— Ch. 1 · Origins And Evolution —
Microsoft Translator.
~4 min read · Ch. 1 of 6
In 1999, Microsoft Research began developing the first version of its machine translation system. This early effort relied on semantic predicate-argument structures called logical forms. The technology originated from a grammar correction feature built into Microsoft Word. By 2000, the team had used this system to translate the entire Microsoft Knowledge Base into Spanish, French, German, and Japanese. That initial project proved that automated translation could handle large-scale technical documentation. A treelet translation system later simplified these logical forms into dependency trees. An order template model followed, which significantly improved processing speed. These shifts allowed engineers to incorporate new target languages much faster than before. In 2007, the consumer-facing site known as Bing Translator launched publicly. It provided free text and website translations directly through a web interface. Users could translate phrases or entire pages using Bilingual Viewer tools. March 2016 marked the introduction of speech translation capabilities. May 2018 brought an API update that made neural machine translation the default method. This update also added transliteration features and a bilingual dictionary for word lookup. Speech translation fully integrated into Microsoft Speech services in September 2018.
Translation Methodologies
Microsoft's approach relies on data-driven algorithms rather than explicit human-written rules. Neural networks mimic brain functions to translate between languages in two distinct stages. First, the system models a specific word based on its context within the full sentence. Second, it translates that internal model into another language while maintaining sentence context. Syntax-based statistical machine translation focuses on translating syntactic units instead of individual words. Engineers used this syntax-based SMT method to translate computer-related texts from English into multiple target languages. Ongoing research produced improvements in word inflections and word ordering. Phrase-based SMT learns correspondence between languages from parallel text without linguist knowledge. This approach generates better translations in less time compared to other systems. Bitext word alignment identifies correspondences between words to train the systems automatically. Microsoft developed both discriminative and generative approaches to word alignment. These methods resulted in faster algorithms and higher quality translations overall. Language modeling uses n-gram models to construct comprehensible translations in the target language. The goal ensures output remains fluent and readable for end users. In November 2016, deep neural networks appeared in nine high-traffic languages including Japanese. Neural networks now provide better translation than industry standard statistical machine learning.