firmNER-v3

firmNER-v3 is a fine-tuned DeBERTa model specializing in Named Entity Recognition (NER) to identify firm (company) names in text documents. It represents an upgraded, higher-accuracy version over the previous v2 models.

Basic Usage

The easiest way to use this model is via the Hugging Face pipeline API. For token classification tasks, using an aggregation strategy is recommended to automatically group sub-word tokens back into full words or entities (see below).

from transformers import pipeline, AutoModelForTokenClassification, AutoTokenizer

# Load model and tokenizer
model_name = "loyoladatamining/firmNER-v3"
model = AutoModelForTokenClassification.from_pretrained(
    model_name, 
    id2label={0: 'O', 1: 'B-ORG', 2: 'I-ORG'}, 
    label2id={'O': 0, 'B-ORG': 1, 'I-ORG': 2}
)
tokenizer = AutoTokenizer.from_pretrained(model_name, model_max_length=1024)

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

# Inference
text = "Orange and Macrosoft announced a new partnership yesterday."
results = nlp(text)
print(results)

Output Format

When using the pipeline with aggregation_strategy="max", the model outputs a list of dictionaries. Each dictionary represents a detected entity (labeled as ORG for organizations/firms) and contains the following structure:

[
  {
    "entity_group": "ORG",
    "score": 0.XXX,
    "word": "Orange",
    "start": 0,
    "end": 6
  },
  {
    "entity_group": "ORG",
    "score": 0.XXX,
    "word": "Macrosoft",
    "start": 11,
    "end": 20
  }
]

Key Fields

• entity_group: The predicted label group (i.e. here, ORG for identified company names).
• score: The model's confidence probability for the predicted entity.
• word: The extracted text string representing the firm name.
• start / end: The character indices indicating the exact position of the text string within the input document.

Citation

If you find firmNER 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}
}