The Stack Overflow Podcast “Translation is the tip of the iceberg”: A deep dive into specialty models
Feb 28, 2025
Olga Barogovaya, VP of AI at Smartling and expert in AI-powered translation solutions, discusses the evolution of natural language processing from rule-based systems to specialized machine learning models. She highlights the importance of fine-tuning in translation and the role of human translators. The conversation covers the advantages of specialized models in various sectors like healthcare and the complexities of handling personally identifiable information. Olga also addresses the challenges of translation workflows in business, emphasizing the need for human oversight.
AI Snips
Chapters
Transcript
Episode notes
LLM Adoption
- Smartling has been using LLMs since transformer models like BERT became available.
- LLMs allow for advancements in natural language processing.
Specialized vs. General LLMs
- Fine-tuning LLMs for specific tasks like translation yields better results than general-purpose models.
- General-purpose models hallucinate and struggle with cultural nuances in translation.
Data and Fine-Tuning
- Smartling leverages labeled data for quality, edit distance, and domain to fine-tune models.
- Translation is a small part of the process, which also includes smoothing, quality estimation, and terminology management.