Mishtar4/persuasion-detector-xlm-roberta

Persuasion Technique Detector — XLM-RoBERTa-large

Fine-tuned XLM-RoBERTa-large for multi-label detection of 23 persuasion techniques at paragraph level.

Model Details

• Base model: xlm-roberta-large
• Task: Multi-label text classification (23 classes)
• Training data: SemEval 2023 Task 3 (~19,900 train / ~6,500 val paragraphs, 6 languages)
• Best micro-F1: ~0.45 on Polish dev set
• Languages: Optimized for Polish, supports 100+ languages

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from huggingface_hub import hf_hub_download
import torch
import numpy as np

model_id = "Mishtar4/persuasion-detector-xlm-roberta"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id).eval()

thresholds = np.load(hf_hub_download(model_id, "thresholds.npy"))
labels = open(hf_hub_download(model_id, "labels.txt")).read().splitlines()

text = "Your paragraph here"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=256)
with torch.no_grad():
    probs = torch.sigmoid(model(**inputs).logits)[0].numpy()

detected = [(labels[i], float(probs[i])) for i in range(len(labels)) if probs[i] >= thresholds[i]]
print(detected)

Classes

23 persuasion techniques from SemEval 2023 Task 3: Appeal_to_Authority, Appeal_to_Popularity, Appeal_to_Values, Appeal_to_Fear-Prejudice, Flag_Waving, Causal_Oversimplification, False_Dilemma-No_Choice, Consequential_Oversimplification, Straw_Man, Red_Herring, Whataboutism, Slogans, Appeal_to_Time, Conversation_Killer, Loaded_Language, Repetition, Exaggeration-Minimisation, Obfuscation-Vagueness-Confusion, Name_Calling-Labeling, Doubt, Guilt_by_Association, Appeal_to_Hypocrisy, Questioning_the_Reputation

Deployment

Used in a bilingual Telegram bot (EN/PL/UA): github.com/Mishtar4/persuasion-bot