Task 4.1: Apply language model text analysis

• APrompt a deployed model to return the summary and topics in a fixed schema

  Correct

  Summaries and topics are a natural fit for generative prompting with structured output, and the guide pairs prompting with Foundry Tools for this kind of extraction.

• BCall Azure Language summarization and key phrase extraction for every ticket

  Incorrect

  Azure Language lists summarization and key phrase extraction as legacy capabilities, supported for existing implementations rather than new projects.

• CTrain a custom NER model to label the summary sentences and topic words

  Incorrect

  Entity recognition labels spans in the text; it doesn't write a summary.

• DRun language detection and keep the first three sentences as the summary

  Incorrect

  Language detection identifies the language; trimming text isn't summarizing it.

Azure Language separates core capabilities, recommended for new work, from legacy ones such as summarization and key phrases. For summaries and topics in a new build, prompt a model and constrain the output to a schema.

• AAzure Translator text translation of each message into English

  Incorrect

  Translation changes the language but carries the names and numbers straight through.

• BAzure AI Content Safety text moderation on each message

  Incorrect

  Text moderation classifies harm such as hate or violence; it doesn't find and mask personal data.

• CA system message asking the chat model to omit personal data

  Incorrect

  That shapes the model's own answers; it doesn't redact what users typed before it is stored, and returns no scores.

• DAzure Language text PII detection with a redaction policy

  Correct

  Text PII detects and redacts sensitive entities in raw strings with low latency, returning categories, offsets and confidence scores.

PII detection is a core Azure Language capability; for inline, synchronous redaction of strings before storage, use text PII.

• ASynchronous document translation, one request per file and language

  Incorrect

  Synchronous translation suits a single file in the response; batch is the method for numerous files in one job.

• BAzure Translator text translation on text extracted from each file

  Incorrect

  Text translation handles strings; the team would have to rebuild every document's layout around the translated text.

• CAzure Translator document translation with an asynchronous batch job

  Correct

  Batch document translation processes many or large files from Blob Storage in parallel and preserves the original document structure.

• DA language model prompted to translate each manual into six languages

  Incorrect

  A model can translate text, but it doesn't return formatted Word and PDF files, and 24,000 outputs would need custom assembly.

Document translation keeps structure and format and comes in two modes: synchronous for one file returned in the response, and asynchronous batch for many or large files staged in Blob Storage.

• AAzure Language language detection on each incoming message

  Incorrect

  Language detection names the language, not the tone or risk.

• BAzure AI Content Safety text analysis for the self-harm category

  Correct

  Content Safety classifies self-harm in text and returns a severity level the team can act on.

• CA language model prompted to return a tone label from a fixed list

  Correct

  Nuanced tone labels such as sarcasm fit generative prompting with a constrained set of outputs.

• DAzure Language entity linking on each message from the app

  Incorrect

  Entity linking matches entities to Wikipedia entries; it says nothing about tone or self-harm.

• EAn Azure AI Content Safety blocklist of words that signal tone

  Incorrect

  A list of tone words can't classify tone or grade self-harm risk.

Split the job: generative prompting handles open-ended judgments like tone, and Content Safety handles safety categories with calibrated severity levels.

• AConversation PII, built for multi-turn conversations

  Correct

  Conversation PII redacts across speakers and turns for chat and transcript workflows; its API and model are in preview.

• BPrebuilt NER, with person names filtered out by hand later

  Incorrect

  NER finds entities but doesn't redact, and manual filtering misses data.

• CLanguage detection on each customer message in the chat

  Incorrect

  Language detection doesn't find or redact personal data.

• DText PII run on each chat flattened into one string

  Incorrect

  Flattening the chat loses the turn and speaker structure the redaction should respect.

Transcripts have turns and speakers; choose the PII feature built for that structure rather than flattening it.

• ASend each post twice, once for each likely source language

  Incorrect

  That doubles cost and still guesses.

• BAlways set the source language to English for every post

  Incorrect

  A wrong source language produces poor translations.

• CLeave the source language unset so that it is detected

  Correct

  If the source language isn't specified, the service autodetects it and returns the detected language.

• DRun prebuilt NER on each post before translating it

  Incorrect

  Entity recognition doesn't identify the language.

Text translation can detect the source language itself; specify it only when you know it, for example to avoid confusion on very short text.

• AtextType set to html, with well-formed complete elements

  Correct

  textType defines whether text is plain or HTML, and HTML must be a well-formed, complete element.

• BDocument translation of each fragment as a separate PDF

  Incorrect

  Converting fragments to files adds work and changes the format.

• CtextType left at plain with tags stripped and re-added later

  Incorrect

  Rebuilding markup by hand is error-prone.

• DThe source language parameter set to html

  Incorrect

  html isn't a language; the content type goes in textType.

Tell the service what it's translating: plain text by default, or HTML so that markup is handled correctly.

• AAssume the results are as accurate as they are for English

  Incorrect

  The docs warn that performance can vary outside the tested set.

• BTest with representative Turkish content and measure errors

  Correct

  Classifiers can handle other languages, but performance can vary, so evaluate false positives and negatives with representative content.

• CTranslate every post to English and moderate only the translation

  Incorrect

  Translation can lose slurs and context, and still needs evaluation.

• DSwitch off the hate category for all of the Turkish posts

  Incorrect

  That removes protection instead of measuring it.

Safety classifiers are multilingual but trained and tested on a set of languages. For others, measure quality on your own data and tune thresholds or add human review.

• AAsk for the most creative and engaging summary possible

  Incorrect

  Creativity works against traceable, faithful summaries.

• BRaise the temperature so that the bullets differ more from each other

  Incorrect

  More randomness makes summaries less faithful.

• CNumber the source sentences and have each bullet cite them

  Correct

  Citing provided sources and stating an output contract make each statement traceable, which compliance needs.

• DRemove the source text and summarize from the title alone

  Incorrect

  Without the source there is nothing to trace back to.

Domain customization often starts in the prompt: a clear role, an output contract, citations to supplied text and a policy for missing information.