What's new in Azure AI Language?

Azure AI Language is updated on an ongoing basis. Bookmark this page to stay up to date with release notes, feature enhancements, and our newest documentation.

July 2025

Expanded .NET SDK support for text and conversation authoring APIs:

June 2025

New version of the Conversational Language Understanding (CLU) training configuration. This new version is aimed at minimizing over-predictions of the None intent—particularly in multilingual contexts—is now available via the REST API using trainingConfigVersion 2025-07-01-preview. For more information, see Train your model: request body data.

Updated Build your conversational agent accelerator project. The update includes a new routing strategy—TRIAGE_AGENT. This strategy uses an agent hosted on Azure AI Foundry Agent Service. It utilizes Conversational Language Understanding (CLU) and Custom Question Answering (CQA) as tools to triage user intent for downstream agent routing. Additionally, these tools help deliver precise answers to specific questions. For more information, see TechCommunity Blog Post

.NET SDKs support. The following .NET SDKs are now available, and support the latest REST API version 2025-15-05-preview:

  • Azure.AI.Language.Text 1.0.0-beta.3 provides inference capabilities for a wide range of language processing tasks. These tasks include language detection, sentiment analysis, key phrase extraction, and named entity recognition (NER). The capabilities also include recognizing and linking personally identifiable information (PII) entities. Additionally, they offer text analytics for healthcare, custom named entity recognition (NER), and custom text classification. Both extractive and abstractive text summarization are also supported.

  • Azure.AI.Language.Conversation 2.0.0-beta.3 provides inference capabilities for conversational PII, conversational language understanding (CLU), and conversation summarization.

Text PII GPU container is now available for integration. You can access this container on the Microsoft Artifact Registry using the tag gpu.

May 2025

2025-05-15-preview release. The latest API preview version includes updates for named entity recognition (NER) and PII detection:

  • New entity type support for DateOfBirth, BankAccountNumber, PassportNumber, and DriversLicenseNumber.
  • Improved AI quality for PhoneNumber entity type.

PII detection enhancements. Azure AI Language introduces new customization and entity subtype features for PII detection:

March 2025

  • Back-end infrastructure for the Named entity recognition (NER) and Text Personally identifiable information (PII) entity recognition models is now updated with extended context window limits.

  • Our Conversational PII redaction service is now powered by an upgraded GA model. This revised version enhances the quality and accuracy of Credit Card Number entities and Numeric Identification entities. These entities include Social Security numbers, Driver's license numbers, Policy numbers, Medicare Beneficiary Identifiers, and Financial account numbers.

February 2025

  • Document and text abstractive summarization is now powered by fine-tuned Phi-3.5-mini! Check out the Announcing Blog for more information.

January 2025

  • .NET SDK for Azure AI Language text analytics, Azure.AI.Language.Text 1.0.0-beta.2, is now available. This client library supports the latest REST API version, 2024-11-01, and 2024-11-15-preview, for the following features:
    • Language detection
    • Sentiment analysis
    • Key phrase extraction
    • Named entity recognition (NER)
    • Personally identifiable information (PII) entity recognition
    • Entity linking
    • Text analytics for health
    • Custom named entity recognition (Custom NER)
    • Custom text classification
    • Extractive text summarization
    • Abstractive text summarization

November 2024

  • Runtime Container for Conversational Language Understanding (CLU) is available for on-premises connections.
  • Both our Text PII redaction service and our Conversational PII service preview API (version 2024-11-15-preview) now support the option to mask detected sensitive entities with a label beyond just redaction characters. Customers can specify if personal data content such as names and phone numbers, that is, "John Doe received a call from 424-878-9192" are masked with a redaction character, that is, "******** received a call from ************" or masked with an entity label, that is, "PERSON_1 received a call from PHONENUMBER_1." More on how to specify the redaction policy style for your outputs can be found in our how-to guides.
  • Native document support gating is removed with the latest API version, 2024-11-15-preview, allowing customers to access native document support for PII redaction and summarization. Key updates in this version include:
    • Increased Maximum File Size Limits (from 1 MB to 10 MB).
    • Enhanced PII Redaction Customization: Customers can now specify whether they want only the redacted document or both the redacted document and a JSON file containing the detected entities.
  • Language detection is a built-in feature designed to identify the language in which a document is written. It provides a language code that corresponds to a wide array of languages. This feature includes not only standard languages but also their variants, dialects, and certain regional or cultural languages. Today the general availability of script detection capability, and 16 more languages support, which adds up to 139 total supported languages is announced.
  • Named Entity Recognition service, Entity Resolution was upgraded to the Entity Metadata starting in API version 2023-04-15-preview. If you're calling the preview version of the API equal or newer than 2023-04-15-preview, check out the Entity Metadata article to use the resolution feature. The service now supports the ability to specify a list of entity tags to be included into the response or excluded from the response. If a piece of text is classified as more than one entity type, the overlapPolicy parameter allows customers to specify how the service handles the overlap. The inferenceOptions parameter enables users to modify the inference process, such as preventing detected entity values from being normalized and added to the metadata. Along with these optional input parameters, we support an updated output structure (with new fields tags, type, and metadata) to ensure enhanced user customization and deeper analysis Learn more on our documentation.
  • Text Analytics for Health (TA4H) is a specialized tool designed to extract and categorize key medical details from unstructured sources. These sources include doctor's notes, discharge summaries, clinical documentation, and electronic health records. Today, we released support for Fast Healthcare Interoperability Resources (FHIR) structuring and temporal assertion detection in the Generally Available API.

October 2024

September 2024

August 2024

July 2024

  • Conversational PII redaction service in English-language contexts is now Generally Available (GA).
  • Conversation Summarization now supports 12 added languages in preview as listed here.
  • Summarization Meeting or Conversation Chapter titles features support reduced length to focus on the key topics.
  • Enable support for data augmentation for diacritics to generate variations of training data for diacritic variations used in some natural languages, which are especially useful for Germanic and Slavic languages.

February 2024

  • Expanded language detection support for added scripts according to the ISO 15924 standard is now available starting in API version 2023-11-15-preview.

January 2024

December 2023

  • Text Analytics for health new model 2023-12-01 is now available.
  • New Relation Type: BodySiteOfExamination
  • Quality enhancements to support radiology documents
  • Significant latency improvements
  • Various bug fixes: Improvements across NER, Entity Linking, Relations, and Assertion Detection

November 2023

July 2023

April 2023

  • The latest model version (2022-10-01) for Language Detection now supports 6 more International languages and 12 Romanized Indic languages.

March 2023

February 2023

  • Conversational language understanding and orchestration workflow now available in the following regions in the sovereign cloud for China:
    • China East 2 (Authoring and Prediction)
    • China North 2 (Prediction)
  • New model evaluation updates for Conversational language understanding and Orchestration workflow.
  • New model version ('2023-01-01-preview') for Text Analytics for health featuring new entity categories for social determinants of health.
  • New model version ('2023-02-01-preview') for named entity recognition features improved accuracy and more language support with up to 79 languages.

December 2022

  • New version (v5.2.0-beta.1) of the text analysis client library is available in preview for C#/.NET:
  • New model version (2022-10-01) released for Language Detection. The new model version comes with improvements in language detection quality on short texts.

November 2022

October 2022

September 2022

August 2022

July 2022

  • New AI models for sentiment analysis and key phrase extraction based on z-code models, providing:

    • Performance and quality improvements for the following 11 languages supported by sentiment analysis: ar, da, el, fi, hi, nl, no, pl, ru, sv, tr
    • Performance and quality improvements for the following 20 languages supported by key phrase extraction: af, bg, ca, hr, da, nl, et, fi, el, hu, id, lv, no, pl, ro, ru, sk, sl, sv, tr

  • Conversational PII is now available in all Azure regions supported by the Language service.

  • A new version of the Language API (2022-07-01-preview) is available. It provides:

    To use this version in your REST API calls, use the following URL:

    <your-language-resource-endpoint>/language/:analyze-text?api-version=2022-07-01-preview

June 2022

May 2022

  • PII detection for conversations.

  • Rebranded Text Summarization to Document summarization.

  • Conversation summarization is now available in public preview.

  • The following features are now Generally Available (GA):

    • Custom text classification
    • Custom Named Entity Recognition (NER)
    • Conversational language understanding
    • Orchestration workflow

  • The following updates for custom text classification, custom Named Entity Recognition (NER), conversational language understanding, and orchestration workflow:

    • Data splitting controls.
    • Ability to cancel training jobs.
    • Custom deployments can be named. You can have up to 10 deployments.
    • Ability to swap deployments.
    • Auto labeling (preview) for custom named entity recognition
    • Enterprise readiness support
    • Training modes for conversational language understanding
    • Updated service limits
    • Ability to use free (F0) tier for Language resources
    • Expanded regional availability
    • Updated model life cycle to add training configuration versions

April 2022

  • Fast Healthcare Interoperability Resources (FHIR) support is available in the Language REST API preview for Text Analytics for health.

March 2022

February 2022

  • Model improvements for latest model-version for text summarization

  • Model 2021-10-01 is Generally Available (GA) for Sentiment Analysis and Opinion Mining, featuring enhanced modeling for emojis and better accuracy across all supported languages.

  • Question Answering: Active learning v2 incorporates a better clustering logic providing improved accuracy of suggestions. It considers user actions when suggestions are accepted or rejected to avoid duplicate suggestions, and improve query suggestions.

December 2021

  • The version 3.1-preview.x REST endpoints and 5.1.0-beta.x client library are retired. Upgrade to the General Available version of the API(v3.1). If you're using the client libraries, use package version 5.1.0 or higher. See the migration guide for details.

November 2021