sofiagzzloz/political-persuasion-model

Political Persuasion Hybrid Model

This repository contains a runnable hybrid multi-label classifier for rhetorical strategy detection in political speech segments.

It predicts six labels:

• emotion_appeal
• authority_appeal
• polarization
• presumption
• exaggeration
• rhetorical_framing

Model Description

This is a hybrid architecture that combines:

1. TF-IDF features (1-2 grams), and
2. DistilBERT embeddings (distilbert-base-uncased, mean pooled),

then trains one RandomForest classifier per label with label-specific thresholds.

Files in This Repo

• artifacts/tfidf_vectorizer.joblib
• artifacts/classifier_<label>.joblib (one per label)
• artifacts/thresholds.json
• artifacts/metadata.json
• hybrid_results.csv
• results.json
• test_predictions.csv

Intended Use

Use this model for research and analysis of rhetorical patterns in political language.

Limitations

• Built from a specific political speech dataset and taxonomy.
• Performance may not generalize to other domains/languages without re-training.
• Should not be used as a sole basis for high-stakes decisions.

Metrics (latest run)

See:

• results.json (summary + per-label metrics)
• hybrid_results.csv (per-label threshold/F1/precision/recall)

How to Run Inference

Inference CLI lives in the project codebase:

python scripts/infer_hybrid_model.py \
  --artifacts-dir results/hybrid_YYYYMMDD_HHMMSS/artifacts \
  --text "We must act now to protect our families."

CSV batch mode:

python scripts/infer_hybrid_model.py \
  --artifacts-dir results/hybrid_YYYYMMDD_HHMMSS/artifacts \
  --input-csv dataset/dataset_for_annotation.csv \
  --text-column text \
  --output results/hybrid_predictions.csv

Training Data

Related dataset repository:

• sofiagzzloz/political-persuasion-dataset

Citation

If you use this model, please cite your project/report and reference this repository.