Domain 5: Implement natural language processing solutions (15-20%)

← Domain 4 · Domain 6 →

---

This domain covers analyzing text and speech using the Azure AI Language and Azure AI Speech services. At 15–20% of the exam, it ties with Domains 1 and 6 as the highest-weighted areas — prioritize it.

5.1 Language Service (Analyze Text)

Pre-built Text Analytics Features

Feature	What It Does	Exam Trigger Phrase
Sentiment Analysis	Returns positive/negative/neutral at document + sentence level	"customer feedback analysis", "sentiment per sentence"
Key Phrase Extraction	Identifies the main topics in a document	"summarize key topics"
NER	Extracts people, organizations, locations, dates	"extract entities from text"
Entity Linking	Disambiguates named entities using Wikipedia (e.g., "Mercury" → planet or element)	"resolve entity ambiguity"
PII Detection	Detects and can redact personal data (emails, SSN, phone numbers)	"remove sensitive data before storage"
Language Detection	Identifies the language of text	"detect language automatically"

Custom Language Models

Model	Use Case	Training Data Needed
Custom NER	Extract domain-specific entities (legal terms, medical codes)	Labeled examples with entity spans
Custom Text Classification	Categorize documents into your own labels	Labeled documents per category

NER vs Entity Linking

NER → identifies entities and their type (Person, Location, etc.) Entity Linking → disambiguates entities by connecting them to a known knowledge base entry. Both can be used together.

---

5.2 Conversational Language Understanding (CLU)

CLU replaces the older LUIS service. It turns spoken or typed utterances into structured intents + entities.

Core Concepts

Concept	Definition	Example
Utterance	What the user says	"Set an alarm for 7am tomorrow"
Intent	The user's goal	SetAlarm
Entity	A parameter extracted from the utterance	Time = 7am, Date = tomorrow
Confidence Score	0–1 probability that the model assigned the correct intent	Use threshold (e.g., > 0.7) to reject low-confidence calls

CLU Lifecycle

Design → Label utterances → Train → Test → Deploy → Consume via SDK/REST

• Deployment slots: Maintain separate production and staging deployments so you can test a new model version before promoting it.
• Export/Import: Model definition exported as JSON — useful for version control or migrating between projects.

CLU vs Custom QA

CLU → maps an utterance to an intent + entities (structured output). Use when your app needs to take action based on what the user wants. Custom QA → matches a question to a stored answer (Q&A pair output). Use when your app needs to return information from a knowledge base.

---

5.3 Custom Question Answering

Custom QA (formerly QnA Maker) builds a knowledge base of Q&A pairs and returns the best matching answer to a user question.

Knowledge Base Sources

Source Type	Example
URLs	Public FAQ web pages, product documentation
Files	PDF manuals, Word documents, Excel/TSV Q&A spreadsheets
Manual entry	Directly authored Q&A pairs

Key Features

Feature	What It Does
Multi-turn conversations	Adds follow-up prompts to an answer (e.g., "Did that help?" → "Yes/No" branches)
Alternate phrasing	Add synonym questions to improve match rate
Chit-chat	Pre-built personality sets (Professional, Friendly, Witty) for small talk
Active Learning	Surfaces low-confidence questions for human review and improvement
Confidence threshold	Rejects answers below a set score — prevents wrong answers from being returned

Deployment

1. Create a project in Language Studio or AI Foundry
2. Add sources and label/edit Q&A pairs
3. Train and test
4. Deploy to a named endpoint
5. Consume via REST API (POST /knowledgebases/{kbId}/generateAnswer)

Multi-turn Q&A

The exam phrase "multi-turn Q&A from documents" maps to Custom Question Answering — not CLU. CLU is for intent detection; Custom QA is for retrieving stored answers.

---

5.4 Speech Services

Core Capabilities

Capability	Description	SDK Method
Speech-to-Text (STT)	Real-time or batch audio transcription	SpeechRecognizer
Text-to-Speech (TTS)	Synthesize neural voices with natural prosody	SpeechSynthesizer
Speech Translation	Real-time STT + translation in one pass	TranslationRecognizer
Intent Recognition	Detect CLU intents directly from spoken audio	IntentRecognizer with CLU model
Keyword Recognition	Local on-device detection of a wake word	KeywordRecognizer

Intent Recognition vs Keyword Recognition

Intent Recognition → understands what the user wants (requires CLU model, cloud call). Keyword Recognition → detects a specific activation word (e.g., "Hey Cortana") — runs locally on device, no cloud needed.

The exam uses "recognize spoken intent / commands" → Intent Recognition + Speech SDK. The exam uses "wake word / offline activation" → Keyword Recognition.

Generative Speaking (GenAI Speaking)

Combines Azure OpenAI with Speech TTS to produce expressive, context-aware spoken responses — used in AI agents that speak to users in real-time.

Batch Transcription

For large audio files that cannot be processed in real-time:

• Submit files stored in Azure Blob Storage
• Use an async pattern: submit → poll status → retrieve transcript
• Returns word-level timestamps and speaker diarization

---

5.5 Translator Service

Feature	Description
Text Translation	Translate text across 100+ languages in a single API call
Transliterate	Convert script without changing language (e.g., Japanese Kanji → Romaji)
Detect	Auto-detect the source language
Dictionary Lookup	Returns alternate translations and examples
Document Translation	Async translation of entire Word/PDF files, preserving layout
Custom Translator	Train a domain-specific translation model using parallel corpora (TMX/XLIFF files)

Document Translation Pattern

Document Translation is async — submit → get Operation-Location header → poll until complete → download translated files. Same 202 → GET pattern as the Read API and batch operations.

---

Flashcards

1 / 8

❓

(Click to reveal)

💡

---

← Domain 4 · Domain 6 →