AI-102 · Domain 5 of 6

AI-102 Domain 5: Implement natural language processing solutions

Text analytics, translation, speech and custom language models in one wide skill area. Most questions hinge on picking the right Language, Translator or Speech feature and knowing its exact setup steps, outputs and limits.

Exam weight
15-20%
Official skill groups
3
Estimated questions estimate
6–12
Suggested study time estimate
15–20 hrs

About 18 min read Reviewed October 2026

Overview

This skill area spans three services. Azure Language analyzes text for key phrases, entities, sentiment, language and personally identifiable information (PII), and hosts custom models: conversational language understanding (CLU) for intents and entities, and custom question answering for knowledge bases. Azure Translator translates text and whole documents and supports custom translation models. Azure Speech converts speech to text and text to speech, shapes voices with Speech Synthesis Markup Language (SSML), trains custom speech models, recognizes intents and keywords, and translates speech.

The outline is long because each feature has its own lifecycle: create, train, evaluate, deploy, consume, back up and export. The exam tests whether you can follow that lifecycle in the right order and pick the feature that matches the input and the output. It connects to Domain 1 for service selection and keys, and to Domain 2, where generative models now handle some of the same jobs.

How it's tested

Questions combine service selection with lifecycle steps. Expect a requirement such as detecting customer frustration, routing chat by intent or translating spoken meetings, and you choose the feature; then expect ordering items for building and publishing a CLU model or knowledge base, plus code and SSML completions.

  • Feature matching

    Input and output are described and you choose between key phrases, NER, entity linking, sentiment, PII, language detection, translation, CLU or question answering.

  • Lifecycle ordering

    You order the steps to create, train, evaluate, deploy and consume a CLU model, a knowledge base or a custom translation model.

  • Speech and SSML details

    You complete a Speech SDK snippet or an SSML document to change voice, speed, pauses or pronunciation, or to translate speech.

Reasoning it through

A bank wants its app to understand 'move 200 from savings to checking' and extract the amount and both accounts. Sentiment and key phrases do not identify an action. Custom question answering returns stored answers to questions, not actions with parameters. A CLU model with a TransferFunds intent and entities for amount and account matches the need. When the user wants the system to do something, think intent plus entities.

Skill groups 3

Official skill groups from the Microsoft outline, in outline order.

5.1 Analyze and translate textKey phrases, entities, sentiment, language detection, PII, and Translator for text and documents.
What it means
Azure Language offers prebuilt text analysis: key phrase extraction, named entity recognition (NER), entity linking, sentiment analysis with opinion mining, language detection and PII detection with redaction. Azure Translator translates text in real time through REST calls and translates whole documents asynchronously while preserving their formatting.
What to understand
Sentiment returns positive, neutral, negative or mixed labels with confidence scores at document and sentence level, and opinion mining links opinions to aspects such as room or price. Language detection returns a language name, ISO 639-1 code and confidence score. PII detection returns entities and a redacted version of the text. Text translation can send one request with several to= target languages and detects the source language if from is omitted. Document translation reads from and writes to Azure Blob Storage containers, authorized with SAS (shared access signature) tokens or managed identity.
How it shows up
Expect a requirement to pick the right analysis feature, a REST call to complete with the right query parameters or headers, or a document translation setup that needs storage containers.
Common confusion
NER and entity linking are mixed up. NER labels a span with a category such as Person or Location; entity linking resolves it to a specific entry in a knowledge base, such as which Paris is meant.
5.2 Process and translate speechGenerative speaking, speech to text, text to speech, SSML, custom speech, intent and keyword recognition, and speech translation.
What it means
Azure Speech converts audio to text and text to natural-sounding neural voices, accepts SSML for fine control, trains custom speech models on your audio and transcripts, recognizes keywords on device and intents from speech, and translates spoken audio into text or synthesized speech in other languages. Generative speaking capabilities pair language models with speech input and output for voice assistants.
What to understand
The SpeechConfig object carries the key and region (or endpoint) plus recognition language and voice name; recognizers and synthesizers do the work. SSML uses speak, voice, prosody (rate, pitch, volume), break, say-as, phoneme and mstts:express-as for speaking styles. Custom speech improves recognition of domain vocabulary or accents using plain text, structured text, or audio with human-labeled transcripts, and is measured by word error rate (WER). Speech translation uses a translation config with a source language and one or more target languages, and can also synthesize the translated speech.
How it shows up
Expect an SSML snippet to complete, a choice between custom speech and a phrase list, or a TranslationRecognizer setup with target languages.
Common confusion
Custom speech and custom neural voice get confused. Custom speech improves recognition (speech to text); custom neural voice creates a branded synthetic voice (text to speech) and requires limited-access approval.
5.3 Implement custom language modelsCLU intents and entities, model lifecycle, custom question answering knowledge bases, and custom translation.
What it means
CLU predicts a user's intent and extracts entities from utterances you label. Custom question answering builds a knowledge base of question-and-answer pairs from URLs, files and manual entries, with multi-turn follow-ups, alternate phrasing and chit-chat. Custom Translator trains translation models on parallel documents for domain-specific terminology.
What to understand
CLU lifecycle: create intents and entities (learned, list, prebuilt, regex), add and label utterances including the None intent, train, review precision, recall and F1 score, deploy to a named deployment, then call it with the project and deployment names. Back up and move a project by exporting and importing its JSON. Question answering lifecycle: create the project, add sources and pairs, test, then publish (deploy) before clients can query; export the project to back it up or move it. Custom Translator trains with aligned source and target documents and reports a BLEU score; clients use its category ID.
How it shows up
Expect ordered steps for building and publishing a model, a choice between CLU and question answering, or a requirement for multi-turn conversation or multiple languages.
Common confusion
Treating CLU and question answering as interchangeable. CLU tells your app what the user wants to do so your code can act; question answering returns a stored answer to a question.

High-yield concepts

The terms, rules and formulas this domain turns into questions most often.

  • Sentiment analysis and opinion mining

    Sentiment returns positive, neutral, negative or mixed labels and confidence scores for the document and each sentence; opinion mining links opinions to specific aspects.

    Remember
    Mixed appears at document level when sentences disagree; enable opinion mining to learn which aspect, such as delivery or price, drew the complaint.
    Tested as
    Scenarios needing the reason for negative feedback point to opinion mining.
    Trap
    Expecting overall sentiment to say which product feature was criticized.

  • Language detection output

    Language detection returns the language name, ISO 639-1 code and a confidence score for each document.

    Remember
    Ambiguous or mixed text returns the predominant language with lower confidence; unrecognized text returns an unknown result.
    Tested as
    Response-reading items ask which field holds the code or how to handle low confidence.
    Trap
    Assuming detection splits a mixed-language document into each language.

  • PII detection

    PII detection finds personal data such as names, phone numbers and identification numbers and returns entities plus a redacted text string.

    Remember
    You can restrict detection to chosen categories and use a health domain for protected health information.
    Tested as
    Compliance scenarios that need masked transcripts point here.
    Trap
    Using general NER for redaction, which categorizes entities but does not mask them.

  • Translator text request

    Text translation is a REST POST to the translate operation with api-version and one or more to parameters, authenticated with the key and, for regional or multi-service resources, a region header.

    Remember
    Omit from to auto-detect the source language; send several to parameters to get several translations in one call.
    Tested as
    Code completions test query parameters and the Ocp-Apim-Subscription-Region header.
    Trap
    Making one request per target language.

  • Document translation

    Document translation asynchronously translates whole files such as Word, PDF and PowerPoint while keeping layout, reading from a source container and writing to a target container.

    Remember
    It needs the Translator resource's custom domain endpoint and storage access through SAS tokens or a managed identity.
    Tested as
    Batch translation of files points here, not to text translation.
    Trap
    Extracting text and calling text translation, which loses formatting.

  • SSML elements

    SSML wraps text in speak and voice elements and uses prosody for rate, pitch and volume, break for pauses, say-as for dates and numbers, phoneme for pronunciation and mstts:express-as for speaking style.

    Remember
    Use say-as interpret-as for how a value is read and phoneme for how a word is pronounced.
    Tested as
    Completion items show an SSML document with one missing element.
    Trap
    Using prosody to fix the pronunciation of a product name.

  • Custom speech

    Custom speech adapts recognition with plain text, structured text, pronunciation data or audio with human-labeled transcripts, deployed to a custom endpoint.

    Remember
    Measure improvement with word error rate (WER), comparing the custom model with the base model on test audio.
    Tested as
    Industry jargon or accent problems in transcripts point here.
    Trap
    Choosing custom neural voice, which changes the output voice rather than recognition.

  • Keyword vs. intent recognition

    Keyword recognition listens on device for a wake phrase using a custom keyword model; intent recognition maps spoken requests to intents through pattern matching or a CLU model.

    Remember
    Wake words run locally before streaming audio to the cloud; intents are evaluated after recognition.
    Tested as
    Voice assistant scenarios ask for one or the other.
    Trap
    Using CLU to detect the wake word.

  • Speech translation

    A SpeechTranslationConfig sets the recognition language and target languages, and a TranslationRecognizer returns translations; a voice name adds synthesized translated audio.

    Remember
    Add one target language per call to add_target_language (or AddTargetLanguage); speech-to-speech needs a synthesis voice.
    Tested as
    Live meeting translation scenarios test this setup.
    Trap
    Chaining speech to text, Translator and text to speech by hand when one recognizer does it.

  • CLU entity types

    CLU supports learned entities labeled in utterances, list entities with fixed values and synonyms, prebuilt entities such as numbers and dates, and regex entities.

    Remember
    Use a list entity for a closed set such as store locations and a prebuilt entity for common types such as datetime.
    Tested as
    Design questions ask which entity type fits a described value.
    Trap
    Labeling hundreds of examples for values a list entity would match exactly.

  • CLU evaluation

    CLU reports precision, recall and F1 score per intent and entity, plus a confusion matrix after training.

    Remember
    Fix confusion between two intents by adding varied utterances to each, and use the None intent for out-of-scope input.
    Tested as
    Questions show two intents often mistaken for each other and ask for the fix.
    Trap
    Adding identical utterances, which does not help the model separate intents.

  • Question answering knowledge base

    A custom question answering project holds question-and-answer pairs from URLs, files and manual edits, with alternate questions, follow-up prompts and optional chit-chat.

    Remember
    Test in the portal, then deploy (publish) before clients can query; export the project for backup or migration.
    Tested as
    Ordering questions often place publish after testing and before client integration.
    Trap
    Querying the knowledge base from an app before it is published.

  • Multi-language question answering

    When creating the first project in a Language resource, you choose whether all projects share one language or each project can have its own.

    Remember
    For several languages, create one project per language or choose per-project language at setup; the choice cannot be changed later for that resource.
    Tested as
    Global support bots test this setup decision.
    Trap
    Expecting one project to answer in several languages automatically.

  • Custom Translator

    Custom Translator trains a translation model on parallel documents aligned sentence by sentence for your domain.

    Remember
    Training reports a BLEU score; publish the model and call Translator with its category ID.
    Tested as
    Requirements for consistent domain terminology in translations point here.
    Trap
    Using a glossary-free generic model when terminology consistency is the stated requirement.

Side by side

Concepts candidates mix up, compared on the point that decides the answer.

CLU vs. custom question answering vs. text analytics

ConceptMeaningKey differenceCommon trap
CLU Predicts intent and extracts entities from a user utterance Your app acts on the result Using it to return FAQ answers
Custom question answering Returns the best stored answer for a question Answers come from a knowledge base Using it to trigger actions with parameters
Prebuilt text analytics Sentiment, key phrases, NER, PII, language No training; generic analysis Building a custom model for something prebuilt already does

Named entity recognition vs. entity linking vs. key phrase extraction

ConceptMeaningKey differenceCommon trap
NER Labels spans with categories such as Person or Location Says what kind of thing it is Expecting disambiguation
Entity linking Resolves entities to a knowledge base entry Says which specific thing it is Confusing it with NER
Key phrase extraction Returns the main talking points No category or identity Using it to find people or places

Custom speech vs. custom neural voice vs. phrase list

ConceptMeaningKey differenceCommon trap
Custom speech Trained recognition model for your vocabulary and audio Improves speech to text accuracy Confusing it with voice creation
Custom neural voice A branded synthetic voice Improves text to speech output; limited access Choosing it for transcription errors
Phrase list Runtime hints for words likely to be spoken No training; quick boost for a few terms Expecting it to fix accents or noisy audio

Text translation vs. document translation vs. speech translation

ConceptMeaningKey differenceCommon trap
Text translation Synchronous REST call for strings Real time, several targets per request Using it for formatted files
Document translation Asynchronous batch translation of files in Blob Storage Preserves layout; needs containers and access Forgetting storage authorization
Speech translation Spoken audio to translated text or speech Uses the Speech service, not Translator directly Chaining three services by hand

Exam traps

Each trap is an answer that looks right. Learn the clue that should change your mind.

  1. FAQ bot for actions

    What candidates pick
    Custom question answering for a bot that books meeting rooms.
    Why it looks right
    Both are conversational and handle natural questions.
    Why it's wrong
    Booking needs an intent plus entities such as room and time for code to act on; question answering only returns stored answers.

    Clue to watch for Verbs like book, transfer, cancel or order with parameters.

  2. Prosody for pronunciation

    What candidates pick
    Adjust prosody rate and pitch to fix how a brand name is spoken.
    Why it looks right
    Prosody is the best-known SSML element.
    Why it's wrong
    Prosody changes speed, pitch and volume, not phonetics; use phoneme or a custom lexicon for pronunciation.

    Clue to watch for A specific word is pronounced wrongly, not too fast or too flat.

  3. Text translation for files

    What candidates pick
    Extract text from PDFs and call the translate operation.
    Why it looks right
    Text translation is simple and synchronous.
    Why it's wrong
    It loses layout and needs custom handling for large files; document translation preserves formatting and runs as a batch over storage containers.

    Clue to watch for Files, formatting or batch volume in the requirement.

  4. Unpublished knowledge base

    What candidates pick
    Connect the app to the knowledge base right after adding sources and testing.
    Why it looks right
    The test pane already returns good answers.
    Why it's wrong
    Clients query the deployed (published) knowledge base; until you deploy, the production endpoint has nothing to serve.

    Clue to watch for An ordering question where publish or deploy is missing between test and consume.

  5. Custom voice for jargon

    What candidates pick
    Train a custom neural voice so medical terms transcribe correctly.
    Why it looks right
    Both features start with 'custom' and involve speech.
    Why it's wrong
    Transcription accuracy is speech to text; custom speech trained with domain text or labeled audio is the fix.

    Clue to watch for Errors appear in transcripts, not in generated audio.

  6. One project, many languages

    What candidates pick
    Add answers in five languages to one question answering project.
    Why it looks right
    It keeps all content in one place.
    Why it's wrong
    A project is tied to a language; supporting several languages needs separate projects with per-project language chosen when the first project is created in the resource.

    Clue to watch for A global support bot that must answer in several languages.

  7. Manual speech translation chain

    What candidates pick
    Speech to text, then Translator, then text to speech in three separate calls.
    Why it looks right
    Each step is familiar.
    Why it's wrong
    The Speech service's translation recognizer does recognition and translation in one stream and can synthesize the result, with lower latency and less code.

    Clue to watch for Live or real-time spoken translation.

Answer framework

Use this sequence for any language, translation or speech question on AI-102.

  1. Classify the input as text, a formatted document or audio, and the output as labels, a stored answer, an action with parameters, translated content or speech.
  2. Use a prebuilt feature first: sentiment, key phrases, NER, entity linking, PII, language detection or Translator, and move to a custom model only when prebuilt cannot meet the requirement.
  3. For actions with parameters choose CLU; for answers to known questions choose custom question answering; for domain terminology in translations choose Custom Translator.
  4. Follow the lifecycle in order: define, label or add sources, train, evaluate, deploy or publish, then consume with the project and deployment names.
  5. For speech, separate recognition fixes such as custom speech and phrase lists from output fixes such as SSML and voices, and use the translation recognizer for spoken translation.

Practice question

A text-to-speech app reads product announcements. The product name Xylar is pronounced incorrectly, while speed and tone are fine. Which SSML change should you make for that word?

Want more like this? The AI-102 practice test has questions for every domain.

Checklist

0 of 14

Tick each one you can do without notes. Your ticks stay in this browser.

Study strategy

Start with the prebuilt Language features, because they are quick to learn and appear often. Run one paragraph of customer feedback through sentiment with opinion mining, key phrases, NER, entity linking, PII and language detection, and write down what each returns. Then call Translator from code with two target languages and set up one document translation job against two storage containers.

Next, build the two custom language models end to end, which is where the highest-priority lifecycle questions come from. Create a CLU project with three intents, a list entity and a prebuilt entity, train it, read the confusion matrix, deploy it and call it. Then build a question answering project from a public FAQ page, add alternate phrasing, a follow-up prompt and chit-chat, publish it, and export it.

Commonly underestimated areas are speech details: SSML elements, custom speech data types and WER, keyword versus intent recognition, and translation recognizer setup. Write two SSML documents by hand and run a speech translation sample.

Afterwards, practice mixed questions that contrast CLU with question answering and custom speech with custom voice, then review Domain 6, where language skills enrich indexed documents.

FAQ

How much of AI-102 covers natural language processing?

Implement natural language processing solutions is weighted at 15-20% in the final AI-102 study guide. It has the longest list of bullet tasks of any skill area, so plan time for breadth.

What is the difference between CLU and custom question answering?

CLU identifies what the user wants to do and extracts entities so your code can act. Custom question answering returns a stored answer to a question from a knowledge base. Many bots use both, routing between them.

Do I need to memorize SSML syntax?

Know the main elements and what each controls: voice, prosody for rate and pitch, break for pauses, say-as for how values are read, phoneme for pronunciation and express-as for speaking style. Questions usually ask you to choose or complete one element.

When should I use custom speech instead of a phrase list?

A phrase list is a quick runtime hint for a handful of words. Custom speech is a trained model for broader vocabulary, accents or acoustic conditions, measured with word error rate.

Why does document translation need storage accounts?

It is an asynchronous batch operation that reads files from a source Blob Storage container and writes translated files to a target container, so the service needs access through SAS tokens or a managed identity.

Can a generative model replace these language features?

Often it can, but the exam favors the purpose-built feature when the requirement stresses consistent structured output, lower cost or a specific capability such as PII redaction or document translation.

How does this domain relate to knowledge mining?

Language features such as entity recognition, key phrases, PII and translation run as built-in skills in Azure AI Search skillsets, enriching documents before they are indexed.

Other Artificial Intelligence & Machine Learning certifications

Where the same subject shows up in other exams. Cards marked "Same subject" open the matching domain.

Test your Domain 5 knowledge

You have just separated prebuilt text analytics from custom language models, walked through the CLU and question answering lifecycles, and matched SSML elements to the problems they fix. Practice questions test whether you can do that quickly across mixed scenarios.

Take the AI-102 practice test Check your AI-102 readiness

Keep going