---
license: apache-2.0
library_name: transformers
tags:
- propaganda-detection
- multi-label-classification
- modernbert
- text-classification
datasets:
- synapti/nci-propaganda-v5
base_model: answerdotai/ModernBERT-base
language:
- en
metrics:
- f1
pipeline_tag: text-classification
---

# NCI Technique Classifier v5.2

Multi-label propaganda technique classifier based on ModernBERT, trained to identify 18 propaganda techniques from the SemEval-2020 Task 11 taxonomy.

## Model Description

This model is part of the NCI (Narrative Coordination Index) Protocol for detecting coordinated influence operations. It classifies text into 18 propaganda techniques with well-calibrated probability outputs.

### Key Improvements in v5.2

- **Reduced False Positives**: Scientific/factual content false positive rate reduced from 35% (v4) to 8.8%
- **Better Calibration**: ASL loss with clip=0.02 provides more discriminative probability outputs
- **Hard Negatives Training**: Trained on v5 dataset with 1000+ hard negative examples (scientific, business, factual content)
- **Document-Level Analysis**: Works well with full documents, no need for sentence-level splitting

### Training Details

- **Base Model**: `answerdotai/ModernBERT-base`
- **Dataset**: `synapti/nci-propaganda-v5` (24,037 samples)
- **Loss Function**: Asymmetric Loss (ASL)
  - gamma_neg: 4.0
  - gamma_pos: 1.0
  - clip: 0.02 (reduced from 0.05 to minimize probability shifting)
- **Training**: 3 epochs, lr=2e-5, batch_size=16
- **Validation**: 4/7 tests passed (57%)

## Techniques Detected

| ID | Technique | Description |
|----|-----------|-------------|
| 0 | Loaded_Language | Words with strong emotional implications |
| 1 | Appeal_to_fear-prejudice | Building support through fear or prejudice |
| 2 | Exaggeration,Minimisation | Overstating or understating facts |
| 3 | Repetition | Repeating messages for reinforcement |
| 4 | Flag-Waving | Appealing to patriotism/national identity |
| 5 | Name_Calling,Labeling | Using labels to evoke prejudice |
| 6 | Reductio_ad_hitlerum | Comparing to Hitler/Nazis |
| 7 | Black-and-White_Fallacy | Presenting only two choices |
| 8 | Causal_Oversimplification | Assuming single cause for complex issues |
| 9 | Whataboutism,Straw_Men,Red_Herring | Deflection techniques |
| 10 | Straw_Man | Misrepresenting opponent's position |
| 11 | Red_Herring | Introducing irrelevant topics |
| 12 | Doubt | Questioning credibility |
| 13 | Appeal_to_Authority | Using authority figures to support claims |
| 14 | Thought-terminating_Cliches | Phrases that end rational thought |
| 15 | Bandwagon | "Everyone is doing it" appeals |
| 16 | Slogans | Catchy phrases for memorability |
| 17 | Obfuscation,Intentional_Vagueness,Confusion | Deliberately confusing language |

## Usage

```python
from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "synapti/nci-technique-classifier-v5.2"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

text = "This is OUTRAGEOUS! They are LYING to you. WAKE UP!"

inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    outputs = model(**inputs)
    probs = torch.sigmoid(outputs.logits)[0]

# Get techniques with probability > 0.5
LABELS = [
    "Loaded_Language", "Appeal_to_fear-prejudice", "Exaggeration,Minimisation",
    "Repetition", "Flag-Waving", "Name_Calling,Labeling", "Reductio_ad_hitlerum",
    "Black-and-White_Fallacy", "Causal_Oversimplification",
    "Whataboutism,Straw_Men,Red_Herring", "Straw_Man", "Red_Herring", "Doubt",
    "Appeal_to_Authority", "Thought-terminating_Cliches", "Bandwagon", "Slogans",
    "Obfuscation,Intentional_Vagueness,Confusion"
]

for i, (label, prob) in enumerate(zip(LABELS, probs)):
    if prob > 0.5:
        print(f"{label}: {prob:.1%}")
```

## Performance

### Validation Results

| Test Case | v5.2 | v4 | Status |
|-----------|------|-----|--------|
| Pure Propaganda | 66.8% | 70.8% | ✓ Detected |
| Neutral News | 6.9% | 5.5% | ✓ Clean |
| SpaceX Factual | 3.7% | - | ✓ Clean |
| Multi-Label Propaganda | 76.5% | - | ✓ Detected |
| Mixed Content | 7.3% | - | - |
| Fear Appeal | 69.9% | - | ✓ Detected |
| Scientific Report | **8.8%** | 35.4% | ✓ Clean |

### Key Metrics

- **Scientific Report FPR**: 8.8% (vs 35% in v4) - 75% reduction
- **Factual News FPR**: 4.6% (vs 29% in v4) - 84% reduction
- **Propaganda Detection**: Maintained (73.7% max confidence on propaganda)

## Citation

```bibtex
@inproceedings{da-san-martino-etal-2020-semeval,
    title = "{S}em{E}val-2020 Task 11: Detection of Propaganda Techniques in News Articles",
    author = "Da San Martino, Giovanni and others",
    booktitle = "Proceedings of the 14th International Workshop on Semantic Evaluation",
    year = "2020",
}
```

## License

Apache 2.0