ritam-m/finbert-company-ner

FinBERT — Company NER (Financial News)

A ProsusAI/finbert model fine-tuned for Named Entity Recognition (NER) to detect company names in financial news articles. Trained on Indian financial news but generalises to broader financial English text.

Model Details

• Base model: ProsusAI/finbert (110M params, BERT pre-trained on financial text)
• Task: Token classification — company name detection using IOB tagging
• Language: English (financial domain)
• Label scheme: O, B-COMPANY, I-COMPANY

Performance

Evaluated on a held-out test set of 167 articles.

Metric	Entity-level	Token-level
Precision	0.43	0.67
Recall	0.58	0.88
F1	0.49	0.76

Confusion matrix (token level):

	Pred O	Pred COMPANY
True O	76,046	948
True COMPANY	262	1,910

This model has high recall (0.88) — it finds the majority of company tokens but fires more false positives than the bert-base-NER variant. Prefer this model when missing a company is more costly than an occasional false alarm.

Training

• Dataset: 1,664 Indian financial news articles from MarketAux API, ground-truthed using MarketAux equity entity tags
• Split: 1,331 train / 166 val / 167 test
• Epochs: 4
• Batch size: 16
• Learning rate: 2e-5
• Weight decay: 0.01
• Hardware: NVIDIA RTX 3070 — ~103 seconds

Usage

from transformers import pipeline

nlp = pipeline(
    "ner",
    model="ritam-m/finbert-company-ner",
    aggregation_strategy="first",
)

text = "HDFC Bank reported strong Q1 earnings, while Infosys maintained its FY27 guidance."
results = nlp(text)

for entity in results:
    if entity["entity_group"] == "COMPANY":
        print(entity["word"], entity["score"])

Limitations

• Ground truth derived from MarketAux API entity tags — not manually verified
• Trained primarily on Indian financial news; performance may vary on other markets
• Entity-level F1 of 0.49 reflects a first training run on ~1,300 examples; more data will improve results
• Compare with ritam-m/bert-base-company-ner which trades recall for precision