Microsoft Translator

• 1. Introduction

  ›

  Microsoft Translator began not as a product aimed at consumers, but as an internal research project buried inside Microsoft's own knowledge base. Between 1999 and 2000, engineers at Microsoft Research built the first version of the system. Their starting point was something already shipping in Word: the grammar correction feature. That heritage shaped everything that followed. The team stripped the grammar logic down to its core and rebuilt it as a translation engine, one based on logical forms called semantic predicate-argument structures. That engine was powerful enough to translate the entire Microsoft Knowledge Base into Spanish, French, German, and Japanese. What made a machine capable of doing that? And how did a tool built from spell-check DNA grow to support 181 languages across everything from Skype calls to Star Trek dialogue?

• 2. From Word to the World

  ›

  Logical forms, the framework that powered that first system, depend on breaking language into dependency trees. Microsoft's engineers found the approach limiting over time. Their experience with it led directly to a treelet translation system that simplified those dependency trees further. From there the team moved to an order template model, which brought meaningful gains in speed and made it practical to add new target languages. The consumer-facing side of the service arrived in 2007, when the site then called Windows Live Translator launched under the Bing brand. It offered free text and website translation on the web. Text appeared translated directly on the page; whole websites passed through something called the Bilingual Viewer, which let users read the original and the translation side by side with synchronized highlights, scrolling, and navigation. Four layout options were available, including side-by-side, top-and-bottom, and two hover modes. In 2011 a cloud-based API extended the service to developers and enterprise customers. Speech translation arrived in March 2016.

• 3. How the Machine Learns

  ›

  Microsoft's translation approach is described by the company as "data driven": rather than coding explicit rules about language, algorithms are trained on existing translated text in parallel. The system draws on five distinct research areas. Neural networks model a word within the full context of its sentence before translating the model of that word, not the word itself, into the target language. Syntax-based statistical machine translation works on syntactic units rather than individual words or phrases. Microsoft has used that method to handle much of its own computer-related text from English into other languages. Phrase-based statistical machine translation skips linguist input entirely, letting the machine find correspondences between languages on its own, which tends to produce faster results. Bitext word alignment focuses on identifying which words correspond across a pair of translated documents. Microsoft has pursued both discriminative and generative approaches here, aiming for faster algorithms and higher quality. Language modeling relies on n-gram models to keep output readable in the target language. In May 2018 an API update made neural machine translation the default method, and added transliteration and a bilingual dictionary that surfaces example sentences alongside alternative translations.

• 4. Measuring What Machines Get Wrong

  ›

  Translating between languages is easier to attempt than to evaluate. Microsoft Translator uses a metric called the BLEU score, short for Bilingual Evaluation Understudy, which measures how closely a machine translation matches a human one. BLEU was among the first automated metrics to show strong correlation with human judgments of quality. It remains one of the most widely used because it is both automated and inexpensive to run. Statistical algorithms, though, do not guarantee accuracy, and Microsoft has built feedback directly into the service. The Collaborative Translation Framework, now deprecated, let readers suggest alternative translations or vote on ones others had already offered. Those alternatives fed back into the algorithms to shape future output. In November 2016 deep neural networks replaced statistical methods for nine of the highest-traffic languages, including all speech languages and Japanese. That switch produced measurably better translations than the statistical industry standard had achieved.

• 5. A Language for Every Use

  ›

  Custom Translator gives enterprises, developers, and language service providers a way to build neural translation systems trained on the vocabulary of their own industries. A manufacturing company can teach the system terms specific to its supply chain; a legal firm can tune it to courtroom language. The customized output flows back through the standard Microsoft Translator API, so existing applications and workflows receive it without modification. The Live feature takes a different approach to customization, focusing on real-time spoken conversation. It supports up to 500 people across multiple devices, multiple languages, and in-person settings simultaneously, and remains free. The Microsoft Translator Hub, now restricted to statistical machine translation and unavailable on the newest API version, added a dimension beyond commerce: communities have used it for language preservation. Translation systems built through the Hub now exist for Hmong, Mayan, Nepali, and Welsh, among others.

• 6. Partners, Parliaments, and Klingon

  ›

  Senedd Cymru, the Welsh Parliament, partnered with Microsoft Translator under its earlier name, the National Assembly for Wales, to support Welsh. The Government of Nunavut contributed to Inuktitut. Auckland University of Technology and Waikato University both worked on Maori. Jawaharlal Nehru University covered Urdu. Translators Without Borders brought Swahili into the system. The CNGL Centre for Global Intelligent Content and the Baltic language specialist Tilde rounded out Estonian, Latvian, and Lithuanian. One partnership stands apart from the rest. Since May 2013, the Klingon Language Institute, which promotes the constructed language used in the Star Trek universe produced by Paramount and CBS Studios, has supported Klingon's presence in Microsoft Translator. The service carries Klingon in both its Latin transliteration and its native pIqaD script. That places a fictional alien tongue in the same catalogue as 181 other languages and language varieties, sitting alongside Querétaro Otomi, supported by the Government of the State of Querétaro, and Fijian, Tahitian, and Tongan, contributed through the language services company Appen.