heavyhelium/electra-base-logic-fallacy

ELECTRA-base fine-tuned for logical fallacy classification

13-way classifier fine-tuned on the LOGIC dataset from Jin et al. 2022, Logical Fallacy Detection (arXiv:2202.13758).

Base model: google/electra-base-discriminator.

Labels

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

Training

• Data: LOGIC train split, 1849 examples / 13 classes (zhijin/zhijingjin splits; dev 300, test 300).
• Hyperparams: 3 epochs, lr 5e-5, weight decay 0.01, warmup ratio 0.1, batch 8, max_len 128, seed 42, best checkpoint by val macro-F1.
• Hardware: CPU (4 threads), ~23 min wall.

Evaluation

In-domain (LOGIC test, n=300)

metric	value
accuracy	0.643
macro-F1	0.552
weighted-F1	0.625

Comparable to the paper's plain-ELECTRA baseline (~0.533 F1 in Table 3).

Zero-shot transfer (LOGICCLIMATE, n=1312)

metric	value
accuracy	0.210
macro-F1	0.183

Sharp drop on out-of-domain transfer, in line with the paper's Table 4 findings (their best model drops from 0.588 to 0.272 F1).

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

repo = "heavyhelium/electra-base-logic-fallacy"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForSequenceClassification.from_pretrained(repo).eval()

text = "Everyone I know drives a Toyota, so Toyotas must be the best cars."
enc = tok(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    pred_id = model(**enc).logits.argmax(-1).item()
print(model.config.id2label[pred_id])  # -> ad populum

Limitations

• Poor cross-domain generalization. Drops ~0.37 macro-F1 from educational text (LOGIC) to climate-change news (LOGICCLIMATE). Do not trust predictions far from the training domain.
• Data imbalance bias. Rare classes (equivocation, ~2% of training data) are under-predicted; equivocation test F1 is 0.00 in both in-domain and transfer settings.
• Short-text bias. Training examples are mostly 1-2 sentence educational quiz items (median ~100 characters). Longer or structurally different text may degrade.
• Single-label. Each input is forced into exactly one of 13 classes; real-world text often exhibits multiple fallacies or none.

Citation

@inproceedings{jin-etal-2022-logical,
    title = "Logical Fallacy Detection",
    author = "Jin, Zhijing and Lalwani, Abhinav and Vaidhya, Tejas and Shen, Xiaoyu and Ding, Yiwen and Lyu, Zhiheng and Sachan, Mrinmaya and Mihalcea, Rada and Sch{\"o}lkopf, Bernhard",
    booktitle = "Findings of EMNLP 2022",
    year = "2022",
    url = "https://arxiv.org/abs/2202.13758",
}