With so many TMS and AI translation tools available, where do you start? What should companies think about when choosing tools?

Dave: Businesses need to consider not just the features and functionality of each tool but also how well they integrate with each other and your desired translation workflow.

For example, integrating AI into the translation process starts with choosing a machine translation engine and deciding whether or not to customize it. This choice directly impacts the other translation tools for consideration and what your translation process will look like. It’s all about fitting the pieces together to get the highest quality output and top productivity gains from the new technology. 

It’s important to understand that there are many tools in the toolbox, and MT is just one of them. For example, AI-powered technology can automate workflow aspects as well. 

Multiple AI translation models exist, including Neural Machine Translation (NMT) and the newer Large Language Models (LLMs) like ChatGPT or Bard. Could you clarify the difference?

Dave: NMT is finely tuned for translation tasks. It’s designed to prioritize accuracy, ensuring the translation is as close to the original text as possible.

ChatGPT and other large language models can also perform translation for many use cases capably, but they are built for a broader range of language tasks. It’s trained on a massive, diverse data set, including multiple languages. This training allows them to generate fluid and natural language.

We’re currently evaluating LLM outputs against traditional NMTs from leading providers. While NMT still performs better in translation quality, LLMs are not far behind, and we’re seeing promising results as we continue testing and training these models. 

However, when we compare LLMs to trained, customized NMTs, the latter comes out ahead. At least for now – who knows what will happen in the next six months! 

Do all businesses need to customize their MT engine? 

Dave: Customization is vital for quality. There are considerable advantages to using a customized MT engine trained on your company data. When we tailor an MT engine to a specific company’s data, we typically see up to a 25% increase in quality of output when compared with the untrained engine. 

However, there are specific data requirements to train these MT engines effectively, so doing this training is not the right choice for every client. For instance, we need at least 10,000 segments to see a substantial improvement. If a client only has around 5,000 segments, it may be too soon to train the engine for their needs.

How does choosing one MT tool over another impact a business’s other translation tools?

Dave: The Translation Management System (TMS) must be compatible if you’re using a customized engine. For example, if you’re using Google’s MT, the TMS should support its general and trained models. The leading MT providers—Google, Amazon, and Microsoft—constantly innovate. You’ll also want a TMS provider that keeps up with them to allow your business to make the most of these advancements.

With so many TMS, MT, and LLM options and the need to integrate them all for smooth and efficient translation processes, can you tell us more about what goes into integrating these systems?

Dave: It’s incredibly complex. We use several platforms with our clients. We’re technology agnostic, provide bespoke customer solutions, and can work with and are open to any system our customers use. These tools often include their own MT solutions and offer connectors to other MT engines that better fit the client’s requirements.

We also have our own suite of tools to check the quality of MT output for specific translation tasks. We use automated and human-led evaluations to benchmark metrics like BLEU and METEOR scores to see how well each NMT engine works for particular clients.

The goal—high-quality translations—is the same no matter what tools you use. However, different tools present information and track performance in slightly different ways. Each system, especially when integrated with others like NMT engines or LLMs, requires extensive user education and sometimes custom development to ensure everything functions cohesively.

The real work is in the details, so understanding each system’s capabilities and teaching our teams and clients how to utilize them effectively is a massive part of what we do. Integrating MT into workflows isn’t a quick ‚plug and play‘ setup; it demands a thorough understanding of how MT is presented to linguists and how it fits into the translation process.

Buying and implementing tools is an investment of time and resources. Can you share with us the benefits? 

Dave: There’s quick ROI regarding cost, speed, quality, and consistency when selecting the ideal tools for the job, setting up workflows, and training users. 

For example, many project management tasks that used to take hours can be automated through a TMS with just a few clicks. In many cases, AI translation tools can give linguists a rough draft to edit, saving even more time. Automation isn’t just about speed, either. Translation quality improves thanks to translation memories that leverage previously translated content and QA tools that check for various quality considerations. Automation is a necessity to scale globally, but there are a lot of moving parts that have to work together to make that happen. If your tools aren’t compatible or your workflows are inefficient, you won’t see as much of a benefit as expected.

You have to set realistic expectations around the complexity of these integrations and the necessary customization required to see ROI.

Is machine translation viable for every language combination? 

Dave: The quality potential of the languages in question must be understood. Each language pair has its own MT quality level. For instance, translating from English into Romance languages like French, Spanish, and Portuguese tends to have higher quality MT output. Germanic and Slavic languages vary more. The quality will be impacted if your source is two of the more difficult languages, like Finnish to Japanese. 

Despite this gap between languages, there has been significant progress in MT quality for languages and language families that have traditionally lagged. Take Korean, for example; it used to lag in quality, but thanks to the efforts of platform providers, we’ve seen a noticeable improvement in the past year. LLMs, on the other hand, may have better results with a broader range of languages because of larger data training sets, inherent flexibility, and contextual understanding. 

What should businesses know about protecting data privacy when using these tools? 

Dave: Public tools like Google Translate are not designed for data privacy. They pose a significant risk. Google uses all the data it receives. None of the content that goes into public domain engines like Google Translate or LLMs like ChatGPT remains proprietary or confidential. This is a significant issue for our enterprise clients who may be handling confidential company data in their translation process. Samsung experienced a data leak in April via ChatGTP and banned the use of generative AI tools in the aftermath.

Companies with data privacy concerns should choose private and secure tools so that none of the data is shared.

With AI assuming more translation tasks, everyone wants to know how linguists will continue to fit in.

Dave: At the end of the day, linguists want to do great work for their clients and produce high-quality localized content. The more tools we can give them to help them achieve this goal, the better. Over time, the job of a linguist may become more technical and specialized as they learn to use these new tools to their total capacity. 

As I’ve experienced in my career, you must continually shift and grow your knowledge. For linguists, this might mean transitioning from traditional translation to roles like prompt engineering, linguistic data management, or post-editing. It’s about moving where the technology leads and embracing continuous learning. Although the roles are changing, the need for skilled linguists isn’t disappearing—it’s just transforming. Their expertise remains critical for training AI, validating its output, and refining translations.

You work closely with tools vendors to stay current with technological trends and discuss needed features. What advancements are you seeing? 

Dave: One exciting new development is the potential to use LLMs as part of the translation quality assessment process. Some CAT tool companies are now integrating LLMs to estimate the quality of translations, which can significantly aid in the post-editing process. Also, we’re seeing continuous, significant quality improvement via NMT. When MT evolved from statistical to neural models, we saw a massive leap in output quality, and it keeps improving. Another trend is the rise of adaptive MT systems, which learn from the corrections linguists make, creating a more intelligent translation process over time.

Beyond translation, AI’s role is expanding into project management tasks. It automates the setup of translation projects and even handles complex tasks like file parsing, which is beneficial for challenging formats like subtitles. This automation streamlines the workflow, making it more efficient for project managers and linguists.

We’ve also noticed a trend towards graphical workflow interfaces that are more user-friendly and intuitive for workflow management. These interfaces allow for a visual representation of the workflow, making it easier for project managers to drag and drop steps and manage the process effectively.