cornelius/lobbyist-classifier-de

Lobbyist classifier (German)

Binary sequence classifier fine-tuned to predict whether a LinkedIn-style job position (title + employer + description) corresponds to a lobbyist (1) or not (0). Trained for the project "Who Becomes a Lobbyist?" (MINISTERIALLOBBY) on Revelio/LinkedIn position text, with labels from the German Bundestag lobby register (DE) or LobbyView (US).

• Base model: distilbert-base-german-cased
• Task: Sequence classification (2 labels: non-lobbyist, lobbyist)
• Max length: 256 tokens

Evaluation (5-fold CV)

• Mean F1: 0.8455 (± 0.0035)
• Fold F1 scores: [0.8467809952206916, 0.8434272955623779, 0.8514680483592401, 0.8410428931875525, 0.8445796460176991]
• Training samples: 17824 (positive: 8912)

Intended use

• Research: classify past or current job positions as lobby vs non-lobby for career-path and panel analyses.
• Not for commercial use without checking compliance with LinkedIn/Revelio terms.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo_id = "cornelius/lobbyist-classifier-de"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSequenceClassification.from_pretrained(repo_id)

def predict(texts, threshold=0.95):
    inp = tokenizer(texts, truncation=True, max_length=256, padding="max_length", return_tensors="pt")
    with torch.no_grad():
        logits = model(**inp).logits
    probs = torch.softmax(logits, dim=1)
    return probs[:, 1].numpy()  # prob lobbyist

# Single position: title + " " + company + " " + description
text = "Senior Public Affairs Manager  Acme Corp  Government relations and advocacy."
prob = predict([text])[0]
print(f"P(lobbyist) = {prob:.2f}")

Citation

If you use this model, please cite the paper "Who Becomes a Lobbyist? Comparative Evidence from the US and Germany" (MINISTERIALLOBBY project, DFG).