Implement text analysis solutions
Language Model Text Analysis and Translation
ImportantExtract entities, topics, summaries, and schema-shaped JSON from text with generative prompting and Azure Language; detect sentiment, tone, unsafe content, and PII; translate with Azure Translator or language models; and adapt outputs to domain tasks such as compliance summaries.
Aligned to the AI-103 study guide with skills measured as of April 16, 2026; guide and product behavior verified October 12, 2026.
Why this matters
Text analysis is where generative models and purpose-built Foundry Tools overlap most. AI-103 rewards knowing when a deterministic Language API is the safer, cheaper choice, when an LLM with structured outputs is more flexible, and how Translator now offers both neural machine translation and LLM translation with tone and terminology controls.
Must Know
- Azure Language in Foundry Tools provides PII detection, language detection, named entity recognition, and Text Analytics for health as its core features; sentiment analysis, key phrase extraction, and summarization remain available but are positioned as legacy features with announced retirement dates.
- PII detection returns entity categories with confidence scores and a redacted text field, and supports redaction policies such as masking with entity labels or synthetic replacement.
- A language model with structured outputs can extract entities, topics, and summaries into a strict JSON schema, which suits fields that no prebuilt NER category covers.
- Sentiment analysis returns positive, neutral, or negative labels with confidence scores at the sentence and document level, and opinion mining links sentiment to specific aspects such as price or service. A language model can also classify tone, such as frustrated or formal, when a schema defines the allowed values.
- Content Safety text analysis and guardrails detect hate, sexual, violence, and self-harm content, and the PII guardrail control (preview) can flag personal data in model output.
- Azure Translator text translation (API version 2026-06-06) can use neural machine translation (NMT) or a language model. The LLM option adds tone and gender controls and requires a Foundry resource, while NMT handles larger request volumes and is billed per character.
- Custom Translator trains custom neural translation models from parallel documents for domain terminology, and adaptive custom translation steers LLM translations with reference sentence pairs. Document translation preserves file layout for whole documents.
- Domain customization for language model outputs uses precise instructions, few-shot examples, retrieved reference material, and output schemas; fine-tuning is reserved for consistent style or format that prompting cannot reach.
Compare and Distinguish
- Prebuilt NER versus custom NER versus generative extraction: fixed entity categories versus entities trained from your labels versus schema-driven extraction by a language model.
- Extractive versus abstractive summarization: selected source sentences versus newly written summary text.
- PII detection versus content safety: personal data identification and redaction versus harmful content classification.
- NMT versus LLM translation in Translator: high-volume, per-character translation versus tone-aware translation billed by tokens.
- Custom Translator versus adaptive custom translation: a trained NMT model versus reference examples that steer LLM output at request time.
Scenario examples
- Scenario: A bank must redact account numbers and names from chat logs before analysts read them. Think: Azure Language PII detection with a redaction policy.
- Scenario: A legal team needs each contract summarized into obligations, parties, and renewal dates as JSON. Think: a language model with a strict structured output schema and domain instructions.
- Scenario: Marketing copy must be translated with a formal tone for German customers. Think: Translator with the LLM translation option and the tone parameter.
Exam traps
- JSON mode or a prompt request for JSON does not guarantee the schema; strict structured outputs do.
- Custom Translator trains NMT models; it does not provide LLM tone controls.
- Content Safety does not redact PII; use Language PII detection or the PII guardrail control.
- Sentiment analysis and summarization in Azure Language have retirement dates, so new designs often use language models for those tasks.
Key takeaways
- Use Language for deterministic PII and entity work, and language models with schemas for flexible extraction.
- Pick NMT or LLM translation in Translator based on volume, tone, and terminology needs.
- Customize domain output with instructions, examples, retrieval, and schemas before fine-tuning.
How it works
- Language APIs accept documents in batches and return typed results with confidence scores and character offsets.
- Translator's 2026-06-06 API accepts an inputs array and returns a value array, choosing NMT or an LLM per request.
Objects and administrative surfaces
- Azure Language REST APIs and SDKs, and the Language MCP server for agents.
- Azure Translator text, document, and custom translation endpoints.
- Foundry playground with structured outputs for prompt-based extraction.
- Content Safety text analysis and Foundry guardrail PII controls.
When to use it
- Use NMT for high-volume baseline translation and LLM translation where tone or nuance matters.
- Use Text Analytics for health when clinical entities and relations are required.
Security and governance implications
- Redact PII before sending text to downstream logs or analytics.
- Review domain summaries with subject matter experts before using them for compliance decisions.
Troubleshooting signals
- If generative extraction omits fields, check that the schema lists them as required and that strict mode is on.
- If translations ignore company terminology, add a custom model or adaptive reference pairs.
More detail
- Extract entities, topics, summaries, and structured JSON with prompts and Foundry Tools.
- Detect sentiment, tone, safety issues, and sensitive content.
- Translate text with Azure Translator or LLM-powered translation flows.
- Customize language model outputs for compliance and domain extraction tasks.
Ready for the quiz?
- When is generative extraction a better fit than prebuilt NER?
- Which feature redacts personal data from text?
- What does the LLM translation option add over neural machine translation?
- How does opinion mining differ from document-level sentiment?
Related objectives
- D4.1.S1 — Implement solutions to extract entities, topics, summaries, and structured JSON outputs by using generative prompting and Foundry Tools
- D4.1.S2 — Configure detection of sentiment, tone, safety issues, and sensitive content
- D4.1.S3 — Build solutions that translate text by using Azure Translator in Foundry Tools or LLM‑powered translation flows
- D4.1.S4 — Customize language model outputs for domain tasks, such as compliance summarization and domain extraction