Skip to content

Entity recognition

Entity recognition (NER) detects named entities — people, organizations, dates, money amounts, and the like — in transcripts. Detected entities are stored on the transcript's words with a confidence value, can be queried in the transcript search grammar (#ENTITY tokens), and feed data redaction rules that target entity types such as card numbers.

NER is administered under Administration › Conversation Analytics › Entity Recognition, a page with two tabs: Engines and Jobs.

Engines

Two engine types are supported:

Engine Languages
MiaRec NER API v1 en-US
Amazon Detect PII en, es

Engines follow the same model as the other analytics engines: name, status, Global or tenant-scoped visibility, and encrypted credentials — see the pattern described in Generative AI engines.

The entity types

The built-in vocabulary is 18 entity types: CARDINAL, DATE, EVENT, FAC, GPE, LANGUAGE, LAW, LOC, MONEY, NORP, ORDINAL, ORG, PERCENT, PERSON, PRODUCT, QUANTITY, TIME, WORK_OF_ART.

The NER job

The Jobs tab manages the Recognize entities job. Its settings follow the standard pipeline-job shape (source data, filtering criteria, post-processing, schedule — see the pipeline overview), plus NER-specific options:

  • Entity Recognition engine — which configured engine to use.
  • Recognize entities — all entity types, or a selected subset.
  • Test only — run detection without storing results.

Chain the job after transcription (its input must have transcripts) and, when redaction depends on entities, before the Redact data job.

Note

As an alternative to the dedicated job, entity recognition can also run as an AI task with the entity-recognition destination, using a generative AI engine instead of a dedicated NER engine.

The user-facing panel flag

Detected entities are shown to end users in a "Recognized Named Entities" panel on the conversation page, with click-to-seek playback. This panel is controlled by the Named Entity Recognition panel flag in Advanced Settings. The administration pages above stay visible regardless of the flag — with the flag off, entities are still detected and usable by redaction and search, but users do not see the panel.