CouncilTopics-PT: A multi-label classifier for Portuguese municipal meeting topics.

Model Description

CouncilTopics-PT is a multilabel classification system designed to identify Portuguese municipal topics across 22 distinct categories. The model is an ensemble composed of 12 diverse base learners (using both lexical and contextual representations), whose predictions are integrated by a meta-learner to improve generalization and robustness to linguistic variation in meeting minutes. Try out the model: Hugging Face Space Demo

Key Features

• 🧠 Meta-Learning: Ensemble combination using stacked generalization
• 📚 12 Base Models: 3 feature sets × 4 algorithms for robust predictions
• 🇵🇹 Portuguese Optimized: Prepared for the Portuguese language
• 🏢 22 Categories: Comprehensive municipal administrative document classification
• 🎯 Dynamic Thresholds: Optimized per-category decision boundaries

Model Details

• Architecture: Stacking with Meta-Learning
• Base Models: 12 diverse classifiers (LogReg, Random Forest, Gradient Boosting)
• Feature Engineering: TF-IDF + BERTimbau embeddings + Statistical features
• Meta-Learner: Ensemble combination algorithm
• Categories: 22 Portuguese administrative topic labels
• Training Method: Cross-validation stacking with dynamic threshold optimization

How It Works

The Council topics system operates in multiple stages:

1. Feature Extraction: Three complementary feature sets

   • TF-IDF vectorization (word and character n-grams)
   • BERTimbau embeddings from neuralmind/bert-base-portuguese-cased
   • Statistical text features

2. Base Model Ensemble: 12 diverse classifiers trained on different feature combinations

   • Logistic Regression (C=1.0, C=0.5)
   • Random Forest
   • Gradient Boosting

3. Meta-Learning: Combination of base model predictions using stacking

4. Dynamic Thresholds: Per-category optimized decision boundaries for multilabel output

Usage

Quick Start with Python

import joblib
import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from scipy.sparse import hstack, csr_matrix

# Load the model components
tfidf_vectorizer = joblib.load("int_stacking_tfidf_vectorizer.joblib")
meta_learner = joblib.load("int_stacking_meta_learner.joblib")
mlb_encoder = joblib.load("int_stacking_mlb_encoder.joblib")
base_models = joblib.load("int_stacking_base_models.joblib")
optimal_thresholds = np.load("int_stacking_optimal_thresholds.npy")

# Prepare text
text = """CONTRATO DE PRESTAÇÃO DE SERVIÇOS
Entre a Administração Pública Municipal e a empresa contratada, 
fica estabelecido o presente contrato para prestação de serviços 
de manutenção e conservação de vias públicas."""

# Extract features
tfidf_features = tfidf_vectorizer.transform([text])

# Generate base model predictions
base_predictions = np.zeros((1, len(mlb_encoder.classes_), 12))
model_idx = 0

for feat_name in ["TF-IDF", "BERT", "TF-IDF+BERT"]:
    for algo_name in ["LogReg_C1", "LogReg_C05", "GradBoost", "RandomForest"]:
        model_key = f"{feat_name}_{algo_name}"
        if model_key in base_models:
            model = base_models[model_key]
            pred = model.predict_proba(tfidf_features)
            base_predictions[0, :, model_idx] = pred[0]
        model_idx += 1

# Meta-learner prediction
meta_features = base_predictions.reshape(1, -1)
meta_pred = meta_learner.predict_proba(meta_features)[0]

# Apply dynamic thresholds
predicted_labels = []
for i, (prob, threshold) in enumerate(zip(meta_pred, optimal_thresholds)):
    if prob > threshold:
        predicted_labels.append({
            "label": mlb_encoder.classes_[i],
            "probability": float(prob),
            "confidence": "high" if prob > 0.7 else "medium" if prob > 0.4 else "low"
        })

# Sort by probability
predicted_labels.sort(key=lambda x: x["probability"], reverse=True)
print("Predicted categories:", predicted_labels)

Categories

The model classifies topics into 22 Portuguese administrative categories:

Category	Portuguese Name
General Administration	Administração Geral, Finanças e Recursos Humanos
Environment	Ambiente
Economic Activities	Atividades Económicas
Social Action	Ação Social
Science	Ciência
Communication	Comunicação e Relações Públicas
External Cooperation	Cooperação Externa e Relações Internacionais
Culture	Cultura
Sports	Desporto
Education	Educação e Formação Profissional
Energy & Telecommunications	Energia e Telecomunicações
Housing	Habitação
Private Construction	Obras Particulares
Public Works	Obras Públicas
Territorial Planning	Ordenamento do Território
Other	Outros
Heritage	Património
Municipal Police	Polícia Municipal
Animal Protection	Proteção Animal
Civil Protection	Proteção Civil
Health	Saúde
Traffic & Transport	Trânsito, Transportes e Comunicações

Evaluation Results

Comprehensive Performance Metrics

Metric	Score	Description
F1-macro	0.5486	Macro-averaged F1 score
F1-micro	0.7379	Micro-averaged F1 score
F1-weighted	0.742	Weighted-averaged F1 score
Accuracy	0.4259	Subset accuracy (exact match)
Hamming Loss	0.0426	Label-wise error rate
Average Precision (macro)	0.608	Macro-averaged AP
Average Precision (micro)	0.785	Micro-averaged AP

Technical Architecture

Base Model Ensemble

• Feature Set 1: TF-IDF (word + character n-grams)
• Feature Set 2: BERTimbau embeddings (768 dimensions)
• Feature Set 3: Combined TF-IDF + BERT features

Algorithms per Feature Set

1. Logistic Regression (C=1.0)
2. Logistic Regression (C=0.5)
3. Gradient Boosting Classifier
4. Random Forest Classifier

Meta-Learning Strategy

• Cross-validation stacking for robust meta-features
• Intelligent combination: 70% meta-learner + 30% simple ensemble
• Dynamic threshold optimization per category using differential evolution

Training Data

The model was trained on a curated dataset of Portuguese municipal council meeting minutes including:

• Administrative contracts and agreements
• Environmental reports and assessments
• Traffic regulations and urban planning documents
• Public health and safety communications
• Cultural and educational program descriptions

Limitations

• Language Specificity: Optimized for Portuguese administrative language
• Domain Focus: Best performance on municipal documents
• Computational Requirements: Requires significant memory for all model components
• Threshold Sensitivity: Performance depends on carefully tuned per-category thresholds
• Class Imbalance: Some categories may have lower precision due to limited training examples

License

This model is released under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License (CC BY-NC-ND 4.0).