Ines Montani - Keynote: Applied NLP in the age of Generative AI | PyData Amsterdam 2024

Modularity and having clear components is crucial when building NLP systems - break down complex workflows into smaller, testable pieces rather than relying on one large model

Focus on solving the actual business problem rather than getting caught up in the latest tech - the goal is to build practical solutions, not chase trendy approaches

Keep humans in the loop during development and validation - this helps catch errors, improve data quality, and ensure systems meet real needs

Data privacy and control are essential considerations - solutions should be designed to keep sensitive data secure and under organizational control

Transfer learning allows leveraging large models while maintaining efficiency - train smaller task-specific components on top of general language understanding

Proper evaluation and testing are critical - without metrics and testing, there’s no way to know if changes improve or break system performance

Start with clear schemas and taxonomies for structured outputs - define what information you want to extract before building complex systems

Consider using retrieval augmented generation (RAG) to combine knowledge bases with generation capabilities

Focus on data quality and iteration in development workflows - good training data and rapid feedback loops are key to successful systems

Stay pragmatic and be willing to use simpler solutions when appropriate - not every problem requires advanced AI approaches