SignBank+: Preparing a Multilingual Sign Language Dataset for Machine Translation Using Large Language Models

TL;DR

This paper presents SignBank+, a cleaned and optimized multilingual sign language dataset that enables effective machine translation between spoken language text and SignWriting, establishing new benchmarks and supporting future research.

Contribution

The creation of SignBank+ dataset, demonstrating that traditional text-to-text translation models perform well on SignWriting translation tasks.

Findings

Models trained on SignBank+ outperform those on the original dataset.

Traditional translation approaches are as effective as complex factorization techniques.

SignBank+ sets new benchmarks for SignWriting-based sign language translation.

Abstract

We introduce SignBank+, a clean version of the SignBank dataset, optimized for machine translation between spoken language text and SignWriting, a phonetic sign language writing system. In addition to previous work that employs complex factorization techniques to enable translation between text and SignWriting, we show that a traditional text-to-text translation approach performs equally effectively on the cleaned SignBank+ dataset. Our evaluation results indicate that models trained on SignBank+ surpass those on the original dataset, establishing a new benchmark for SignWriting-based sign language translation and providing an open resource for future research.