Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Apps and Agents Developer/Course Guide/topics:
The existing ContactInfo entity combines several distinct entity patterns into one heterogeneous label. Phone numbers, email addresses, and social-media handles have substantially different syntax, contextual clues, and boundary patterns.
Splitting the broad entity into Phone , Email , and SocialMedia gives the custom NER model a clearer and more internally consistent training signal. Every existing tagged span should then be relabeled using the correct specialized entity class before retraining.
This can improve precision because the model no longer needs to learn one extremely broad decision boundary that covers unrelated forms of contact data. More precise labels also help the model learn the correct start and end boundaries of each entity, addressing the reported false positives in which adjacent text is included.
Adding more examples under the same ambiguous ContactInfo schema may reinforce the underlying modeling ambiguity. Auto-labeling reduces annotation effort but does not correct the schema problem. Lowering a confidence threshold would generally cause more uncertain predictions to be accepted, tending to lower rather than improve precision.
Study Guide alignment: implement solutions that extract entities and customize language-model outputs for domain-specific extraction tasks . Microsoft Learn
====================
