The Week in Review: March, 31, 2023
We’ve got stories on language access and education, along with press releases from XL8, GLOBO, and more in this week’s recap.
→ Continue ReadingT
hanks to marked improvements in large language models (LLMs) over the past few years, generative artificial intelligence (GenAI) has gone mainstream. People from all over the world and all walks of life are experimenting with GenAI to increase their productivity at work or enhance their personal lives. Businesses, too, are wondering whether and how to incorporate GenAI into their products and workflows.
The language industry might have a leg up over other sectors, as language service providers (LSPs) and language technology companies already employ other forms of AI, such as neural machine translation (NMT). Still, with GenAI in its nascent stage, language companies are certainly in uncharted territory.
To find out how the industry is utilizing this emerging technology, MultiLingual reached out to some of the most prominent and influential companies in the space. Ten responded: Acolad, Argos Multilingual, Bureau Works, Gateway Translations, Keylingo, LILT, Lionbridge, Pangeanic, Phrase, and Translated.
The responses showed that one of the most common uses of GenAI is improving internal processes, with multiple companies providing employee training on the topic. Some companies have gone further and are already integrating GenAI into their products and services through tailored LLMs. In their own words, here is how some of the top companies are leveraging GenAI for their teams and customers.
Advertisement
What role does GenAI play for your organization? How does your team specifically use it to fulfill your services?
Acolad (Pavel Soukenik, Head of Global Solutions): GenAI isn’t just the future for Acolad; it’s our present and has been a significant part of our past. Our AI journey began in 2016, and since then, our dedicated Research & Development and Technology & Solutions teams have leveraged our extensive experience in machine learning (ML) to harness GenAI’s transformative potential for both customers and internal teams. Integration and automation are key in our approach.
Internally, we utilize AI as a powerful productivity tool, encouraging our teams to explore its vast possibilities. AI is used to optimize workflows through our automation platform, which integrates traditional technologies with ML and GenAI modules. Our proprietary prompt platform offers a secure, General Data Protection Regulation (GDPR)-compliant, authentication-based environment. This enables our teams to work more efficiently while ensuring the secure handling of confidential data.
For our customers, we seamlessly integrate AI into tailored, innovative solutions that empower them with personalized and efficient communication. Our service portfolio leverages AI across all stages of the content life cycle to enhance cost efficiency, reduce turnaround times, and improve the quality of deliverables and existing translation assets. Acolad’s AI-driven services include automated post-editing (PE), product descriptions, content rating, mono- and bilingual content quality management, engine evaluation in MT Hub, voiceovers, and solutions for regulated industries. Furthermore, we partner with customers to develop or refine their own models, utilizing our in-house AI expertise to achieve exceptional outputs.
Argos Multilingual (Libor Safar, Vice President of Growth): GenAI plays the role of a brilliant, promising intern who is already doing a great job — one who everybody loves, warts and all. An intern who just needs some more polishing, like a diamond, and some more time to realize her full potential.
Our dedicated AI data services team helps enterprises build smart, multilingual LLMs. This means pre-training models and fine-tuning or testing them. Specifically, our linguists and domain experts help ensure diverse perspectives and linguistic nuances are captured in the models. The goal is to enrich the quality and authenticity of prompts and responses. This includes multilingual annotation using customized multidimensional quality metrics (MQM) to assess both source and target texts.
We’ve also embedded GenAI into our hybrid AI translation workflow, which essentially uses AI for translation and quality control (QC), with humans in full command of quality verification and editing. In our eyes, the balance is all about risk tolerance. Some enterprises may prioritize the speed of AI with minimal human intervention, while others prefer a more hands-on human approach for sensitive content. We don’t see only one way here. So the AI workflow can be designed to experience the best of both worlds: rapid, AI-enhanced translations with the option to increase human involvement at any point. The benefit is full control over quality and risk.
Bureau Works (Gabriel Fairman, Chief Executive Officer): GenAI allows our users to go from mind-numbing PE to thoughtful teaching. In the process, they translate 22% more effectively than the best NMT tools available. Our framework allows users to teach their own private engines without any setup other than determining their linguistic corpus. These personalized engines help users translate just like themselves, but faster and with fewer errors.
Gateway Translations (Markus Seebauer, Managing Director): Many people think that GenAI tools like ChatGPT are experimental and inferior to proven MT engines like DeepL. In a decade of working on translations for a large US tech company using a custom language model, we have run experiments to improve AI output. We have achieved quick wins in increased and consistent translation quality by creating customer-specific editing instructions and project management (PM) guidelines that extend beyond QC. When we used AI to assist in translating GitHub’s technical documentation, AI enabled faster time-to-market, provided multiple options for linguists, and gave client-side localization managers better visibility into editing ratios to negotiate future cost savings.
Keylingo (Virginia Careto, Vice President of Operations): GenAI plays an important role within Keylingo, mainly by enhancing our marketing and strategy efforts. It remains a tool on the bleeding edge of technology, and therefore it’s not yet a tool that we use for services we deliver to end clients. However, we consider it a valuable tool for optimizing our internal processes, particularly in the realms of ideation, strategy development, and copywriting. It is similar to having an additional teammate that we can count on for valuable contributions that we often implement following a complete human review.
In our marketing efforts, GenAI assists us every day by brainstorming fresh ideas for campaigns, social media posts, internal initiatives, and blog articles. This ensures our content stays engaging, relevant, and consistent with our identity. It aids in creating compelling headlines, persuasive ad copy, and search engine optimization (SEO)-friendly content, thus maintaining a consistent brand voice and enhancing efficiency.
In our strategy development, GenAI supports us by simulating various market scenarios for better anticipation of challenges and opportunities, which leads us to more robust strategic plans. It allows us to identify trends and patterns, enhancing our decision-making with comprehensive and accurate information. In addition, AI generates diverse solutions and creative approaches to business challenges, aiding in innovation and problem-solving.
LILT (Spence Green, Chief Executive Officer): As LILT is an AI-powered enterprise software platform for content translation and generation, GenAI plays a foundational role in our business model; it serves as the core of our platform and touches every word translated by LILT. GenAI provides prompts (translation suggestions) to human linguists, who verify accuracy word by word before confirming and finalizing the translation.
The LILT platform creates a unique, customized LLM per language (or domain) per customer; these train in real time on the customer’s preferences, tone, and style. This customer-specific LLM then powers all translation for that customer in that language or domain, whether to assist human translation or instantly machine-translate content. Each customer’s bespoke model powers translation across content types such as corporate websites, product manuals, social media, customer support chat, e-learning content, legal documentation, and product user interfaces (UIs).
Each customer’s fine-tuned models can then be paired with LILT Create, a generative tool that uses each customer’s data as prompts to generate high-quality, brand-aligned content in target language for content creators. This tool saves content creators 90% of the upfront content creation time, allowing them to jump ahead to refining and finalizing content to business preference, then publishing in global markets.
LILT also interoperates with third-party LLMs, ensuring top-level performance through the fine-tuning of custom models in all languages. LLM integrations include Google Translate, Amazon Translate, GPT, DeepL, and Amazon Bedrock.
Lionbridge (Will Rowlands-Rees, Chief Product Officer): Lionbridge has always been a leader in language innovation, and GenAI is no exception. Our goal is to work with this innovative technology to enhance our global content workflow for greater speed and personalization using LLMs. This integration has resulted in shorter project turnaround times and unlocked a new level of suitability and customer satisfaction.
To address the evolving needs of our customers, we utilize GenAI strategically in traditional localization. This technology enables us to offer customers more control over their content with greater speed and accuracy. We use it upstream to create high-quality content at the source and streamline the entire content process, from creation to review and delivery. This allows us to develop new products and services that expedite the traditional review process, where speed and cost are crucial.
We provide training on GenAI tools for all employees, and in 2024, all employees set a GenAI goal. They are not just experimenting with the technology occasionally; employees use GenAI to assist with their jobs regularly. We leverage its power to supercharge prospecting and sales, streamline onboarding, automate repetitive tasks in operations, and uncover financial billing trends. With AI as a cornerstone of daily activities, we have built a deep familiarity with the technology that benefits our team and customers alike.
Pangeanic (José Miguel Herrera Maldonado, Head of Machine Learning): GenAI plays a crucial role for Pangeanic, moving some of our solutions like MT to the next level. While the stated goal of translators, localization departments, and LSPs is to help humans understand across cultures and languages, the language industry has been working at the segment level for the past 30 years — it’s nonsensical. Humans don’t communicate at a segment level.
The obvious use case for GenAI is generating reports, project documentation, news, headlines, and summaries. But in the MT field, we utilize GenAI to improve the quality and fluency of translations at a chapter, page, and paragraph level. It’s like having a human reviewer or post editor. GenAI helps identify linguistic and contextual patterns thanks to its large context windows, resulting in more accurate and natural translations. It’s not just ChatGPT; there are other models you can run to better control output and privacy. Combined with good custom NMT, retrieval augmented generation (RAG) systems, and terminology management, you can obtain amazing results if you use it as a post-editor.
Phrase (Simone Bohnenberger-Rich, Chief Product Officer): GenAI is integral to Phrase, driving significant advancements in our localization processes. It augments and automates complex tasks that were previously challenging for AI technologies. In the Phrase tech stack, we deploy GenAI at critical points to hyper-automate manual processes, maximizing cost-effectiveness. For instance, MT quality is enhanced, dramatically reducing the need for human editing by using LLMs to produce high-quality, audience-specific translations. It also automates costly quality assurance (QA) steps like language quality assessment, which some organizations skip due to cost. This allows us to ensure quality while saving up to 90% of expenses. Additionally, GenAI helps adjust the tone of voice and other stylistic elements to align with the brand’s voice. This ensures that content creation is more adaptive and responsive. By integrating GenAI in these ways, we leverage technology to enhance service quality and operational efficiency, fostering scalable and effective innovation in localization.
Translated (John Tinsley, Vice President of AI Solutions): AI — be it generative, predictive, or some other form — plays an integral role at Translated, and has done so for the past 20 years. Our MT solution, ModernMT, was the first commercial application of the transformer technology back in 2017, with the advent of NMT. Transformers have since been popularized more broadly in today’s GenAI landscape as the technology that underpins LLMs (note the T in GPT stands for transformer).
As we move on from the first wave of foundational models to more application-specific implementations of GenAI, MT itself is evolving to become more generative in nature. This is massively expanding the scope of its capabilities, to use more and more context and to take precise instructions in the same way a translator might use a style guide. This makes it extremely powerful and brings us ever closer to the language singularly.
Considering that MT is AI, it is therefore pervasive across Translated. All human translations use suggestions from AI to varying degrees, while thousands of customers use it directly as a product itself via ModernMT. In these workflows, we also use AI-powered quality estimation to help determine where we should focus human efforts, allowing us to be more efficient in delivering services.
AI also underpins our audio-visual services, which has led to it becoming the fastest-growing area of the business over the past two years. This includes our automated speech recognition, speech synthesis, and increasingly popular voice cloning technologies. There are already many examples of this being applied effectively and at scale across our enterprise client base.
We make use of AI in less obvious ways across the business too, such as helping our Community and PM teams select the most appropriate human resources to work on given projects from clients. Again, these are not necessarily new developments, though their application and effectiveness are expanding!
What role do humans play in these processes?
Acolad: At Acolad, we advocate an expert-in-the-loop approach — emphasizing transparency and responsibility. While our specialized development teams focus on creating and deploying innovative solutions, our AI Ambassadors Program promotes AI literacy and adoption among clients and employees. This reflects our “no person left behind” policy aimed at empowering everyone to utilize the technology and positioning AI as a productivity enhancer, unlocking opportunities for growth and innovation. We are also committed to upskilling our employees for new AI-driven roles, complemented by hiring engineers, product managers, and solutions architects with AI backgrounds who bring additional expertise and fresh perspectives.
Argos Multilingual: Enterprises are transitioning to an AI-generated content-creation process. The advantages are obvious, but so are the risks. This is where humans are indispensable for verifying and editing a wide range of content types. This includes plagiarism checks, fact verification, and regular feedback on the prompts used.
Bureau Works: Humans are at the center of all of this. The engineers who iterate on this concept are viscerally connected with the translator’s experience, and the translators are connected to the tech the engineers build. We see GenAI as a portal that allows users to exchange information dialogically and fluidly with their engine. The semantic capabilities of the language models are much more fluid and responsive when compared to rigid, syntax-oriented legacy approaches. Humans are not just in the loop, they are in complete control of the process. They can choose what GenAI features they enable or disable, what they follow or ignore, and more. The engine is there to unlock higher creative and critical potential, rather than just allowing them to move faster. It’s a process of continuously fine-tuning the tech to minimize edit distances while maximizing the authorship experience for the translator.
Gateway Translations: Human expertise remains crucial for optimizing prompts, addressing source errors, and adding cultural context. Effective prompts allow UI designers to flag non-inclusive language for future review. We oversee the revision process and review the efficiency of the implemented AI systems, such as specific LLMs, by comparing raw outputs with revised final texts. A combination of technology and native speakers who are subject matter experts has proven most effective.
Keylingo: Human expertise remains crucial in our AI-enhanced marketing and strategic processes, and our team plays an essential role in several areas. Experts oversee and refine AI-generated content to ensure it aligns with our brand values and strategic goals, maintaining quality and relevance. Strategists interpret AI-generated data and insights to create actionable plans, effectively integrating them into our broader strategic framework. Our tech team uses AI-powered tools at various stages of the innovation cycle, particularly when understanding the problem and creating prototypes or minimum viable products (MVPs). This approach is in line with our agile mindset and enables our experts to focus on tasks such as decision-making and validation.
LILT: Humans play a foundational role in LILT’s business model. Human linguists in the LILT platform serve the role of verifiers, confirming accurate translations word by word. As the human interacts with the model prompt, the platform captures that interaction as training data. That data is then used to fine-tune that customer’s model for their preferences, tone, terms, and style.
After the model is fine-tuned for customer preferences, demand for linguists is dependent on volume of source content and of new differentiated content (for example, customers frequently send new content types, conduct brand refreshes, and introduce new terminology). LILT measures the impact of model fine-tuning via Word Prediction Accuracy (WPA), a score that measures the ability of the fine-tuned model to translate source content as a human linguist would. Our customers have achieved 90%+ WPA scores on their custom models, meaning that human linguists translate or correct less than 10% of the words presented to them. However, this fine-tuning would not be possible without human linguists verifying accuracy of translation and creating training data to be used for model fine-tuning. We refer to this as having a human “in the driver’s seat,” wherein the human guides and supervises the model and ultimately decides the best output for the content use case at hand.
LILT Create pairs third-party models for content generation with bespoke LILT models and curated datasets that customers have fine-tuned in LILT. The output is generation of highly customized content that is aligned to a company’s voice, tone, and terms. However, when accuracy and quality expectations are high, a human creator should still review, tweak, and finalize content before it is published.
Lionbridge: Human expertise drives the translation industry, and this holds true even as the field evolves technically. Customers depend on our deep understanding of localization and cultural nuances to ensure accurate and appropriate communication, especially when navigating the complexities of highly sensitive or critical content. Human experts can ensure the translated message conveys the right tone and avoids misinterpretation.
We must also remember that the translation industry involves so much more than just converting words and that customers value the human touch. Project managers handle logistics, language specialists offer industry-specific knowledge, and engineers ensure technology runs smoothly. These roles work together to provide linguistic and industry knowledge, develop and manage technology, and guide workflows to deliver a successful translation project and ensure customer satisfaction.
Pangeanic: For our marketing team, GenAI is a must-have, an indispensable tool to craft headlines and summaries. It eliminates “the fear of the blank page,” not to mention the ability to create images and even basic code!
However, creating good marketing material, code for specific applications, and insightful articles is beyond the technology’s reach. GenAI is good at what we call “shallow” human tasks. We’ve seen so many shallow newsletters on LinkedIn and in the media. The importance of humans in these processes cannot be understated. GenAI cuts production time, whether in automated translations, PE, or content creation. Depending on the task, the “human-in-the-loop” approach is moving to a “human-in-the-end” one, which allows us to assess the quality of generation and ensure it meets standards of accuracy, terminology coherence, and style.
Phrase: Despite the powerful capabilities of GenAI, human oversight remains crucial to ensure it delivers a real return on investment (ROI). Humans are essential for managing the overall process and defining the business logic and limitations — such as speed, volume, and quality — within which GenAI operates. We are implementing GenAI capabilities with very clear and transparent QCs, like the Phrase Quality Performance Score, to ensure these standards are met. Additionally, humans set the overall parameters for GenAI’s operation and govern or police GenAI models to ensure the outputs are accurate. This blend of advanced technology and human oversight is vital for maximizing the effectiveness and reliability of our processes.
Translated: Our motto at Translated is “We believe in humans,” so suffice to say humans are a cornerstone in almost everything we do. Humans create, transform, revise, assess, and approve content. Sometimes they do this from scratch, but increasingly frequently, they are doing this complemented by AI. This is a symbiotic relationship because the more people work with AI, the better it gets at helping people do the work, be it translation, revision, QA, subtitling, or dubbing.
That’s not necessarily to say it’s taking away from work that humans would have done otherwise — far from it. Enterprises are translating more than ever! In the past, the decision of what to translate or localize was based on what was possible in terms of time and budget. Now that those constraints are removed, they’re going all in. Yes, some of the work is fully automated, but this is allowing the humans — the subject matter experts — to bring their skillset to places where it has the biggest impact and is most critical.
Certainly, the role of humans in the content creation and localization process is evolving and changing as technology evolves. But this is a trend we’ve seen many times since the industrial revolution! The focus of human experts will be less on a broad set of tasks, including those mundane repetitive ones, but rather on a smaller set of higher-value activities.
Advertisement
Anything else you’d like to mention?
Acolad: Our expertise in GenAI has culminated in the development of our own AI platform, which is currently being piloted with customers and will soon be launched market-wide. This is an exciting milestone for us, and we are thrilled to continue advancing the state of the art in AI technology.
Argos Multilingual: The bottom line is that, regardless of the process, humans remain the ultimate arbiters of what quality looks like, what is right, and what is wrong.
Bureau Works: Tech is seen as an opportunity for some, a threat to others. Particularly for those who are being impacted, the change is not welcome. Unwanted change feels like an enemy that must be destroyed. This creates a conflict-based relationship between disruptor and disrupted that negatively impacts the possibility of mutually successful outcomes. We need common sense, thoughtful opinions, and open dialogue to maximize all of our odds of being better off in the future.
Gateway Translations: Although AI is lacking in specific domain expertise and cultural knowledge, it still significantly aids our work. The latest LLMs, such as ChatGPT-4o, have shown notable improvements in speed and text quality. Localization specialists are still crucial to making sure that engineers understand the limits of application programming interface (API)-driven unedited AI translation, which is tempting for them, and to getting continuously educated on internationalization best practices. Until we reach the next frontier in automated PE, we continue running experiments in cross-functional teams involving end clients and our global teams, with experienced linguists remaining a critical success factor.
Keylingo: While GenAI is available to basically everyone, it’s clear to Keylingo that there’s not a strong product market fit for our ideal customer. That said, Keylingo is leveraging GenAI for multiple internal processes, and therefore, learning and improving our ability to prompt this important tool. As GenAI matures, we will most certainly begin to test its implementation on customer deliverables following a detailed and comprehensive human review process. Finally, what is clear to Keylingo is that GenAI will dramatically impact the language services industry, as it will countless others. Our role at this stage, we believe, is to learn and develop expertise as much as we can so that at the proper time, we will be ready to implement this tool at a more comprehensive level.
While at Keylingo we strongly believe that GenAI is a powerful tool to optimize our processes, the human element remains indispensable for the success of our business. Our team’s expertise, creativity, and oversight ensure that we leverage GenAI effectively while maintaining the quality and integrity of our work and trademark to ensure an intuitive, real, and customer-centered experience.
LILT: As GenAI technology evolves, the question will become not if the human should be involved, but rather when and where. In the foreseeable future, humans will still be necessary to verify GenAI outputs for accuracy and purpose fit. Companies that strategically apply GenAI for their workflows will distribute repetitive, mundane tasks to AI whenever and wherever possible while ensuring that humans drive tasks and functions where human reasoning and verification produces the best output. The customers that we see taking advantage of this are thinking strategically about end use cases by content type, determining the required level of accuracy, and using that to inform the production workflow for that content type (AI only, human only, or AI plus human review).
Lionbridge: The question is no longer who controls AI but how humans leverage AI to work together more effectively. At Lionbridge, we believe this starts with empowering our employees to understand AI and harness its capabilities. That is why we have embarked on a GenAI journey internally, by developing custom curated training programs, building a centralized knowledge and innovation hub known as the AI Lab, and creating bespoke technical infrastructure. Lionbridge’s AI sandbox, a secure AI environment for employees to experiment with, has over 1,000 users daily. Around 4,200 people have completed our GenAI trainings so far, or about 70% of employees.
Building this culture of AI literacy and excellence across our global organization has ignited a true paradigm shift in the ways Lionbridge operates. By encouraging every team member to experiment daily with GenAI, we have produced wins, case studies, and hands-on use cases that have scaled across the enterprise and to customers.
Pangeanic: Let’s not forget the “Gen” part in GenAI. These systems are decoder-only and have been trained to generate. They are not trained on parallel corpora as is the case with NMT (which is very good if you have TMX files). They’re not thought to be production engines, but to work in context. We are seeing amazing gains with hybrid systems where context is provided and they act as a reviewer or post-editor. The human touch is indispensable for ensuring quality and consistency in the final output. At Pangeanic, we value collaboration between humans and machines to deliver high-quality linguistic services and meet the specific needs of our clients.
Phrase: While GenAI introduces remarkable advancements, it’s crucial to recognize the extensive infrastructure required for optimal performance. GenAI isn’t a standalone solution; achieving the highest quality in automated processes like translation necessitates a comprehensive infrastructure. This includes having access to additional data sources for fine-tuning and maximizing quality, as well as a robust LLMOps framework to detect and manage changes, ensuring trustworthiness.
Moreover, there is a need to differentiate between the hype and actual ROI of GenAI. While GenAI can significantly enhance certain processes, it is not a panacea for all challenges in localization. Recognizing its limitations and setting realistic expectations is essential for leveraging its full potential without overpromising results.
With the exponential growth in content produced by both GenAI and user contributions, there is an inevitable and natural shift toward technological solutions to manage what is beyond human capacity. By strategically integrating GenAI into the Phrase tech stack and complementing it with other AI tools, we position ourselves as leaders in reliable and scalable innovation within the localization industry. This approach is further enhanced by maintaining a balanced human-AI collaboration, enabling us to efficiently and effectively deliver high-quality content at scale that is culturally nuanced.
Translated: It’s a very exciting time, but we’re also at the peak of inflated expectations. There’s a lot of hype, but in practice, at least in our industry, there is not a huge amount of practical applications yet. The main reason is that this first generation of foundational models is not designed to be task-specific. Thus, their interfaces don’t offer a great user experience — hence the need for workarounds like prompt engineering. There are also challenges around performance and reliability at enterprise scale, not to mention cost.
These hurdles have held things up, but slowly they will fall down as task-specific models are released by various players. These models will inherently be smaller and more manageable. We will then start to see more practical applications at scale that are more cost-effective, and that’s when the fun will really begin!
EDDIE ARRIETA is CEO of MultiLingual Media.
CATHY MARTIN is managing editor at MultiLingual Media.
Advertisement
Related Articles
We’ve got stories on language access and education, along with press releases from XL8, GLOBO, and more in this week’s recap.
→ Continue ReadingOver the next five years, Taiwan plans to put nearly $1 billion toward English education in the country. It’s a pretty hefty sum of cash,…
→ Continue ReadingTranslated, the international company that pioneered AI-powered professional translation, is excited to announce that audiovisual localization pioneer Fabio Minazzi has entered the company as Director…
→ Continue Reading