Systems for real life

• basic (or advanced) workflows in practical NMT systems
• data, training and evaluation
• critical issues, typical errors (fluency vs adequacy, robustness, under or overtranslation, hallucination)
• practical solutions (named entities, placeholders, pre- and post-processing)
• high quality NMT system HOWTO or: how to win the WMT (http://statmt.org/wmt21/) news task

Methods to improve translation quality, system efficiency and service

• data centric and model centric approaches
• backtranslation, domain adaptation
• transfer learning, teacher student learning, knowledge distillation
• 'auxiliary' tasks (as machine learning problems)
• QA (Quality Estimation)
• APE (Automatic Post-Editing)
• future directions
• document level MT
• document based dynamic domain adaptation
• adaptive MT

Questions (with or without answers) - What are the most important current research problems in NMT? - How does SOTA research results carry over to (our) real life systems? - Do we need something special to deal with Hungarian? - Does linguistics have a place in practical MT systems?

Slides