Tudorx95/NER_Economic_Political_Spacy

spaCy NER Fine-tuned for Political & Economic Entities

Fine-tuned spaCy en_core_web_trf (RoBERTa-base backbone) on a custom politico-economic NER dataset. Trained to recognize 11 entity types.

Entity types

POLITICIAN, POLITICAL_PARTY, POLITICAL_ORG, FINANCIAL_ORG, ECONOMIC_INDICATOR, POLICY, LEGISLATION, MARKET_EVENT, CURRENCY, TRADE_AGREEMENT, GPE

Performance (Test Set)

spaCy native evaluation:

• Precision: 0.8507
• Recall: 0.7768
• F1: 0.8121

nervaluate ent_type (comparabil cu GLiNER):

• Precision: 0.8633
• Recall: 0.7891
• F1: 0.8245

Per label (ent_type):

Label	Precision	Recall	F1
POLITICIAN	0.813	0.685	0.743
POLITICAL_PARTY	0.914	0.934	0.924
POLITICAL_ORG	0.779	0.567	0.656
FINANCIAL_ORG	0.886	0.750	0.813
ECONOMIC_INDICATOR	0.571	0.800	0.667
POLICY	0.750	0.750	0.750
LEGISLATION	1.000	0.800	0.889
MARKET_EVENT	0.968	0.968	0.968
CURRENCY	0.630	0.567	0.596
TRADE_AGREEMENT	0.684	0.929	0.788
GPE	0.879	0.824	0.851

Usage

import spacy

nlp = spacy.load("path/to/model")
# sau dupa descarcare de pe HuggingFace:
# from huggingface_hub import snapshot_download
# nlp = spacy.load(snapshot_download("Tudorx95/NER_Economic_Political_Spacy"))

doc = nlp("The Federal Reserve raised rates after President Biden signed the new bill.")
for ent in doc.ents:
    print(ent.text, "->", ent.label_)

Training Details

• Framework: spaCy 3.8.14
• Backbone: en_core_web_trf (RoBERTa-base)
• Strategy: Frozen transformer + NER head fine-tuning
• Train examples: 5747
• Dev examples: 1228
• Test examples: 2124
• Best dev F1: 0.8374