The unfulfilled promise of MTPE
Localization professionals are no strangers to using technology to streamline the translation process. It’s standard practice, for instance, for translators to use software to split a text into manageable segments, apply previous translations, and run quality assurance checks. In addition, translators and businesses alike have relied on machine translation (MT) to generate rough translations for years. Improvements in MT have also led to increased interest in MTPE (or machine translation post-editing) — that is, MT output edited by a human.
Even so, many believe that MTPE produces inferior results compared to having one human translate a piece of content and another edit it. As a result, businesses sensitive to quality have been reluctant to embrace MT and MTPE. Though more expensive, the traditional human-driven approach has traditionally been favored, especially by companies in heavily regulated industries such as pharmaceuticals, life sciences, legal, and finance.
In short, there hasn’t been a generally accepted substitute for a human-driven translation workflow — until now.
There’s a tectonic shift underway in the translation and localization industry. The maturation of neural machine translation (NMT), advances in artificial general intelligence (AI), and the introduction of large language models (LLMs) have made things that were once unthinkable possible.
Revolutionize your translation mix: AI-Powered Human Translation
AI-Powered Human Translation (AIHT) fixes the quality gap between MTPE and human translation. Our AIHT solution delivers a remarkable MQM score of 98+, equivalent to humans, while reducing per word costs by up to 50% and cutting delivery time in half compared to traditional human translation.
AIHT seamlessly integrates the strengths of AI, Large Language Models (LLMs), machine translation, and human expertise into a unified workflow, enhancing every step with AI-driven efficiency.
Furthermore, advances in AI are also enabling Machine Translation to produce outstanding quality scores of up to MQM 93 for a fraction of a penny per word cost!
Laying the Foundation
Traditionally, the workflow begins with the application of translation memory, a saved record of all your previous translations. New content is compared to this record and, where there is a match of a certain confidence level, the saved translation is used instead of needing to translate from scratch. Translation memory can save companies anywhere from 30-70% on per word costs.
The difference with Smartling is using AI to expand the coverage of your translation memory to improve cost savings by enhancing lower confidence matches to make them more accurate. We perform a process called “fuzzy match repair” using our in-house Smartling MT engines to improve the fit of these lower confidence matches.
Content that translation memory couldn't address is sent to our AI translation step. Our machine learning quality evaluation tool assesses the output of multiple machine translation engines and selects the highest-quality option to be used.
With initial translations ready, it's time for post-processing, a critical step to ensure ready to use and brand compliant translations. We begin with the application of MT and LLM-optimized glossaries (your list of important terms and how they should be translated), which goes beyond substitution. The distinction lies in ensuring that these substitutions fit the context of the string and are grammatically correct, rather than simply replacing the term.
We also automate content and format cleanup, addressing issues such as whitespaces, missing or extra tags, and placeholders. This formatting step is important because sometimes MT engines alter formatting, causing errors when strings are re-ingested into the platform or impacting the cleanliness of the translation memory.
In-Context Linguistic Review
Lastly, the translation is reviewed in context by an expert linguist handpicked for the project. This is what we call human-in-the-loop validation.
Our technology has already done the heavy lifting – it translated the text and ensured grammatical accuracy while making adjustments to align with brand guidelines. And that frees the professional linguist to focus on higher-level validation and polishing.
All work is completed within the Smartling platform. So every change is tracked and saved in real time, eliminating version control confusion. If questions arise, the linguist can communicate directly with you in Smartling.
Human Quality at Half the Cost
All of this sounds good but it only works if human quality can truly be delivered.
In our latest report, we shared our approach to measuring translation quality and the results we achieved across workflows. Smartling's rigorous process includes random sampling across multiple languages each month and a thorough review according to the Multidimensional Quality Metrics (MQM) framework - the industry standard for quality evaluation.
Our research found that all of Smartling’s translation workflows consistently achieve high MQM scores, including AIHT.
Industry benchmarks for human translation ranges between 95% and 97%. With an average MQM score of 98, AI-Powered Human Translation is actually exceeding traditional human translation outputs from many language service providers. And, it does this while reducing per word cost by 50% and improving time to market by 2x.
AIHT is having a huge impact on our customers.
For example, a major enterprise, with an annual translation volume exceeding 20 million words, revolutionized their localization strategy using Smartling’s AI-Powered Human Translation. This cutting-edge approach delivered substantial cost savings, accelerated turnaround times, and maintained exceptional human translation quality. Before switching to Smartling, their translation needs were met through traditional human translation services.
Want to learn more about how AI-Powered Human Translation can improve your localization program? Get in touch.