NLU Training Data Generation — Utterance Augmentation at Scale

I will generate a high-quality NLU training dataset for your intent and entity system — using LLM-based paraphrase generation, back-translation augmentation, and template-based variation to produce diverse, realistic utterance sets that cover the natural language variation your users actually produce. NLU models trained on limited training data created by one or two people produce models that only recognise phrasings those specific people thought of — missing the genuine diversity of how different users express the same intent, which is what causes production NLU systems to misclassify utterances that seemed obvious to the people who built the system.

The data generation covers seed utterance review, LLM-based paraphrase generation with diversity controls, back-translation augmentation for additional variation, template-based systematic variation, deduplication and quality filtering, and delivery in your NLU framework's training format (Rasa YAML, JSON, or CSV).