HTML 77.2%
TypeScript 10.5%
Python 9.6%
JavaScript 2.5%
1<!DOCTYPE html><html lang="en-US" class="no-js localnav-dark scroll-smooth"><head><meta charSet="utf-8" data-next-head=""/><meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover" data-next-head=""/><link rel="icon" href="/favicon.ico" data-next-head=""/><link rel="license" href="/licenses.json" data-next-head=""/><link rel="alternate" type="application/rss+xml" title="RSS feed for Apple Machine Learning Research" href="/rss.xml" data-next-head=""/><meta property="og:type" content="website" data-next-head=""/><meta property="og:image" content="https://mlr.cdn-apple.com/media/Home_1200x630_48225d82e9.png" data-next-head=""/><meta property="og:image:alt" content="Apple Machine Learning Research" data-next-head=""/><meta property="og:image:width" content="1200" data-next-head=""/><meta property="og:image:height" content="630" data-next-head=""/><meta property="og:locale" content="en_US" data-next-head=""/><meta property="og:site_name" content="Apple Machine Learning Research" data-next-head=""/><title data-next-head="">DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation - Apple Machine Learning Research</title><meta name="robots" content="index,follow" data-next-head=""/><meta name="googlebot" content="index,follow" data-next-head=""/><meta name="description" content="Sign language processing systems have traditionally operated at the sentence level, ignoring critical discourse phenomena fundamental to…" data-next-head=""/><meta property="og:title" content="DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation" data-next-head=""/><meta property="og:description" content="Sign language processing systems have traditionally operated at the sentence level, ignoring critical discourse phenomena fundamental to…" data-next-head=""/><meta property="og:url" content="https://machinelearning.apple.com/research/discosign-gloss-translation" data-next-head=""/><link rel="canonical" href="https://machinelearning.apple.com/research/discosign-gloss-translation" data-next-head=""/><meta name="omni_page" content="DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation - Apple Machine Learning Research" data-next-head=""/><link rel="preconnect" href="https://www.apple.com"/><link rel="preconnect" href="https://securemetrics.apple.com"/><link rel="preload" href="/_next/static/css/e8feeb57cc43bc97.css" as="style"/><link rel="preload" href="/_next/static/css/e911e3077d4a1fb5.css" as="style"/><link rel="stylesheet" href="https://www.apple.com/wss/fonts?families=SF+Pro,v3|SF+Pro+Icons,v3"/><script type="text/javascript">(()=>{document.documentElement.classList.remove("no-js"),document.documentElement.classList.add("js");var e=/(iPad|iPhone|iPod)/g.test(navigator.userAgent);if(e){document.documentElement.classList.add("ios");var n=function(){if(/iP(hone|od|ad)/i.test(navigator.platform)){var e=navigator.appVersion.match(/OS (\d+)_(\d+)_?(\d+)?/);return[parseInt(e[1],10),parseInt(e[2],10),parseInt(e[3]||0,10)]}}();n&&n[0]<13&&document.documentElement.classList.add("old-ios")}function d(e){9===e.keyCode&&(document.body.classList.add("kb-user"),window.removeEventListener("keydown",d),window.addEventListener("mousedown",o))}function o(){document.body.classList.remove("kb-user"),window.removeEventListener("mousedown",o),window.addEventListener("keydown",d)}document.documentElement.classList.add(navigator.userAgent.toLowerCase().indexOf("android")>-1?"android":"no-android"),document.documentElement.classList.add(navigator.userAgent.toLowerCase().indexOf("edge")>-1?"edge":"no-edge"),window.addEventListener("keydown",d),e&&(window.onpageshow=function(e){e.persisted&&window.location.reload()})})();</script><link rel="stylesheet" href="/_next/static/css/e8feeb57cc43bc97.css" data-n-g=""/><link rel="stylesheet" href="/_next/static/css/e911e3077d4a1fb5.css" data-n-p=""/><noscript data-n-css=""></noscript><script defer="" noModule="" src="/_next/static/chunks/polyfills-42372ed130431b0a.js"></script><script src="/_next/static/chunks/webpack-5b1dd238f0d0e2cb.js" defer=""></script><script src="/_next/static/chunks/framework-7e56a3cb0b7f58f1.js" defer=""></script><script src="/_next/static/chunks/main-c88564b921566f59.js" defer=""></script><script src="/_next/static/chunks/pages/_app-9ae9cf95ec7b2da2.js" defer=""></script><script src="/_next/static/chunks/3464-15f4a44903f1e6cb.js" defer=""></script><script src="/_next/static/chunks/8353-96e22b1877333961.js" defer=""></script><script src="/_next/static/chunks/8316-9c73e4877a7fee96.js" defer=""></script><script src="/_next/static/chunks/8212-dba4f0cb8c7ad020.js" defer=""></script><script src="/_next/static/chunks/2995-77a9b623861b3048.js" defer=""></script><script src="/_next/static/chunks/7316-ee5160fba2ffc896.js" defer=""></script><script src="/_next/static/chunks/4270-4c758dc099e30aa0.js" defer=""></script><script src="/_next/static/chunks/4379-3529fa6f9904b310.js" defer=""></script><script src="/_next/static/chunks/pages/research/%5Bslug%5D-f21f27fade5530c4.js" defer=""></script><script src="/_next/static/SFgMl2lRh8vCRBNsaVqKa/_buildManifest.js" defer=""></script><script src="/_next/static/SFgMl2lRh8vCRBNsaVqKa/_ssgManifest.js" defer=""></script></head><body class=""><div id="__next"><input type="checkbox" id="localnav-menustate" class="localnav-menustate"/><nav id="localnav" class="localnav localnav-dark localnav-scrim" data-sticky="true" aria-label="Global" role="navigation"><div class="localnav-wrapper"><div class="localnav-background"></div><div class="localnav-content"><div class="localnav-title"><a href="/" aria-label="Machine Learning Research" class="icon icon-before icon-apple">Machine Learning Research</a></div><div class="localnav-menu"><a href="#localnav-menustate" class="localnav-menucta-anchor localnav-menucta-anchor-open" id="localnav-menustate-open"><span class="localnav-menucta-anchor-label">Open Menu</span></a><a href="#" class="localnav-menucta-anchor localnav-menucta-anchor-close" id="localnav-menustate-close"><span class="localnav-menucta-anchor-label">Close Menu</span></a><div class="localnav-menu-tray"><ul class="localnav-menu-items"><li class="localnav-menu-item"><a href="/" class="localnav-menu-link">Overview</a></li><li class="localnav-menu-item"><a href="/highlights" class="localnav-menu-link">Research Highlights</a></li><li class="localnav-menu-item"><a href="/research/" class="localnav-menu-link">Publications</a></li><li class="localnav-menu-item"><a href="/updates/" class="localnav-menu-link">Events</a></li><li class="localnav-menu-item"><a href="/work-with-us" class="localnav-menu-link">Work with us</a></li></ul></div><div class="localnav-actions localnav-actions-center"><div class="localnav-action localnav-action-menucta" aria-hidden="true"><label for="localnav-menustate" class="localnav-menucta"><span class="localnav-menuicon-bread localnav-menuicon-bread-top"><span class="localnav-menuicon-bread-crust localnav-menuicon-bread-crust-top"></span></span><span class="localnav-menuicon-bread localnav-menuicon-bread-bottom"><span class="localnav-menuicon-bread-crust localnav-menuicon-bread-crust-bottom"></span></span><span class="localnav-menucta-chevron d-none"></span></label></div></div></div></div></div></nav><label id="localnav-curtain" for="localnav-menustate"></label><main class="main main-default" role="main"><section><div class="padding nav-padding header-border section-content"><div class="column large-3 small-12"><div class="tag-header"><a href="/research/?domain=Accessibility" class="tag-domain" style="color:var(--sk-fuchsia)"><span class="a11y">research area </span>Accessibility</a>, <a href="/research/?domain=Speech%20and%20Natural%20Language%20Processing" class="tag-domain" style="color:var(--sk-dark_blue)"><span class="a11y">research area </span>Speech and Natural Language Processing</a><span class="post-separator" aria-hidden="true"></span><span><a href="/research/?event=EMNLP" class="tag-conference"><span class="a11y">conference<!-- --> </span>EMNLP</a></span></div></div><div class="column large-8 large-offset-1 small-12 small-offset-0"><div class="font-weight-bold typography-tout"><span><span class="text-capitalize"><a href="/research/" class="tag-type"><span class="a11y">content type </span>paper</a></span><span class="post-separator" aria-hidden="true"></span><span class="a11y">published </span>September 2026</span></div><h1 class="post-title-markdown font-weight-bold typography-headline-reduced"><p>DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation</p></h1><p><span class="a11y">Authors</span>Vasileios Baltatzis‡, Mert Inan‡†**, Connor Gillis, Raja Kushalnagar§**, Lorna Quandt§**, Leah Findlater, Colin Lea</p></div></div></section><section><div class="padding section-content"><div data-testid="links" class="column large-3 small-12 small-order-1"><p><a href="https://arxiv.org/abs/2609.02796" target="_blank" rel="noopener nofollow" class="icon icon-after icon-external" aria-label="View publication - opens in a new window">View publication</a></p><p><button class="link hide-on-old-ios js-only" aria-label="Copy bibtex to clipboard">Copy Bibtex</button><span id="_R_19j5m_" style="position:fixed" hidden=""></span></p><p class="hide-on-small"><button aria-label="share DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation via LinkedIn" class="social-icon js-only"><svg xmlns="http://www.w3.org/2000/svg" width="30" height="30" role="img"><path d="M22.5 7.5H9A1.5 1.5 0 0 0 7.5 9v13.5A1.5 1.5 0 0 0 9 24h13.5a1.5 1.5 0 0 0 1.5-1.5V9a1.5 1.5 0 0 0-1.5-1.5M12.716 21h-2.213v-7.119h2.213zm-1.13-8.137a1.29 1.29 0 1 1 0-2.58 1.29 1.29 0 0 1 0 2.58M21.004 21h-2.211v-3.462c0-.826-.015-1.888-1.15-1.888-1.151 0-1.328.9-1.328 1.828V21h-2.211v-7.119h2.123v.973h.03c.295-.56 1.017-1.15 2.093-1.15 2.24 0 2.654 1.475 2.654 3.392z"></path></svg></button><button aria-label="share DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation via X" class="social-icon js-only"><svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" role="img"><path d="m7.812 8.45 5.898 8.847-5.692 6.753h1.7l4.741-5.626 3.753 5.626h5.198l-6.047-9.072 5.501-6.528h-1.7l-4.551 5.402-3.601-5.402zm2.43 1.3h2.074l8.666 13h-2.074z"></path></svg></button><button aria-label="share DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation via email" class="social-icon js-only"><svg viewBox="0 0 34 34" role="img"><path d="M8 11.57a.6.6 0 0 0 0 .16v10.54a.6.6 0 0 0 0 .12L13.42 17Zm9 7.88a.83.83 0 0 0 .59-.24l8-8H8.44l8 8a.83.83 0 0 0 .59.24m1 .15a1.38 1.38 0 0 1-2 0l-2.23-2.23-5.41 5.42a.6.6 0 0 0 .17 0h16.89a.5.5 0 0 0 .19 0l-5.41-5.42Zm8-8L20.62 17 26 22.35a1 1 0 0 0 0-.08V11.73a.6.6 0 0 0 0-.13"></path></svg></button><button aria-label="Copy link to clipboard" class="social-icon hide-on-old-ios js-only"><svg viewBox="0 0 34 34" role="img"><path d="m14.59 22.28-.91.9a2 2 0 0 1-2.86 0 2 2 0 0 1 0-2.82L14.17 17c.69-.69 2-1.7 3-.76a1.12 1.12 0 1 0 1.53-1.54c-1.62-1.61-4-1.31-6.1.76l-3.35 3.33a4.21 4.21 0 0 0 0 6 4.28 4.28 0 0 0 6 0l.91-.9a1.12 1.12 0 1 0-1.57-1.58M24.75 9.36A4 4 0 0 0 19 9.14l-1.14 1.13a1.12 1.12 0 1 0 1.57 1.58l1.13-1.13a1.81 1.81 0 0 1 2.64.22 2 2 0 0 1 0 2.82l-3.6 3.55c-1.64 1.62-2.4.86-2.73.54a1.12 1.12 0 0 0-1.57 1.58 3.52 3.52 0 0 0 2.51 1.11 4.9 4.9 0 0 0 3.37-1.65l3.58-3.55a4.21 4.21 0 0 0 0-6"></path></svg></button><span id="_R_adj5m_" style="position:fixed" hidden=""></span></p></div><div class="column large-8 large-offset-1 small-12 small-offset-0 small-order-0"><div class="show-on-small mb"><button aria-label="share DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation via LinkedIn" class="social-icon js-only"><svg xmlns="http://www.w3.org/2000/svg" width="30" height="30" role="img"><path d="M22.5 7.5H9A1.5 1.5 0 0 0 7.5 9v13.5A1.5 1.5 0 0 0 9 24h13.5a1.5 1.5 0 0 0 1.5-1.5V9a1.5 1.5 0 0 0-1.5-1.5M12.716 21h-2.213v-7.119h2.213zm-1.13-8.137a1.29 1.29 0 1 1 0-2.58 1.29 1.29 0 0 1 0 2.58M21.004 21h-2.211v-3.462c0-.826-.015-1.888-1.15-1.888-1.151 0-1.328.9-1.328 1.828V21h-2.211v-7.119h2.123v.973h.03c.295-.56 1.017-1.15 2.093-1.15 2.24 0 2.654 1.475 2.654 3.392z"></path></svg></button><button aria-label="share DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation via X" class="social-icon js-only"><svg xmlns="http://www.w3.org/2000/svg" width="32" height="32" role="img"><path d="m7.812 8.45 5.898 8.847-5.692 6.753h1.7l4.741-5.626 3.753 5.626h5.198l-6.047-9.072 5.501-6.528h-1.7l-4.551 5.402-3.601-5.402zm2.43 1.3h2.074l8.666 13h-2.074z"></path></svg></button><button aria-label="share DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation via email" class="social-icon js-only"><svg viewBox="0 0 34 34" role="img"><path d="M8 11.57a.6.6 0 0 0 0 .16v10.54a.6.6 0 0 0 0 .12L13.42 17Zm9 7.88a.83.83 0 0 0 .59-.24l8-8H8.44l8 8a.83.83 0 0 0 .59.24m1 .15a1.38 1.38 0 0 1-2 0l-2.23-2.23-5.41 5.42a.6.6 0 0 0 .17 0h16.89a.5.5 0 0 0 .19 0l-5.41-5.42Zm8-8L20.62 17 26 22.35a1 1 0 0 0 0-.08V11.73a.6.6 0 0 0 0-.13"></path></svg></button><button aria-label="Copy link to clipboard" class="social-icon hide-on-old-ios js-only"><svg viewBox="0 0 34 34" role="img"><path d="m14.59 22.28-.91.9a2 2 0 0 1-2.86 0 2 2 0 0 1 0-2.82L14.17 17c.69-.69 2-1.7 3-.76a1.12 1.12 0 1 0 1.53-1.54c-1.62-1.61-4-1.31-6.1.76l-3.35 3.33a4.21 4.21 0 0 0 0 6 4.28 4.28 0 0 0 6 0l.91-.9a1.12 1.12 0 1 0-1.57-1.58M24.75 9.36A4 4 0 0 0 19 9.14l-1.14 1.13a1.12 1.12 0 1 0 1.57 1.58l1.13-1.13a1.81 1.81 0 0 1 2.64.22 2 2 0 0 1 0 2.82l-3.6 3.55c-1.64 1.62-2.4.86-2.73.54a1.12 1.12 0 0 0-1.57 1.58 3.52 3.52 0 0 0 2.51 1.11 4.9 4.9 0 0 0 3.37-1.65l3.58-3.55a4.21 4.21 0 0 0 0-6"></path></svg></button><span id="_R_a6j5m_" style="position:fixed" hidden=""></span></div><div class="postBody"><p>Sign language processing systems have traditionally operated at the sentence level, ignoring critical discourse phenomena fundamental to sign language comprehension. We introduce DiscoSign, a computational approach for discourse-aware text to sign language gloss translation grounded in linguistic research. We address three key phenomena within our modular Large Language Model (LLM)-based translation framework: (i) spatial coreference resolution, where entities maintain consistent spatial locations throughout discourse; (ii) Question-Answer Clauses (QACs), pseudocleft structures serving specific discourse functions; and (iii) concept-gloss consistency, ensuring stable mappings between English concepts and American Sign Language (ASL) signs. Traditional translation metrics fail to capture discourse-level quality, so we introduce a suite of novel evaluation metrics designed to assess each dimension of discourse coherence addressed by our framework. Experiments on sentence-level and discourse-level datasets show that our approach for discourse-aware processing significantly improves spatial consistency and entity tracking relative to sentence-only translation, while maintaining competitive single-sentence gloss translation quality. Our work establishes the first systematic framework for discourse-level text to sign language gloss translation with corresponding evaluation methodology.</p><ul class="links-stacked"><li>‡ Equal contribution</li><li>† Northeastern University</li><li>§ Gallaudet University</li><li>** Work done while at Apple</li></ul><p></p></div></div></div></section><section class="padding"><div class="header section-content"><h2 class="font-weight-bold typography-headline-reduced">Related readings and updates.</h2></div><div class="cards-alt cards-equal section-content"><div class="cards"><div class="card" data-testid="card-sign-language-annotations"><div class="card-content"><div class="reverse-display-order"><a href="/research/sign-language-annotations" class="header mt-1 post-link"><h3 class="post-title-markdown typography-eyebrow-super"><p>Bootstrapping Sign Language Annotations with Sign Language Models</p></h3></a><p class="mb-0 font-weight-semibold typography-body-reduced"><span>April 30, 2026</span><span class="post-separator" aria-hidden="true"></span><a href="/research/?domain=Accessibility" class="tag-domain" style="color:var(--sk-fuchsia)"><span class="a11y">research area </span>Accessibility</a>, <a href="/research/?domain=Computer%20Vision" class="tag-domain" style="color:var(--sk-sky_blue)"><span class="a11y">research area </span>Computer Vision</a><span class="post-separator" aria-hidden="true"></span><span><a href="/research/?event=CVPR" class="tag-conference"><span class="a11y">conference<!-- --> </span>CVPR</a></span></p></div><div class="card-body"><p>AI-driven sign language interpretation is limited by a lack of high-quality annotated data. New datasets including ASL STEM Wiki and FLEURS-ASL contain professional interpreters and 100s of hours of data but remain only partially annotated and thus underutilized, in part due to the prohibitive costs of annotating at this scale. In this work, we develop a pseudo-annotation pipeline that takes signed video and English as input and outputs a ranked…</p></div><a href="/research/sign-language-annotations" class="button button-outline align-tweaks" aria-label="Read more about Bootstrapping Sign Language Annotations with Sign Language Models">Read more</a></div></div><div class="card card-right" data-testid="card-ai-sign-language-generation"><div class="card-content"><div class="reverse-display-order"><a href="/research/ai-sign-language-generation" class="header mt-1 post-link"><h3 class="post-title-markdown typography-eyebrow-super"><p>Towards AI-Driven Sign Language Generation with Non-Manual Markers</p></h3></a><p class="mb-0 font-weight-semibold typography-body-reduced"><span>March 7, 2025</span><span class="post-separator" aria-hidden="true"></span><a href="/research/?domain=Accessibility" class="tag-domain" style="color:var(--sk-fuchsia)"><span class="a11y">research area </span>Accessibility</a>, <a href="/research/?domain=Human-Computer%20Interaction" class="tag-domain" style="color:var(--sk-green)"><span class="a11y">research area </span>Human-Computer Interaction</a><span class="post-separator" aria-hidden="true"></span><span><a href="/research/?event=CHI" class="tag-conference"><span class="a11y">conference<!-- --> </span>CHI</a></span></p></div><div class="card-body"><p>Sign languages are essential for the Deaf and Hard-of-Hearing (DHH) community. Sign language generation systems have the potential to support communication by translating from written languages, such as English, into signed videos. However, current systems often fail to meet user needs due to poor translation of grammatical structures, the absence of facial cues and body language, and insufficient visual and motion fidelity. We address these…</p></div><a href="/research/ai-sign-language-generation" class="button button-outline align-tweaks" aria-label="Read more about Towards AI-Driven Sign Language Generation with Non-Manual Markers">Read more</a></div></div></div></div></section><section class="full-bleed"><div class="banner-footer double-invert" style="background-image:url(https://mlr.cdn-apple.com/media/Discover_1440x420_2x_9c465d585e.jpg)"><img src="https://mlr.cdn-apple.com/media/Discover_1440x420_2x_9c465d585e.jpg" alt="Bottom banner" loading="lazy"/><div class="section-content"><div class="column large-4 large-last medium-8 medium-centered small-10 small-centered"><h2 class="typography-callout">Discover opportunities in Machine Learning.</h2><p>Our research in machine learning breaks new ground every day.</p><p><a href="/work-with-us" class="button">Work with us</a></p></div></div></div></section></main><footer id="footer" class="footer theme-dark" role="contentinfo" aria-label="Machine Learning Research Footer"><div class="footer-content"><nav class="footer-breadcrumbs" aria-label="Breadcrumbs" role="navigation"><div class="footer-breadcrumbs-path"><ol class="footer-breadcrumbs-list"><li class="footer-breadcrumbs-item"><a href="/" class="home footer-breadcrumbs-link"><span class="footer-icon"></span>Machine Learning Research</a></li><li class="footer-breadcrumbs-item"><a href="/research" class="footer-breadcrumbs-link">Publications</a></li><li class="post-title-markdown footer-breadcrumbs-item typography-inherit"><p>DiscoSign: Discourse-Aware Text to Sign Language Gloss Translation</p></li></ol></div></nav><section class="footer-mini" vocab="http://schema.org/" data-typeof="Organization"><div class="footer-mini-legal"><ul class="footer-mini-legal-links"><li class="footer-mini-legal-links-item"><a class="footer-mini-legal-link" href="https://www.apple.com/legal/privacy/">Privacy Policy</a></li><li class="footer-mini-legal-links-item"><a class="footer-mini-legal-link" href="https://www.apple.com/legal/internet-services/terms/site.html">Terms of Use</a></li><li class="footer-mini-legal-links-item"><a class="footer-mini-legal-link" href="https://www.apple.com/legal/">Legal</a></li></ul><div class="footer-mini-locale">Copyright © <!-- -->2026<!-- --> <a href="https://www.apple.com">Apple Inc.</a> All rights reserved.</div></div></section></div></footer></div><script id="__NEXT_DATA__" type="application/json">{"props":{"pageProps":{"posts":[{"documentId":"e6rs3hkat6lmn8lzb85yrm5m","slug":"discosign-gloss-translation","title":"\u003cp\u003eDiscoSign: Discourse-Aware Text to Sign Language Gloss Translation\u003c/p\u003e","subtitle":null,"description":null,"body":"\u003cp\u003eSign language processing systems have traditionally operated at the sentence level, ignoring critical discourse phenomena fundamental to sign language comprehension. We introduce DiscoSign, a computational approach for discourse-aware text to sign language gloss translation grounded in linguistic research. We address three key phenomena within our modular Large Language Model (LLM)-based translation framework: (i) spatial coreference resolution, where entities maintain consistent spatial locations throughout discourse; (ii) Question-Answer Clauses (QACs), pseudocleft structures serving specific discourse functions; and (iii) concept-gloss consistency, ensuring stable mappings between English concepts and American Sign Language (ASL) signs. Traditional translation metrics fail to capture discourse-level quality, so we introduce a suite of novel evaluation metrics designed to assess each dimension of discourse coherence addressed by our framework. Experiments on sentence-level and discourse-level datasets show that our approach for discourse-aware processing significantly improves spatial consistency and entity tracking relative to sentence-only translation, while maintaining competitive single-sentence gloss translation quality. Our work establishes the first systematic framework for discourse-level text to sign language gloss translation with corresponding evaluation methodology.\u003c/p\u003e\u003cul class=\"links-stacked\"\u003e\u003cli\u003e‡ Equal contribution\u003c/li\u003e\u003cli\u003e† Northeastern University\u003c/li\u003e\u003cli\u003e§ Gallaudet University\u003c/li\u003e\u003cli\u003e** Work done while at Apple\u003c/li\u003e\u003c/ul\u003e\u003cp\u003e\u003c/p\u003e","published":"2026-09-11","type":"paper","paper":null,"publication":null,"publicationPaper":null,"layout":"default","theme":"light","social":null,"organizations":[],"links":[{"type":"publication","url":"https://arxiv.org/abs/2609.02796"}],"code":null,"datasetUrl":null,"tags":[{"documentId":"k8ic47qoyuhsm26kyx435q9u","name":"Accessibility","type":"domain","color":"fuchsia"},{"documentId":"ftc4mkzrerxam95lb9hrcs09","name":"Speech and Natural Language Processing","type":"domain","color":"dark_blue"}],"events":[{"documentId":"igned077nx7clq4bcov3hi92","name":"EMNLP","type":"conference"}],"authors":[],"authorsOrdered":"Vasileios Baltatzis‡, Mert Inan‡†**, Connor Gillis, Raja Kushalnagar§**, Lorna Quandt§**, Leah Findlater, Colin Lea","postZone":[],"features":{"aos":false,"katex":false}}],"related":[{"documentId":"c8nhoherh74a3kbgei58lo8d","slug":"sign-language-annotations","title":"\u003cp\u003eBootstrapping Sign Language Annotations with Sign Language Models\u003c/p\u003e","description":null,"body":"\u003cp\u003eAI-driven sign language interpretation is limited by a lack of high-quality annotated data. New datasets including ASL STEM Wiki and FLEURS-ASL contain professional interpreters and 100s of hours of data but remain only partially annotated and thus underutilized, in part due to the prohibitive costs of annotating at this scale. In this work, we develop a pseudo-annotation pipeline that takes signed video and English as input and outputs a ranked…\u003c/p\u003e","published":"2026-04-30","type":"paper","tags":[{"documentId":"k8ic47qoyuhsm26kyx435q9u","name":"Accessibility","type":"domain","color":"fuchsia"},{"documentId":"yygqscpz3696dh1uvow95c90","name":"Computer Vision","type":"domain","color":"sky_blue"}],"events":[{"documentId":"vvp6lpx8qw7yzu62b9aci1v9","name":"CVPR","type":"conference"}]},{"documentId":"ugf7ksve6ik8brx2xgde8v5n","slug":"ai-sign-language-generation","title":"\u003cp\u003eTowards AI-Driven Sign Language Generation with Non-Manual Markers\u003c/p\u003e","description":null,"body":"\u003cp\u003eSign languages are essential for the Deaf and Hard-of-Hearing (DHH) community. Sign language generation systems have the potential to support communication by translating from written languages, such as English, into signed videos. However, current systems often fail to meet user needs due to poor translation of grammatical structures, the absence of facial cues and body language, and insufficient visual and motion fidelity. We address these…\u003c/p\u003e","published":"2025-03-07","type":"paper","tags":[{"documentId":"k8ic47qoyuhsm26kyx435q9u","name":"Accessibility","type":"domain","color":"fuchsia"},{"documentId":"xtkyjc3amx9o417x0wzzu6s3","name":"Best Paper Honorable Mention","type":"award","color":null},{"documentId":"xo35kqwbzgxvfa2gfw8jeyhd","name":"Human-Computer Interaction","type":"domain","color":"green"}],"events":[{"documentId":"z93207gwn2kge10cxt6l0md0","name":"CHI","type":"conference"}]}],"title":"\u003cp\u003eDiscoSign: Discourse-Aware Text to Sign Language Gloss Translation\u003c/p\u003e","disableSmoothScroll":false},"__N_SSG":true},"page":"/research/[slug]","query":{"slug":"discosign-gloss-translation"},"buildId":"SFgMl2lRh8vCRBNsaVqKa","isFallback":false,"gsp":true,"scriptLoader":[]}</script><script>/* RSID: */2 var s_account="awdmachinelearning"</script><script src="https://developer.apple.com/assets/metrics/scripts/analytics.js"></script><script>3 s.pageName= AC && AC.Tracking && AC.Tracking.pageName();45 /************* DO NOT ALTER THE NEXT LINE ! **************/6 var s_code=s.t();if(s_code)document.write(s_code)7 </script></body></html>