6.4. Reflection Checkpoint

Key Takeaways

Before proceeding, ensure you can:

• Apply Text Analytics features (sentiment, NER, key phrases, PII) to appropriate scenarios
• Configure Azure Translator with both required headers (key AND region)
• Interpret BLEU scores for translation quality (40-59 = high quality)
• Implement STT (SpeechRecognizer) and TTS (SpeechSynthesizer) with error handling
• Use SSML elements (<prosody>, <break>, <say-as>, <mstts:express-as>) for speech control
• Address WER issues with appropriate solutions (substitution → custom vocabulary)
• Design CLU projects with intents, entities, and the critical "None" intent
• Choose between Custom Question Answering and RAG based on content update frequency

Connecting Forward

Phase 7 combines NLP capabilities with search and document extraction. Azure AI Search uses skillsets that call Language services you learned here (NER, key phrases) to enrich content during indexing. Document Intelligence extracts structured data that can then be indexed and searched.

Self-Check Questions

1. A speech recognition system for a medical application has high substitution errors for drug names and medical terminology. What specific training data approach would reduce these errors?

2. A banking chatbot needs to understand queries like "transfer $500 from savings to checking" and extract the amount, source account, and destination account. Which custom language model should be used, and what entity types would you define?