Initiative overview
OBJECTIVE: Develop a machine translator for StatCan products, trained on public StatCan products, metadata, and other internal documents. It outputs initial translation drafts for various texts, including sensitive and pre-release material, that language experts will review and edit.
OBSERVATION: Blind test results reveal that customised machine translators outperform off-the-shelf options in translation quality.
TARGET GOAL: Integrate StatCan’s custom machine translator into Translation Services’ existing computer-assisted translation tools to generate initial translations for language experts to review and edit. The tool will be accessible through a user interface that can (a) translate sentences, (b) translate full documents and reproduce the initial format, and (c) be used within existing tools. It is intended to increase operational efficiency and reduce manual workload when and where it is fit for use.
The tool operates in a protected environment accessible to StatCan’s internal language experts. It supports the consistent use of terminology, international standards, and StatCan’s writing style. Machine-translated texts are not perfect, but provide a draft that can be easily edited.
DATA & METHODS: Azure neural machine-translation models were fine-tuned on 395,000 segments of public StatCan products. They were evaluated through expert review using four quality metrics and automated metrics. Quality is planned to be monitored by reviewing translations of recurring StatCan documents, calculating post-editing scores, and collecting quality issues in a form. All machine-translated texts are reviewed. The language expert is at the core of decision-making, and the tool’s versatility allows the expert to choose the best tool and process for each text.
Results, outcomes and impacts
The current baseline of 1,500 words translated per day could increase to 2,500 words per day, which would equal a 66.7% increase, in addition to cost savings from reduced use of external translators. The work distribution strategy is still being fine-tuned, and more data will be available as time progresses and as the tool is fully implemented within the Language Services teams. Data on the number of words post-edited are available, and we expect to see a sharp increase over time. We are working on integrating time metrics as well.




























