Domain 4 · Implement text analysis solutions
Text analysis is the domain where the platform's older task-specific services meet generative models doing the same jobs through prompting. The exam reflects that overlap: many objectives can be met either with a Foundry Tool or with a language model, and the interesting questions are about which is appropriate. Speech then extends the same material into audio, which matters because agents increasingly talk rather than type.
Language model text analysis
The core extraction work is pulling entities, topics, summaries, and structured JSON out of unstructured text, using generative prompting and Foundry Tools. The structured-JSON requirement is the one to keep in mind: an answer that has to feed a downstream system needs a schema, not prose, and that constrains how the task is framed.
Classification covers sentiment, tone, safety issues, and sensitive content — note that safety and sensitive-content detection sit here as analysis tasks, not only as guardrails.
Translation is available through Azure Translator in Foundry Tools or through LLM-powered translation flows. Translator is the specialised, predictable option; an LLM flow is the one that can carry context and tone across the translation, which is why both remain on the objectives.
Finally, outputs are customized for domain tasks — compliance summarization and domain-specific extraction being the named examples. Generic summarisation and compliance summarisation are not the same task, and the distinction is the point.
Speech solutions
Speech begins with the two conversions: speech-to-text and text-to-speech, which together let an agent hold a spoken interaction rather than a typed one. Integrating speech as an agent modality is treated as a design decision in its own right, and custom speech models exist for the case where domain vocabulary — product names, jargon, accents — defeats the general model.
Beyond transcription, multimodal reasoning from audio inputs lets a model work from the audio directly rather than from a transcript, preserving signal that transcription discards. And speech translation closes the loop, rendering spoken input into another language through language models and Foundry Tools.
Quiz · Domain 4
A retail team submits product reviews to Azure AI Language sentiment analysis. At the document level the service returns a 'mixed' overall sentiment, but you need to know which individual statements drove the negative signal. Which characteristic of the sentiment feature lets you do this?