AI-103 Practice Questions: Implement text analysis solutions Domain

Correct answer (A): Entity extraction and key phrase extraction are designed to identify important items and concepts in text. That matches the need for tags and metadata. Summarization is for condensing content, sentiment analysis measures opinion, and speech-to-text is irrelevant because the input is already text.

Why the other options are wrong:
- Option B: Summarization creates shorter prose, but the team wants tags and metadata rather than condensed narratives.
- Option C: Sentiment analysis measures attitude or opinion, not named items and concepts.
- Option D: Speech-to-text is for spoken audio input, but the scenario already starts with text messages.

Correct answer (B): This is an offline transcription scenario with no live assistance requirement, so a batch-oriented speech-to-text pipeline is the best fit. The reference guidance is that real-time voice scenarios favor low-latency processing, while offline archive workflows can prioritize throughput and cost efficiency.

Why the other options are wrong:
- Option A: Streaming is more appropriate for live assistance or low-latency scenarios, not overnight batch processing.
- Option C: Text-to-speech converts text into audio, but the team needs audio converted into text.
- Option D: Translation is unnecessary because the requirement is transcripts, not output in another language.

Correct answer (B): Speech translation is the right choice when spoken input must be translated into another language, which is exactly the need for an English text version of spoken Italian. Text-to-speech is then used to synthesize the assistant's reply as spoken Italian. This matches the required speech and language conversion flow.

Why the other options are wrong:
- Option A: Speech-to-text would transcribe the audio, but it would not provide the required English text version of the spoken Italian input.
- Option C: Text-to-speech synthesizes audio from text. It does not convert incoming speech into text or translation.
- Option D: Text analysis capabilities operate on text, not raw audio. The audio must first be transcribed or translated from speech.

Correct answer (A): Sentiment analysis is the correct fit because the requirement is specifically to classify each comment as positive, negative, mixed, or neutral. The stem also makes it clear that structured business fields and summaries are not needed, so extraction and summarization would solve the wrong problem.

Why the other options are wrong:
- Option B: Entity extraction is for identifying items such as IDs, names, or organizations, but the scenario does not require field extraction.
- Option C: Summarization condenses text, but it does not reliably return the sentiment categories needed for dashboarding.
- Option D: Structured JSON extraction is useful when downstream systems need predictable fields, not when the main goal is sentiment classification.

Correct answer (A): The app needs two things: the right capability and the right security pattern. A dedicated PII detection capability fits the requirement to identify sensitive personal data, and managed identity with Azure RBAC is the preferred Azure-hosted access method because it avoids embedded secrets and supports least privilege.

Why the other options are wrong:
- Option B: Sentiment analysis is the wrong capability for identifying personal data, and storing API keys in source control directly violates the no-embedded-secrets requirement.
- Option C: Content safety filtering is not the best tool for dedicated PII identification, and using service keys still leaves secret management in the application path instead of using managed identity.
- Option D: Selecting or deploying a model endpoint does not by itself implement PII detection, secure the application, or configure least-privilege RBAC access.

Correct answer (B): Language detection is the direct and efficient first step because the messages already exist as text and the requirement is to identify the language before routing. Translating everything first adds cost and latency and is unnecessary when the goal is simply to detect the input language.

Why the other options are wrong:
- Option A: Translating first solves a different problem. It adds cost and latency when the actual requirement is just to identify the language of the source text.
- Option C: The source material is already text, so converting it to speech and then analyzing audio would add an unnecessary and inefficient step.
- Option D: Text-to-speech creates spoken audio from text. It does not identify the language for pipeline routing.

Correct answer (A): Sentiment analysis with opinion mining is the best fit because the requirement is standard text analysis on already-available text: classify sentiment and identify which aspects or features are being praised or criticized. This is exactly the kind of task Azure AI text analysis capabilities are designed for, and it is more direct than using open-ended generation.

Why the other options are wrong:
- Option B: A generative model can produce text, but rewriting comments as marketing copy does not solve the requirement to classify sentiment and identify praised or criticized features in a consistent analytics pipeline.
- Option C: Translation changes language, not meaning extraction. The stem does not say translation is required, so translating first adds unnecessary work and does not directly perform sentiment or opinion mining.
- Option D: Speech-to-text is for spoken audio. The source material is already text, so introducing a speech step is unnecessary and does not directly provide sentiment analysis.

Correct answer (B): The requirement is to identify main topics from multilingual text, not to translate it. If the chosen analysis capability supports the source languages, running key phrase extraction directly on the source text is the better design because it avoids unnecessary translation cost and latency. Translation should be added only when a translated output is actually needed.

Why the other options are wrong:
- Option A: This adds an unnecessary translation stage. The stem explicitly says the analysis capability supports the source languages and the business does not primarily need translated output.
- Option C: The reviews are already text, so moving to a speech workflow would introduce unnecessary processing rather than reduce it.
- Option D: Content safety filtering checks for harmful content categories. It does not extract the main discussion topics from reviews.

Correct answer (B): A production evaluation plan should cover both quality and operations. The app needs metrics for extraction accuracy or summary relevance, safety outcomes, latency, throughput, token usage, malformed output rate, and cost. Content safety is important, but it does not prove correctness or groundedness, so it cannot be the only gate.

Why the other options are wrong:
- Option A: Content safety pass rates do not prove that summaries are accurate or that extracted fields are correct, and request counts alone do not capture latency, malformed output, or cost issues.
- Option C: Prompt review is useful, but it is not enough for production readiness. The team still needs measurable quality and operational metrics for the running application.
- Option D: Changing or deploying a model endpoint does not automatically validate output quality, implement monitoring, or satisfy governance requirements.

Correct answer (B): When downstream automation depends on reliable structure, plain prompting is not enough. A schema-constrained response or tool/function calling is more robust for custom business fields because it guides the model toward valid machine-readable output and reduces malformed JSON compared with prompt-only instructions.

Why the other options are wrong:
- Option A: Prompt wording can help quality, but it does not guarantee valid JSON or required fields when the workflow depends on strict structure.
- Option C: Sentiment analysis is a built-in text analysis task for tone, not a method for extracting a custom business schema from free-form emails.
- Option D: Content safety filtering helps detect harmful content categories, but it does not validate that the JSON is complete, well formed, or correct.