From: redpointai
DeepL, a company recently valued at $2 billion, supports over 100,000 businesses in AI translation worldwide [00:00:02]. Its mission is to solve language and translation problems, specifically for businesses, leveraging AI [00:05:02].
The Global Language Landscape
While English is often perceived as a dominant language, only 18% of the world speaks English [00:05:12]. Even in the United States, which is arguably the largest English-speaking region, 20% of people speak languages other than English at home [00:05:17]. This linguistic diversity highlights the critical need for effective translation solutions for businesses to communicate seamlessly across various languages [00:05:26], both for internal operations and global expansion [00:05:36].
DeepL’s Approach to Business Translation
DeepL focuses on high-value use cases where accuracy and quality are paramount [00:06:03]. The industries employing DeepL’s services are diverse, ranging from law firms, manufacturing, and retail to media and government [00:05:54].
Key use cases include:
- External Communication: A major media company based in Japan, which owns the Financial Times, uses DeepL to translate thousands of articles across Japanese, English, and Chinese to reach new readers globally [00:06:13]. This demonstrates the market potential across various industries for AI translation.
- Internal Operations: Multinational companies, particularly in highly regulated industries, rely on DeepL for internal communication. For instance, a German legal contract can be translated for a specialist in the US [00:06:48], or a Japanese car manufacturer uses it to ensure information flow from European and US customers influences R&D in Japan [00:07:08]. This highlights the impact of large language models on business operations on efficiency.
Specialized Models vs. General Models
DeepL’s success is attributed to its focus on specialized translation models rather than very general ones [00:36:34]. While large language models (LLMs) can perform a wide range of tasks, specialized models trained specifically for translation, particularly with good labeled reinforcement learning, often outperform general models [00:38:16].
- Tailored Solutions: DeepL builds a vertically integrated, “ready-made solution” that can be plugged in and used by employees, avoiding the need for internal AI engineers [00:37:21].
- Data-Driven Adaptation: The company runs different sets of models depending on languages and language pairs, given the varied data sizes available for different languages [00:12:45]. For example, there’s more translated material for German-English than Polish-English [00:12:59]. Models may be grouped based on language similarities to learn from each other [00:13:15].
- Human-in-the-Loop: The influence of human data has been continuously rising [00:14:48]. DeepL employs thousands of human translators globally for training models, performing quality assurance, and providing feedback [00:15:18]. This in-house data labeling ensures high, steady quality, especially for specialized models [00:15:31].
- In-House Infrastructure: DeepL has opted to build its own data centers since its inception, rather than solely relying on hyperscalers [00:27:04]. This allows for cost efficiency and access to the newest hardware for training and inference, which is crucial for staying at the forefront of innovation and for building and utilizing large language models at scale [00:29:10].
Evaluation of Translation Quality
Evaluating translation quality has evolved. While synthetic metrics like BLEU score were state-of-the-art initially [00:34:12], they quickly became insufficient as AI translation quality surpassed their capture capabilities [00:34:25]. The “real test” for DeepL remains human evaluation [00:34:55]. Thousands of translators judge translations on accuracy, nuance, and native feel, often in a comparative manner against other models [00:35:05].
The Future of Translation
AI translation is far from a solved problem, especially for less-resourced languages where data is scarce [00:30:46]. While one-to-one communication like emails is largely “solved” for well-resourced languages [00:31:35], complex tasks like publishing a marketing website or operating manuals for critical infrastructure still require human oversight and accountability [00:31:50].
Synchronous Speech Translation and Language Learning
A significant future frontier is synchronous speech translation and voice translation [00:39:11].
- Impact on Business: The technology could allow global teams to communicate in their native languages in real-time, regardless of location [00:40:48]. This would facilitate access to education, learning resources, and knowledge for employees worldwide [00:41:11].
- Challenges: Spoken language is more complicated than text because it’s stream-based, less structured, and people speak carelessly [00:45:42]. Key technical challenges include latency and handling ambiguity in speech [00:46:18].
- Learning Languages: While AI translation will bridge communication gaps, the personal and cultural value of learning languages will remain [00:41:30]. The average person might learn fewer languages in the future, but those who do will pursue it out of personal interest and for the intellectual challenge, similar to playing chess despite AI mastery [00:43:36].
- AI language learning platforms: This technology could democratize language learning by making it more accessible and affordable, compared to traditional in-person teachers [00:47:25]. Models in this space don’t need to be perfectly accurate as long as they facilitate speaking and dialogue [00:47:53].