Implement Text Analysis Solutions
The Azure AI Language toolbox
Know which feature answers which question. Sentiment analysis returns document- and sentence-level scores; opinion mining is required to attach sentiment to specific aspects or targets. Key phrase extraction surfaces salient terms; NER returns typed entities (person, location, organization, datetime, quantity); entity linking resolves mentions to a knowledge-base identity; PII detection finds and can redact sensitive spans using a configurable redaction policy (character mask, replacement, or removal). Summarization comes in extractive and abstractive flavors. Language detection is short-text sensitive - very short or mixed strings can return low confidence or an ambiguous result.
Conversational and custom models
Custom question answering builds a knowledge base from FAQ pages, documents, and URLs, supports multi-turn follow-ups, and is right when users ask free-form questions against a fixed body of answers. CLU (conversational language understanding) is for utterance → intent plus entity slots, trained with labeled utterances and deployed as a version. Translator covers text and document translation as a separate service.
Operational traps
Batch size and character limits per document cause truncation, not errors - split long inputs. Confidence scores below a threshold should route to a fallback or human review, not straight to automation. Frequent 429s mean the resource is throttled; reduce concurrency or raise quota.
Rule of thumb
FAQ-style answers to natural questions → custom question answering. Command routing with parameters → CLU. Redacting before storage or logging → PII detection, and check whether the policy preserves structure.
Note: verify exact per-document character and batch limits in current documentation before sizing a design.