Persistent MT on software technical documentation - a case study

Persistent MT on software technical documentation - a case study Mar´ıa Concepci´on Laguardia Senior Information Developer Citrix Systems Cambridge, UK conchita.laguardia@citrix.com Abstract We report on the features and current challenges of our on-going implementa- tion of a Persistent MT workflow for Cit- rix Product Documentation, to increase localization coverage to 100% content in docs.citrix.com, into our Tier-11 lan- guages. By the end of 2019, we had processed seven million words of English documen- tation with this model, across the 24 doc sets of the Citrix portfolio (Digital Workspace, Networking, and Analytics), and raised localization coverage from 40% to 100% of the content of our documenta- tion repositories. The current implementation requires a pro- cess of Light Post-editing (LPE) for all lan- guages, in order to fix over-translations, out-of-domain words, inline tags, and markdown errors in the raw output. 1 Background The Localization team at Citrix Systems intro- duced Machine Translation in its documentation workflows in 2013, with an in-house implementa- tion of Moses engines from English into German, French, and Spanish. Statistical MT was the basis for a full post-editing workflow performed by in- house linguists and a small pool of trusted contrac- tors. With the advent of Neural Machine Transla- tion in 2017, we switched to Google generic en- c⃝ 2020 The authors. This article is licensed under a Creative Commons 3.0 licence, no derivative works, attribution, CC- BY-ND. 1Citrix Tier-1 languages: German, French, Spanish, Japanese and Simplified Chinese. gines, and later on incorporated Amazon Trans- late and Microsoft Custom Translator. The adop- tion of cloud Translation Management Platforms with built-in connectors for the main generic Neu- ral MT providers, allowed us to mature the post- editing workflows and expand them to Japanese and Simplified Chinese. This now classic work- flow of MT pre-translation followed by full human post-editing has since been our default model for the documentation use-case. Productivity increased dramatically since intro- ducing NMT as a base for full post-editing. This had a positive impact on time-to-market and reduc- tion of localization costs. However, the actual cov- erage of our localization infrastructure remained at 40% of all the Citrix Product Documentation vol- ume, mainly into Tier-1 languages, with a few doc sets into Korean and Brazilian Portuguese. The cost of sustained localization of the remaining con- tent using MT plus human full post-editing was prohibitively high. It was evident that just integrat- ing MT in the workflows was not enough to pro- duce global-ready content at the increasing speeds and volumes necessary in today’s competitive mar- ket. 2 The plan Encouraged by the better fluency and adequacy of NMT output, at the end of 2018, it was proposed to expand our localization coverage to the remain- ing 60% wordcount of our documentation by using raw MT (MT without post-editing). By then we were familiar with Microsoft’s suc- cessful implementation of raw MT in their tech- nical and end-user support documentation, which Microsoft now considers ”proven for all support content types” (Schmidtke and Groves, 2019). We particularly liked …

Content truncated to 500 words