Microsoft Azure AI Fundamentals (AI-900) practice exam

Key phrase extraction is correct because it identifies the most important words and phrases in a body of text, which is exactly how you surface recurring topics across many survey responses. Sentiment analysis tells you whether each response is positive or negative, not what topics the responses are about. Text-to-speech does the opposite of what is needed: it converts text into spoken audio. Language detection only identifies which language the text is written in (e.g., English or Spanish); it does not extract topics.

'Unsupervised learning using clustering' is correct: when there are no labels and the goal is to find natural groupings in the data, clustering (an unsupervised technique) is the appropriate choice. 'Supervised learning using classification' is wrong because classification requires labeled examples that define the categories in advance, which this team does not have. 'Supervised learning using regression' is wrong because regression predicts a continuous numeric target from labeled data, and there is neither a numeric target nor labels here. 'Reinforcement learning using a reward signal' is wrong because that approach learns through trial-and-error interactions guided by rewards, which does not match the task of grouping a static set of records.

Requiring the system to behave consistently and safely, including under unexpected conditions and after rigorous testing, is the reliability and safety principle. Transparency is wrong because it is about helping people understand the system, not about robust and safe operation. Inclusiveness is wrong because it concerns making the system usable by people of all abilities, not its operational dependability. Fairness is wrong because it concerns equitable treatment across groups; a system can be fair yet still behave unreliably under edge cases, so the two principles are distinct.