README.md

metadata

license: apache-2.0
datasets:
  - fake_news_english
language:
  - en
metrics:
  - roc_auc
pipeline_tag: text-classification

Pretrained Language Model for Fake News Detection

This repository contains a pretrained language model for fake news detection. The model was developed using PyTorch and the Hugging Face Transformers library, and was fine-tuned on a dataset of news articles to classify each article as either "fake : 1" or "Satire : 0".

Usage

To use the pretrained model for fake news detection, you can follow these steps:

1- Install the required dependencies, including PyTorch, Transformers, and scikit-learn.

2- Load the pretrained model using the from_pretrained() method in the Transformers library.

3- Tokenize your input text using the tokenizer.encode_plus() method.

4- Pass the tokenized input to the model's forward() method to get a prediction.

Here's an example code snippet that demonstrates how to use the model:

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load the pretrained model and tokenizer
model_name = "Karim-Gamal/Roberta_finetuned_fake_news_english.pt"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# Tokenize the input text
input_text = "This is a fake news article"
inputs = tokenizer.encode_plus(input_text, padding=True, truncation=True, max_length=512, return_tensors="pt")

# Get the model's prediction
outputs = model(inputs["input_ids"], attention_mask=inputs["attention_mask"])
predictions = torch.softmax(outputs.logits, dim=1).detach().numpy()

Performance

I have conducted eight experiments using eight different models in various settings to determine the optimal performance, and the outcomes are as follows.

The model was evaluated on a test set of news articles, and achieved an AUC score of 95%. This indicates that the model is able to effectively distinguish between fake and real news articles.

Model name	Roc Auc
jy46604790/Fake-News-Bert-Detect	88 %
ghanashyamvtatti/roberta-fake-news	95 %
gpt2 / reward_model	82 %
gpt2 / imdb-sentiment-classifier	79 %
microsoft/Multilingual-MiniLM-L12-H384	86 %
hamzab/roberta-fake-news-classification	84 %
mainuliitkgp/ROBERTa_fake_news_classification	86 %
ghanashyamvtatti/roberta-fake-news after cleaning	82 %

Based on the provided results, it seems like the ghanashyamvtatti/roberta-fake-news model performed the best with a ROC AUC of 95%. This model was specifically designed for detecting fake news, which explains its high performance on this task.

Additionally, the microsoft/Multilingual-MiniLM-L12-H384 model had a respectable performance with a ROC AUC of 86% while being a lightweight model. Therefore, it was used in our paper which is ( Federated Learning Based Multilingual Emoji prediction ) despite having slightly lower performance than other models.

On the other hand, the GPT2 models (gpt2/reward_model and gpt2/imdb-sentiment-classifier) had lower performance compared to the other models. This may be due to the fact that GPT2 models were pre-trained on different tasks and not specifically designed for fake news detection.

It is worth noting that even though the ghanashyamvtatti/roberta-fake-news after cleaning model had a lower performance (82%) than the original ghanashyamvtatti/roberta-fake-news model (95%), it might still be useful in certain scenarios, especially after cleaning the data.

Finally, it is important to test the performance of the selected model after loading it from Huggingface to make sure it is functioning properly in the desired environment.

Model Card

For more information about the model's architecture, The original model link