Machine Translation of Human Languages in the Age of LLMs: Is This the End of the Language Barrier?
summary
Large LLMs now drive machine translation research and production, yielding near-human scores for many head languages while still failing on the long tail; Google reports coverage around 200–300 languages and research shows automated training data contains about 1.5% bilingual signal (~37M sentences) with roughly 14 million extractable parallel pairs. Benchmarks (WMT 2023–25) place LLMs top but not solved, low-resource outputs show catastrophic errors (colors, animals), and multimodal anchors (images, video, cross‑species corpora) are proposed to bootstrap translation where parallel text is absent.
AI-written summary · about
chapters (12)
- 0:00Historical context of MT
- 1:41Evolution of MT paradigms
- 7:30Markus Freitag's background
- 12:10State of machine translation
- 13:37Conference and evaluation
- 21:28Unsupervised MT and LLMs
- 23:58Incidental bilingualism
- 27:25Multimodal signal creation
- 28:42The role of evaluation
- 31:46Scaling to more languages
- 33:47Summary and future outlook
- 36:01Q&A session
related talks
16mBest Practices for Open Multilingual LLM Evaluation
49mA Practical Approach to Using LLMs: Peculiarities and Pitfalls
48mBridging the data gap between LLMs and children
45mCut Translation Costs by 50× for Millions of Product Sheets…
30mWhere there's a will, there's a way: The mad genius of LLMs as a classifiers
57mArtificial Languages are Dead. Long Live Artificial Languages!
44mKeynote: The Art of Language: Mastering Multilingual Challenges in LLMs
37mExpanding the Capabilities of LLMs Through Community Collaboration