Serving a wide range of sectors, TransPerfect is a world leader in language and technology solutions. Since its founding in 1992, TransPerfect has expanded to become a company with over 10,000 workers spread throughout 140 cities on six continents. Translation, localization, interpretation, multicultural marketing, website globalization, voiceovers, subtitling, and legal support services are just a few of the many services the company provides. Additionally, TransPerfect leverages state-of-the-art technology to provide AI-powered language solutions, like its in-house GlobalLink translation management system.
This article details the collaboration between TransPerfect and the AWS Customer Channel Technology – Localization Team to include Amazon Bedrock into the GlobalLink translation management system, a cloud-based tool that assists businesses in managing their multilingual content and translation processes. TransPerfect’s service enables businesses to leverage AI to quickly produce and distribute content at scale in several languages.
The deployment and administration of generative AI models are made easier with Amazon Bedrock, a fully managed service. It provides developers with access to a range of foundation models (FMs), facilitating the effective development and scaling of AI systems. Because of its high scalability, security, and ease of integration with other AWS services, Amazon Bedrock can be used for a wide range of applications, including language translation.
The localization team of AWS Customer Channel Technology has been a TransPerfect customer for a long time. At AWS, the team oversees the complete translation process for digital material, including banners, movies, ebooks, technical documentation, webpages, and more. The AWS team manages digital assets with billions of words in several languages. The AWS team must handle an ever-increasing load and a broader range of languages due to the growing demand for multilingual content from globally minded companies and new local cloud adoption journeys. The group uses the GlobalLink technology suite to automate and optimize translation procedures in order to achieve this.
The Difficulty
Each year, billions of words are translated and delivered thanks to the efficient bespoke workflows and toolkits developed by the AWS team and TransPerfect. A minimum of asset handoff, asset preparation, machine translation, post-editing, quality review cycles, and asset handback comprise the multi-step process of content localization. These procedures are frequently time-consuming, expensive, and manual. AWS and TransPerfect are always working to streamline this process so that more content may be processed at a reduced cost and that the time to market for those assets is shortened, allowing non-English speaking customers to access valuable, salient content more quickly. Transcreating creative content also presented a special problem because it has historically required highly qualified human translators and was difficult to automate, which led to increased expenses and longer turnaround times. TransPerfect collaborated with AWS to assess generative AI-driven projects for transcreation and automated post-editing within TransPerfect’s GlobalLink architecture in order to address these problems.
Data Safety And Security
Amazon Bedrock assists in ensuring that data is not utilized to enhance base models or shared with FM providers. Because Amazon Bedrock is a FedRAMP-authorized service and complies with important compliance standards including ISO and SOC, it is appropriate for government contracts. TransPerfect is able to comply with strict auditability standards thanks to Amazon Bedrock’s robust monitoring and logging features.
Responsible AI is just one of several considerations, even though data security is a crucial need. TransPerfect was able to create and modify honesty safeguards for the automated post-edit service thanks to Amazon Bedrock Guardrails. Hallucinations can cause large language models (LLMs) to produce inaccurate information. If the responses are contradictory or factually inaccurate, Amazon Bedrock’s contextual grounding checks can identify and filter hallucinations. For a translation solution that demands flawless accuracy, this is an essential element.
Using LLMs To Perform Post-Editing Automatically
AWS team workflows leverage machine translation powered by Amazon Translate to translate at scale. Machine translation procedures are used for segments whose translations cannot be reused from translation memory, which are databases of prior excellent human translations. Amazon either employs machine translation post-edit workflows or machine translation-only workflows, where material is translated and published without human intervention, depending on the language or content. Post-editing is the process by which a linguist checks that the machine-translated output of a particular segment accurately captures the original sentence’s meaning and conforms to established glossaries and AWS style rules. Automating any or all of this procedure would significantly reduce costs and turnaround times because it can add days to the translation timeframe.
The machine translation workflow is depicted in the following diagram.
The following elements make up the workflow:
Translation memory, or TM, is a client-specific database of previously translated and authorized material. It maximizes the reuse of pre-existing translations and is always applied first.
Machine translation (MT): Amazon Translate is used to handle fresh content through machine translation once pre-existing translations have been applied.
APE (automatic post-edit): Machine-translated content is edited, enhanced, and corrected by an LLM.
Human post-editing, or HPE, is the process by which a linguist with subject-matter expertise edits and refines machine-translated text.
Several years ago, TransPerfect started collaborating with generative AI and LLMs because they anticipated that AI will revolutionize the translation sector. As anticipated, localization workflows are aiming toward “no human touch” models and have mostly moved to “expert in the loop” models. To better automate and improve these processes, TransPerfect decided to integrate Amazon Bedrock into its GlobalLink Enterprise system. By design, Amazon Bedrock offers data security and ownership. For clients of TransPerfect, particularly those in delicate sectors like banking or the biological sciences, this is an essential aspect.
Machine-translated content is now sent through one of the LLMs offered by Amazon Bedrock for automated post-editing thanks to GlobalLink and Amazon Bedrock. The LLM is pushed to enhance current machine translations by utilizing style guides, pertinent samples of authorized translations, and examples of mistakes to avoid. This post-edited material is either used in “no human touch workflows” to significantly enhance the output or sent to a linguist for a lighter post-edit (a less challenging task). As a result, there is an overall improvement in quality, and post-editors can concentrate on more valuable corrections.
Over 95% of the post-editing changes recommended by Amazon Bedrock LLMs demonstrated noticeably better translation quality, which reduced the overall cost of translations for Transperfect by up to 50% and freed up human linguists for more complex work.
Using LLMs To Facilitate Transcreation
Machine translation has typically done poorly when it comes to creative content that leans toward complexity, subtlety, humor, descriptiveness, and cultural references, despite its strong performance in technical, formal, and instructive information. When translated by a machine, creative information may sound artificial or stiff. For this reason, TransPerfect has historically relied on manual transcreation of this kind of information by human linguists.
The process of translating a communication from one language to another while preserving its context, tone, style, and intent is known as transcreation. For instance, Nike’s slogan, “Just do it,” is translated into German as “Du tust es nie nur für dich,” which translates to “you never do it just for yourself.”
A communication that has been successfully transcreated has the same connotations and elicits the same feelings in the target language as it does in the original. To increase their effect in a particular industry, the AWS team leverages transcreation for extremely innovative marketing assets. However, because transcreation is a highly specialized and creative process, it has not historically benefited from the automated solutions utilized in other forms of localization operations. This indicates that generative AI has generated a lot of attention in an attempt to reduce the expenses and duration of transcreation.
TransPerfect aimed to reduce the time and expenses usually involved in transcreation by utilizing LLMs. Instead of using a completely automated or all-human method, translations are generated using Amazon Nova Pro on Amazon Bedrock or Anthropic’s Claude, prompting the creation of several possible translations with minor modifications. Instead of creating the translation from start, the human linguist selects the best suited translation within the translation editor.
An LLM-powered transcreation in the GlobalLink Translate online editor is displayed in the screenshot below.
Users are experiencing up to 60% increases in linguist productivity when they use GlobalLink powered by Amazon Bedrock for transcreation.
In Conclusion
Customers from a variety of industries, including manufacturing, banking, and medical sciences, have reported up to 80% shorter project turnaround times and up to 40% cost savings in their translation workflows as a result of LLM-powered transcreation and post-editing. Furthermore, adding an automatic post-edit step to workflows that simply use machine translation significantly improves the output’s quality without requiring human intervention.
By prohibiting sharing with FM providers and keeping it out of model updates, Amazon Bedrock protects data. Responsible AI is crucial, even beyond data security. TransPerfect may customize honesty protections for post-editing using Amazon Bedrock Guardrails. It provides contextual grounding checks to detect and filter errors, which are essential for generating accurate translations, in order to combat AI hallucinations.
For your personal use case, test out Amazon Bedrock’s LLM-powered post-editing and transcreation features. Leave a comment with your thoughts and queries.