Exploiting Textual Information for Fake News Detection

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Machine Learning

Background:

• The deliberate dissemination of deceptive news, or "fake news," poses a significant challenge to public trust and information integrity.
• Existing style-based techniques for fake news detection rely on textual features but can be enhanced for greater accuracy and insight.
• Machine learning (ML) and deep learning models offer powerful tools for analyzing textual data and identifying patterns indicative of misinformation.

Purpose of the Study:

• To assess the accuracy of various ML algorithms in detecting fake news using a style-based approach.
• To propose and evaluate an enhanced linguistic feature set combining Named Entity Recognition (NER) and Frequent Pattern (FP) Growth.
• To compare the performance of ML algorithms with Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks.

Main Methods:

• Extracted linguistic features, including part-of-speech counts and enhanced features from NER and FP Growth.
• Constructed document embeddings using pre-trained word embeddings and TF-IDF weighting.
• Combined document embeddings with linguistic features to create diverse training/test sets.
• Applied recursive feature elimination to identify optimal linguistic characteristics.
• Trained and fine-tuned ML algorithms, CNNs, and LSTMs for fake news classification.

Main Results:

• Convolutional Neural Networks (CNNs) demonstrated superior accuracy when utilizing pre-trained word embeddings.
• Support Vector Machines (SVMs) achieved comparable accuracy across a broader range of input feature sets.
• The enhanced linguistic feature set improved the insight into sentence-level structure.
• Style-based techniques, while yielding lower accuracy, offered explainable insights into authorial writing style.

Conclusions:

• Combining advanced techniques like NER and FP Growth enhances the effectiveness of style-based fake news detection.
• Deep learning models, particularly CNNs with appropriate embeddings, show strong potential for accurate fake news identification.
• Style-based approaches remain valuable for providing interpretable results regarding writing style in misinformation.