From: redpointai
The role of human translators is evolving in the era of Artificial Intelligence (AI), particularly with advancements in language models and translation technologies. While AI can automate many aspects of translation, human expertise remains crucial for quality, nuance, and specialized tasks.
The Evolving Role of Human Translators
DeepL, an AI translation company, integrates human data and translators extensively into its operations [00:14:48]. The influence of human data has been consistently rising and is expected to become even more important [00:14:51].
Key areas of human involvement include:
- Model Training and Quality Assurance DeepL has run large-scale data annotation projects internally for years, utilizing human translators to train models on desired translation styles [00:15:03]. They also play a vital role in quality assurance (QA) for specialized models, ensuring high and steady quality, which is crucial for business customers [00:15:30].
- Personalization and Interactivity While AI provides quick translations, users often seek to personalize and refine the output, working interactively with the AI to create a translation that is “mine” rather than an average internet translation [00:03:04].
- Embedding Terminology DeepL allows customers to embed specific terminology into their models, which is a critical feature for businesses that need to control the language used in technical manuals or internal communications [00:11:36]. This requires careful management to ensure grammatical rules are followed and to address words with multiple meanings [00:12:11].
- Handling Nuance and Context Human translators are essential for discerning nuances in language, especially when optimizing for accuracy in technical texts versus fluency in marketing content [00:04:26].
- Specialized Data Labeling For tasks requiring high-quality, specialized data, it’s beneficial to have a small set of the best possible quality and to individually select the annotators [00:17:45]. This ensures accuracy and consistency, as human performance can fluctuate [00:18:07].
The Future of Translators and Language Learning
The need for human translators, particularly native speakers from around the globe, will remain central to DeepL’s operations, despite the challenges in hiring [00:16:01].
Impact on Language Learning
The speaker believes that while AI can bridge communication gaps, the cultural aspect and personal connection derived from learning languages will retain significant value [00:41:30].
- Learning more or less languages? The speaker anticipates that the average person might learn less languages in the future due to advanced translation capabilities [00:43:34]. However, those who do learn languages will do so out of personal interest and enjoyment, similar to playing chess [00:43:40]. Languages are seen as an intellectual challenge and a core human interest [00:43:53].
- Accessibility to Language Education AI technology has the potential to democratize access to learning to speak a language, as it was previously a privileged and expensive endeavor requiring in-person teachers [00:47:25]. Models can enable fluent conversations for learning, where full accuracy isn’t as critical [00:47:46].
Overcoming “AI Replacing Humans” Narrative
DeepL has successfully built a strong brand within the translator community because professionals have recognized that AI can significantly enhance their productivity by handling more mundane tasks, allowing them to focus on adding value where it’s most needed [00:55:02]. The sheer volume of content created globally now demands AI for localization, as humanity would be unable to cope manually [00:55:25].
Accountability in High-Stakes Translations
For high-stakes applications like operating manuals for nuclear power plants, human accountability and responsibility for the final translation result are paramount, even if AI assists in the process [00:32:03]. This reflects a general concern with AI: while it unloads work, caution is needed regarding ultimate responsibility [00:32:24].
Evaluation of AI Models
In professional translation contexts, evaluation is simpler than in literature, where translation itself can be considered an art form [00:33:20]. While synthetic metrics like BLEU score were used in 2017, they quickly became insufficient to capture the high quality achieved by AI translation models [00:34:12]. The real test remains human evaluation, where thousands of translators judge translations for accuracy, nuance, and native feel, often in a comparative way against next-gen models or competitors [00:34:55]. This human feedback loop is crucial for building better, more tailored products [00:11:18].