Do online Machine Translation Systems Care for Context? What About a GPT Model?

Example-Based Machine Translation from Text to a Hierarchical Representation of Sign Language ´Elise Bertin-Lem´ee1, Annelies Braffort2, Camille Challant2, Claire Danet2, Michael Filhol2 1SYSTRAN, 5 rue Feydeau, Paris, France, elise.bertinlemee@systrangroup.com 2Universit´e Paris-Saclay, CNRS, LISN, Orsay, France, {annelies.braffort, camille.challant, claire.danet, michael.filhol}@lisn.upsaclay.fr Abstract This article presents an original method for Text-to-Sign Translation. It compensates data scarcity using a domain-specific pa- rallel corpus of alignments between text and hierarchical formal descriptions of Sign Language videos. Based on the de- tection of similarities present in the source text, the proposed algorithm recursively exploits matches and substitutions of ali- gned segments to build multiple candidate translations for a novel statement. This helps preserving Sign Language structures as much as possible before falling back on literal translations too quickly, in a gene- rative way. The resulting translations are in the form of AZee expressions, desi- gned to be used as input to avatar synthe- sis systems. We present a test set tailored to showcase its potential for expressive- ness and generation of idiomatic target lan- guage, and observed limitations. This work finally opens prospects on how to evaluate this kind of translation. 1 Introduction Rosetta 1 is a French project that aimed to study accessibility solutions for audiovisual content. One of the experiments consisted in designing an au- tomatic translation system from text to Sign Lan- guage (SL) displayed through animation of a vir- tual signer. The three main contributions concerning SL in © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribu- tion, CC-BY-ND. 1. https://rosettaccess.fr/index.php/ home-page-english/ this project were 1) the constitution of Rosetta- LSF (Bertin-Lem´ee et al., 2022), an aligned corpus of text and SL captured using a mocap system, 2) a translation system from text to AZee (a represen- tation of SL content), and 3) a system allowing to generate virtual signer animations from AZee in- put (Dauriac et al., 2022). This article describes the second contribution : the translation system from text to AZee. After an overview of the issues and recent works in the field, we explain our method and design choices, and describe the implementation of the translation system. Finally, we give preliminary results and discuss the questions raised for evaluation. 2 Text-to-Sign translation The automatic translation of content from a spo- ken language into a SL is a fairly recent and still largely unexplored research topic. Here we are in- terested in the translation of text as the source lan- guage, in our case in French, and video or 3D ani- mation as the target language, in our case French Sign Language (LSF). In this section, we look at the main challenges encountered with text-to-sign translation. 2.1 Need for bilingual corpora Machine translation (MT) was first developed for spoken languages in their written form using bilingual dictionaries and rule-based systems, that were not easy to develop and maintain. Access to parallel corpora of aligned examples has led to the rise of data-driven approaches, such as Sta- tistical Machine Translation (SMT) that …

Content truncated to 500 words