Complete Technical Paper Content — Ai Misinformation Detection
1. Executive Abstract, Problem Formulation & Background
Misinformation Detection using BERT Embeddings and Logistic Regression Overview This project focuses on detecting misinformation using Natural Language Processing (NLP). We build a strong text-based baseline model by leveraging BERT embeddings combined with a Logistic Regression classifier. The goal is to establish a solid baseline for misinformation detection using text data, which can later be extended to a multimodal system incorporating both text and images. Problem Statement Misinformation detection is a challenging task, especially in real-world scenarios where content often includes both textual and visual information. Relying solely on text may not be sufficient, as images can significantly alter the meaning or intent of content. In this project, we explore how effectively a model can perform using only textual data. Approach We follow a standard NLP pipeline: Load and preprocess text data Convert text into dense vector representations using BERT (DistilBERT) Train a classifier (Logistic Regression) Evaluate performance using standard classification metrics This pipeline provides a simple, interpretable, and effective baseline model. Dataset We use the IMDB movie review dataset: Binary classification (positive / negative sentiment) Balanced dataset Subset of 3000 samples used for computational efficiency Although this dataset is not directly misinformation-related, it serves as a reliable benchmark for text classification and model validation. Feature Extraction We use DistilBERT (distilbert-base-uncased) to convert text into embeddings: Each text sample is transformed into a 768-dimensional vector We extract the representation of the [CLS] token as the embedding No fine-tuning is applied (feature-based approach) Model We use Logistic Regression as a baseline classifier due to: Simplicity Interpretability Strong performance with high-quality embeddings Train-Test Split 80% training data 20% testing data Random state fixed for reproducibility Evaluation Metrics We evaluate the model using: Accuracy Precision Recall F1-score Confusion Matrix Results The model achieved an accuracy of approximately 74% on the test set. Confusion Matrix Insights The model performs reasonably well on both classes However, there are still misclassifications, particularly in borderline cases Results Confusion Matrix Performance Accuracy: 74% Results Summary The model achieves solid baseline performance, demonstrating that BERT embeddings are effective for text classification tasks. However, the results indicate that text-only features are not sufficient for capturing more complex patterns related to misinformation detection. This highlights the importance of incorporating additional modalities such as images. How to Run Clone the repository git clone https://github.com/your-username/misinformation-detection-bert.git cd misinformation-detection-bert Install dependencies pip install -r requirements.txt Launch Jupyter Notebook jupyter notebook Open and run the notebook 1-text-baseline-bert.ipynb Limitations Uses only textual data No fine-tuning of BERT (feature extraction only) Dataset not directly related to misinformation Moderate performance (baseline level) Future Work To improve performance, the model can be extended to a multimodal system : Extract image embeddings using CLIP Combine text and image features Train a multimodal classifier Compare performance against the text-only baseline This is expected to significantly improve performance in real-world misinformation detection scenarios. Reproducibility To reproduce this project: Install required libraries: pandas numpy scikit-learn transformers torch Run all notebook cells sequentially The dataset is automatically loaded via URL Conclusion In this project, we developed a text-based misinformation detection system using BERT embeddings and Logistic Regression. The model achieved an accuracy of approximately 74%, demonstrating the effectiveness of modern NLP techniques even with simple classifiers. At the same time, the results highlight the limitations of relying solely on textual data. Overall, this project establishes a strong baseline and provides a solid foundation for future improvements through multimodal approaches. Author IOANNIDIS NIKOLAOS Project developed as part of a personal portfolio for MSc applications in Artificial Intelligence / Data Science.
This project focuses on detecting misinformation using Natural Language Processing (NLP). We build a strong text-based baseline model by leveraging BERT embeddings combined with a Logistic Regression classifier.
The goal is to establish a solid baseline for misinformation detection using text data, which can later be extended to a multimodal system incorporating both text and images.
2. Related Work, Literature Survey & Algorithmic Foundations
Misinformation detection is a challenging task, especially in real-world scenarios where content often includes both textual and visual information.
Relying solely on text may not be sufficient, as images can significantly alter the meaning or intent of content.
In this project, we explore how effectively a model can perform using only textual data.
3. Dataset Ingestion, Preprocessing & Stochastic Augmentations
Although this dataset is not directly misinformation-related, it serves as a reliable benchmark for text classification and model validation.
The model achieves solid baseline performance, demonstrating that BERT embeddings are effective for text classification tasks.
However, the results indicate that text-only features are not sufficient for capturing more complex patterns related to misinformation detection.
4. Neural Network Backbone, Layer Specifications & Loss Formulations
git clone https://github.com/your-username/misinformation-detection-bert.git cd misinformation-detection-bert
This is expected to significantly improve performance in real-world misinformation detection scenarios.
In this project, we developed a text-based misinformation detection system using BERT embeddings and Logistic Regression.
5. Training Protocol, Convergence Behavior & Experimental Evaluation
The model achieved an accuracy of approximately 74%, demonstrating the effectiveness of modern NLP techniques even with simple classifiers.
At the same time, the results highlight the limitations of relying solely on textual data.
Overall, this project establishes a strong baseline and provides a solid foundation for future improvements through multimodal approaches.
Model Figures & Evaluation Curves (4 Images)
Deep Learning Execution Pipeline
(Preprocessing)
(Feature Extraction)
(Self-Attention)
(Loss Optimization)