Study Azure AI-900 NLP Tasks: key concepts, common traps, and exam decision cues.
NLP questions on AI-900 are mostly about separating what is being analyzed and what form the language takes. If the requirement starts from written text, audio, translation, or next-word prediction, that clue usually decides the workload.
| Workload | Strong clue | Typical result |
|---|---|---|
| sentiment analysis | “Is this feedback positive or negative?” | tone classification |
| entity recognition | “Pull names, dates, or organizations from text” | structured items from text |
| key phrase extraction | “Find the main phrases in this article” | important phrases |
| translation | “Convert text or speech to another language” | output in a different language |
| speech recognition | “Turn spoken words into text” | transcript |
| speech synthesis | “Read this text aloud” | generated speech audio |
| language modeling | “Predict likely word sequences or language patterns” | text-probability behavior |
| If the input or output starts with… | Strong first lane |
|---|---|
| written text | language workload |
| audio or spoken words | speech workload |
Translation can appear in both lanes. If the scenario is about typed text, think text translation. If the scenario begins with spoken input or spoken output, speech is often the stronger first fit.
| Pair | Keep this distinction clear |
|---|---|
| sentiment analysis vs generative AI | analyze tone in existing text vs create new text |
| speech recognition vs OCR | audio-to-text vs image-to-text |
| entity recognition vs classification | extract structured items vs assign a category |
| language modeling vs search | predicting language patterns vs retrieving documents |