olivetti03/lobby-ner-spain-bert-v8-10

Lobby-NER Spain v8.10

NER model for detecting registered lobbying organisations in Spanish news text.

Architecture

• Base: es_dep_news_trf (RoBERTa-BNE, 570GB Spanish pre-training)
• Fine-tuned: class-incremental NER adding LOBBY label
• Framework: spaCy 3.8+ with curated-transformers

Gazetteer (v5, 8,259 terms)

Built from 4 Spanish transparency registers (Catalan, CNMV, Madrid, Valencia).

Removed: 48 person names, ambiguous entries (Fulton, Premier, Calafat, Cogen, Valcarce, Telefonica). Added: UGT, CCOO, Movistar, ANFAC, Banco Sabadell. Converted: ~215 filial names stripped of country suffix. Cleaned: societary suffixes (S.A., S.L., unipersonal).

Training Data

• Silver corpus: 293,239 positive sentences (gazetteer-matched)
• Hard negatives: 24,468 sentences, 100 percent with INE-verified person names
• Source: 870,804 news articles from 19 Spanish outlets (May 2023 - Jan 2025)

Methodology

1. Gazetteer-based silver labelling (PhraseMatcher, case-insensitive, token-level)
2. Scraping noise filter (JS/CSS/URLs/DOIs/templates)
3. ALL-CAPS non-lobby filter (>2 words in caps that are not lobby = promotional)
4. Tag/header filter (Temas, Etiquetas, Branded Content, 2+ consecutive caps)
5. Document-level split 70/15/15 (zero leakage)
6. Two-stage training: NER head 3 iter lr=0.001, full fine-tuning 12 iter lr=5e-5

Evaluation (TEST greedy)

• Precision: 0.9975
• Recall: 0.9964
• F1: 0.9969
• False Positives: 127
• Test sentences: 49,184

Known Limitations

Homographs (2 active)

• Airbus: company vs aircraft model (Airbus A350)
• SAP: software company vs syndrome abbreviation

Coverage gaps

• CaixaBank, Bankinter, Cepsa, PwC (short), ONCE, MasOrange
• Not in the 4 regional registers used

Residual FP

• Person names outside INE (Basque/Catalan rare names) may trigger capitalization bias

Usage

import spacy nlp = spacy.load("olivetti03/lobby-ner-spain-bert-v8-10") doc = nlp("La CEOE y UGT negociaron con el Ministerio.")

Entities: CEOE (LOBBY), UGT (LOBBY)

License: MIT