Interview Questions for Olga Beregovaya, Vice President of Artificial Intelligence at Smartling:
Refreshed Article:
Embrace a tour behind the scenes of Smartling, an AI-driven powerhouse that revolutionizes content delivery, all the while ensuring top-notch translations, even for the most under-resourced languages. In this engaging chat, we delved into the enigmatic world of Smartling with Olga Beregovaya, their VP of AI.
Makaryan: Alright, let's dive right in! Can you give us a lowdown on Smartling and what it brings to the table?
Beregovaya: Smartling is more than just a translation service. We're a dynamic, AI-powered solution that seamlessly catapults businesses into global markets with ease. With a knack for Romance and Nordic languages, we deliver translations nearly indistinguishable from those crafted by human hands. But we don't stop there; we help businesses adapt their content to diverse audiences, offering multimedia localization, workflow automation, and project management services. Essentially, we're the total package for end-to-end global content delivery—and translation is just the start!
To maintain the highest translation standards, we embrace a human-in-the-loop approach. Our AI models, including neural machine translation, are not always infallible; they can err or create "hallucinations." Thus, our tools empower human translators to refine the AI outputs, ensuring exemplary precision.
As VP of AI at Smartling, my role is multifaceted. I spearhead research and development efforts, guide a cutting-edge deployment team, and shape the strategic and product vision for AI advancements. Initially, I was the VP of Machine Translation and AI, but as we realized that machine translation is merely a subset of AI, we merged the two into a comprehensive AI strategy. My day-to-day revolves around guaranteeing Smartling's AI offerings are cutting-edge, scalable, and cater to our clients' unique needs.
Makaryan: So, what's been the biggest hurdle Smartling has faced with the rise of Generative AI?
Beregovaya: The past couple of years have seen a wave of AI experimentation, with some projects soaring, while others have faltered unexpectedly. According to Gartner, 75% of AI projects flop, but we've managed to stay in the game's winning 25%.
One colossal challenge has been minimizing latency, which is the time it takes for AI-generated translations to be delivered. Large language models (LLMs) consume substantial computational resources, which can slow down response times—a big no-no for real-time content delivery. To expedite responses, we've augmented our systems' efficiency, processing multiple tasks simultaneously, employing diverse AI models in unison, and enhancing the way we transmit and deliver translations.
Privacy and security have also been significant obstacles as companies become increasingly watchful over their data's safety and handling. Many of our clients' legal departments have levied stringent data governance requirements, pushing us to toughen up our SLAs, data privacy policies, and security frameworks.
Choosing the right AI model for each use case is another challenge we face. Early on, many in the industry leaned on a single LLM, but we've opted for a provider-neutral stance. Instead of relying on just one provider, we meticulously evaluate and combine the optimal available models, such as GPT-4, Google Gemini, and Claude, to fit the specific tasks at hand.
Hallucinations remain an ongoing issue. AI models can churn out fluent yet factually incorrect or utterly irrelevant translations. For example, a seemingly simple translation task might result in a lavish Italian dating profile—completely off-topic from the original text. This highlights the importance of continuous monitoring and adjustment in our AI models.
Makaryan: How does the Smartling platform cater to the specific demands of its clients?
Beregovaya: Our clientele is diverse, spanning sectors like tech, streaming, entertainment, e-commerce, and law. Our platform tailors its AI tools and workflow automation to each industry's distinct localization challenges.
Technology companies, for instance, require the efficient translation of UI text for apps and websites, often containing incomplete, inconsistent, or grammatically inaccurate phrases. E-commerce sites typically grapple with translating user-generated content, which can be messy, hyper-informal, or loaded with abbreviations. In the legal industry, we manage the intricate translation of patents, which require exact terminology and subject matter diversity, spanning from consumer products to ground-breaking technologies.
Our platform also streamlines the quality assurance process. Many clients devote substantial resources to internal language evaluation, and our AI-driven tools make this process more efficient by automatically identifying essential issues, enabling human reviewers to focus on essential details. This synergy of automation and AI-led quality assurance helps clients trim costs and boost efficiency in their content delivery processes.
Makaryan: What does the future hold for Smartling as AI trends continue to evolve?
Beregovaya: Historically, Smartling has functioned as a SaaS platform, with customers actively managing their translation workflows, combining AI tools and human oversight. However, AI is reshaping the localization landscape, and we're transitioning towards "service-as-software." AI will handle more of the translation process autonomously, requiring less manual heavy-lifting from our clients.
To realize this vision, we're investing in fully automated global content delivery, leveraging machine translation with automated post-editing and smaller, fine-tuned AI models specialized for translation tasks. General-purpose models like GPT can translate, but specialized models reduce errors and improve accuracy.
Another key trend is multi-modality, where AI systems process and generate content in multiple formats, including text, audio, and visuals. This broadens the horizon beyond translation, paving the way for content creation and adaptation.
Makaryan: Expanding into under-resourced languages poses challenges concerning data acquisition. How does Smartling struggle with this?
Beregovaya: The scarcity of available data for training machine learning models is a significant hurdle when venturing into new languages with minimal digital resources. while languages like English boast vast training datasets, others like Swahili are data-sparse, making it tough to create high-quality translation models.
To bridge this gap, Smartling employs several strategies, such as data scraping, collecting local audio recordings, and generating synthetic data to expand smaller datasets.
We also leverage data from languages within the same language family. Using acoustic models and shared linguistic patterns, we improve translation accuracy for these under-resourced languages. We collaborate intimately with clients, too, developing training datasets through human feedback. For instance, we partnered with the African Language Lab to bolster data for African languages, bolstering resources for long-tail languages.
The good news is that progress is afoot. Initiatives like Facebook's "No Language Left Behind" project are helping to provide more variety and relevance in training data for under-resourced languages. At Smartling, we persist in fine-tuning our models and refining our training processes while recognizing the importance of a diverse data science crew. Different cultural viewpoints offer valuable insights when crafting language models. Our former Chief Data Scientist, who hailed from China, underscored this point by contributing to the selection and fine-tuning of datasets for diverse languages.
Over time, Smartling is seeing substantial improvements in translation quality even for languages once shrouded in obscurity.
- Smartling's role extends beyond translation, encompassing multimedia localization, workflow automation, and project management, delivering a total package for end-to-end global content delivery.
- Maintaining high translation standards is critical for Smartling, which employs a human-in-the-loop approach, refining AI outputs with human translators to achieve exemplary precision.
- As VP of AI at Smartling, Olga Beregovaya spearheads research, development, and strategic planning, guiding a team that deploys cutting-edge AI solutions, while continually improving their scalability and client-centricity.
- Privacy and security have been significant challenges faced by Smartling, requiring stringent data governance, SLAs, and robust security frameworks to meet client expectations in a data-conscious world.
- In the quest for optimal AI models, Smartling adopts a provider-neutral stance, evaluating and combining best-fit models such as GPT-4, Google Gemini, and Claude for specific tasks, minimizing hallucinations and enhancing accuracy.
- The future of Smartling revolves around automated global content delivery, with a focus on fine-tuned AI models for translation tasks, multi-modality, and improved adaptation for various formats (including text, audio, and visuals).
