Custom Training of AI for Language Services: Translation is Just the Beginning
summary
Custom models run as per-customer instances called Halos human-augmented language OS that ingest translation memories via automated pipelines containing hundreds of thousands to millions of sentences. Public generic MT hits about BLEU 40 → 50–60 after customer-data fine-tuning. Human reviewers stay in the loop and the no-edit rate for critical docs rises from ~20% to ~40%. Production work includes a basic translation app and staged capabilities for summarization, text generation and speech-to-text, all fed back to improve models.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 0:41The Answer to the Universe
- 1:16The Stone and Iron Age
- 1:42AI is our tool
- 2:09Halos
- 3:59Translation Memories
- 5:11Blue Score
- 6:55Translation is Just the Beginning
- 7:31Challenges
- 8:31Conclusion
- 9:16Questions
related talks
18mBuilding an AI service without training – AI translations with LLMs
28mAI’s Next Great Challenge: How to create AI that truly understands us?
40mHow Can Artificial Intelligence use Big Data for Translating Documents?
40mDeploying Your Own Language Models: A Use Case Demo with Open Source Tools
18mAI translation + AI agents = i18n made easy
57mMachine Translation of Human Languages in the Age of LLMs: Is This the End of the Language Barrier?
42mPost-Training Small Language Models: the adventures of a practitioner
47mAya: Expanding the Worlds AI serves