Bert_NER_Ausa

A lightweight token-classification (NER) model that extracts structured entities from short health-assistant utterances — appointments, symptoms, allergies, routines/medication, and profile details. It is the entity-extraction stage of the AUSA Hub voice/text assistant, paired with the SetFit intent router aadiausa/Set_Fit_Ausa.

Model details

• Architecture: BertForTokenClassification (TinyBERT — 4 hidden layers, hidden size 312, 12 attention heads, ~14.5M parameters)
• Tokenizer: WordPiece, uncased (bert-base-uncased vocab, 30522 tokens)
• Max sequence length: 512 tokens
• Tagging scheme: BIO — 55 labels = O + B-/I- for 27 entity types

Entity types (27)

Domain	Entities
Symptoms & allergies	ALLERGY, SYMPTOM, SEVERITY, ONSET, DOSAGE
Scheduling	PROVIDER, DATE, START_TIME, END_TIME, ROUTINE, FREQUENCY, SCHEDULED_TIME, START_DATE, END_DATE, INTERVAL, DURATION, DAY_OF_WEEK
Profile / contacts	FULL_NAME, EMAIL, PHONE, ADDRESS, GENDER, HEIGHT, WEIGHT, RELATION, INVITE_METHOD, PERMISSION

Usage

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

model_id = "aadiausa/Bert_NER_Ausa"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForTokenClassification.from_pretrained(model_id)

ner = pipeline("token-classification", model=model, tokenizer=tok,
               aggregation_strategy="simple")

ner("book an appointment with Dr. Patel next Monday at 3pm")
# -> [{'entity_group': 'PROVIDER', 'word': 'dr. patel', ...},
#     {'entity_group': 'DATE', 'word': 'next monday', ...},
#     {'entity_group': 'START_TIME', 'word': '3pm', ...}]

Intended use & limitations

• Designed for short, first-person health-assistant commands in English. Performance on long-form clinical notes or other domains is not guaranteed.
• The model detects spans that may correspond to personal data (FULL_NAME, EMAIL, PHONE, ADDRESS); it does not validate, store, or de-identify them — downstream handling is the integrator's responsibility.
• Not a medical device and not intended for diagnosis or treatment decisions.