IMPLEMENTATION OF THE TEXT SIMILARITY EVALUATION METHOD WITH A WEIGHTED APPROACH TO IMPROVE THE ACCURACY OF FAKE NEWS CLASSIFICATION

Abstract

The article examines how to enhance automated fake news detection accuracy using a novel semantic-aware text-similarity evaluation method. Traditional frequency-based models often fail to capture subtle semantic manipulations in the modern context of digital disinformation and hybrid warfare. This study addresses this gap by combining GloVe word embeddings with a specialized weighting approach. The goal of this study is to develop a highly effective classification methodology that integrates pre-trained GloVe embeddings with a norm-based weighted similarity metric to improve the identification accuracy of fake news. The achievement of this goal is quantitatively validated through measurable improvements in Accuracy, F1-score, Precision, Recall, and Log-loss compared to standard unweighted baselines. The tasks to be addressed include formulating a robust mathematical method to calculate word importance based on the L2-norm of semantic vectors, applying Manhattan distance for nuanced similarity scoring, and seamlessly integrating this feature into a comprehensive machine learning pipeline to ensure flexibility across various algorithms. The methods used in this study encompass extensive data preprocessing on a balanced dataset of 44,898 news records. Text vectorization is performed using 50-dimensional GloVe embeddings. Then, the semantic features are used to train and evaluate four supervised classifiers: Logistic Regression, Random Forest, XGBoost, and Support Vector Classifier (SVC) across incremental dataset sizes and under simulated noise conditions. Conclusions. The empirical results demonstrate that the proposed weighted similarity approach significantly improves fake news classification. Specifically, the method yields an average increase of 3–4% in Accuracy and up to a 4% improvement in F1-scores across all evaluated models. Furthermore, the proposed approach significantly reduces the impact of linguistic noise and maintains strong probabilistic confidence. The scientific novelty lies in replacing uniform semantic pooling with an embedding-norm-based weighting mechanism, offering a computationally efficient and highly accurate solution for real-world disinformation detec