Isabelbinu/roberta-large-argumentscheme-fallacy-classifier

ArgueBot — Unified Argument Scheme & Fallacy Classifier

A fine-tuned RoBERTa-large model that classifies text into one of 24 categories: - 11 argument scheme types (valid argumentative patterns from Walton's taxonomy) - 13 logical fallacy types (common informal fallacies) The model determines both whether an argument is valid or fallacious and which specific type it is — in a single inference pass.

Model Details

Property	Value
Base model	roberta-large
Task	24-class text classification
Scheme classes	11
Fallacy classes	13
Datasets	EthiX + Macagno (argument schemes), Fallacy dataset (13 types)
Epochs trained	5 (early stopping)
Best val metric	1.8465 (val_loss)

---

Labels

✅ Argument Schemes (valid arguments)

• argument from alternatives
• argument from analogy
• argument from cause to effect
• argument from commitment
• argument from example
• argument from expert opinion
• argument from negative consequences
• argument from positive consequences
• argument from practical reasoning
• argument from sign
• argument from values

⚡ Fallacy Types

• ad hominem
• ad populum
• appeal to emotion
• circular reasoning
• equivocation
• fallacy of credibility
• fallacy of extension
• fallacy of logic
• fallacy of relevance
• false causality
• false dilemma
• faulty generalization
• intentional

---

How to Use

from transformers import RobertaTokenizer, RobertaForSequenceClassification
import torch, json
 
model_id = "your-username/arguebot-argument-fallacy-classifier"
tokenizer = RobertaTokenizer.from_pretrained(model_id)
model     = RobertaForSequenceClassification.from_pretrained(model_id)
model.eval()
 
# Load label metadata
import requests
meta      = requests.get(
    f"https://huggingface.co/{model_id}/resolve/main/metadata.json"
).json()
label_map    = {int(k): v for k, v in meta["label_map"].items()}
scheme_ids   = set(meta["scheme_ids"])
fallacy_ids  = set(meta["fallacy_ids"])
 
def predict(text):
    enc = tokenizer(text, return_tensors="pt",
                    truncation=True, max_length=128)
    with torch.no_grad():
        logits = model(**enc).logits
    probs   = torch.softmax(logits, dim=1).squeeze()
    pred_id = int(probs.argmax())
    label   = label_map[pred_id]
    verdict = "Valid Argument" if pred_id in scheme_ids else "Fallacy"
    return {
        "verdict":    verdict,
        "label":      label,
        "confidence": float(round(probs[pred_id].item(), 4)),
    }
 
print(predict("According to NASA, global temperatures will rise 2°C by 2050."))
# {'verdict': 'Valid Argument', 'label': 'argument from expert opinion', 'confidence': 0.94}
 
print(predict("Don't trust him — he was caught lying before."))
# {'verdict': 'Fallacy', 'label': 'ad hominem', 'confidence': 0.88}

---

Training Details

• Deduplication: exact + near-duplicate removal, min 5 words per sample
• Class balancing: sklearn compute_class_weight("balanced") + weighted cross-entropy loss
• Batch strategy: custom InterleavedSampler alternates scheme/fallacy samples per batch
• Early stopping: patience=3, min_delta=0.001, monitor=val_loss
• Optimiser: AdamW, lr=3e-05, weight_decay=0.01
• Scheduler: linear warmup (15% of steps)

---

Intended Uses

• Debate analysis and argumentation quality assessment
• Educational tools for teaching critical thinking and informal logic
• AI-assisted fact-checking and media literacy tools
• Research in computational argumentation

Limitations

• Trained on English text only
• Short texts (< 5 words) may produce unreliable predictions
• Some fallacy types (e.g. intentional, equivocation) are harder to distinguish without broader context
• Not suitable for legal or medical decision-making

---

Citation

If you use this model in your research, please cite: @misc{arguebot2025, title = {ArgueBot: Unified Argument Scheme and Fallacy Classification}, author = {Isabel}, year = {2025}, url = {https://huggingface.co/your-username/arguebot-argument-fallacy-classifier} }

Built with RoBERTa-large · EthiX + Macagno datasets · Walton's Argumentation Schemes