wage-ner-v2

wage-ner-v2 is a fine-tuned token classification model specializing in Named Entity Recognition (NER) to extract minimum and maximum compensation ranges from text. It is built on top of a deberta-v3-base model.

Basic Usage

For extracting specific pay ranges, use the ner pipeline. Using aggregation_strategy="simple" is highly recommended to merge multi-token numerical values cleanly.

from transformers import pipeline, AutoModelForTokenClassification, AutoTokenizer

model_name = "loyoladatamining/wage-ner-v2"
model = AutoModelForTokenClassification.from_pretrained(
    model_name, 
    id2label={0: 'O', 1: 'B-MIN', 2: 'B-MAX'}, 
    label2id={'O': 0, 'B-MIN': 1, 'B-MAX': 2}
)
tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length=128)

# Create the NER pipeline
nlp = pipeline(
    "ner", 
    model=model, 
    tokenizer=tokenizer, 
    aggregation_strategy="simple"
)

# Inference
text = "The pay range is $25.00 to $35.00 per hour depending on experience."
results = nlp(text)
print(results)

Output Format

When using aggregation_strategy="simple", the pipeline outputs a list of dictionaries providing the detected pay scales:

[
  {
    "entity_group": "MIN",
    "score": 0.9942,
    "word": "$25.00",
    "start": 17,
    "end": 23
  },
  {
    "entity_group": "MAX",
    "score": 0.9942,
    "word": "$35.00",
    "start": 27,
    "end": 33
  }
]

Key Fields

• entity_group: MIN (the minimum threshold of the wage statement) or MAX (the maximum range of the wage statement, if available).
• word: The extracted text segment indicating the specific pay amount.

Citation

If you find wage-ner-v2 useful in your work, please consider citing:

@article{meisenbacher2025extracting,
  title={Extracting O* NET Features from the NLx Corpus to Build Public Use Aggregate Labor Market Data},
  author={Meisenbacher, Stephen and Nestorov, Svetlozar and Norlander, Peter},
  year={2025}
}