Savoyard and Gallo.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.