Recognizing fake news based on natural language processing using the BM25 algorithm with fine-tuned parameters

Abstract

The object of the research is the method of natural language processing (NLP) with balanced parameters of the BestMatch25 (ВМ25) algorithm to recognize and classify fake news based on natural language processing (NLP). The unsatisfactory accuracy and speed of existing methods for detecting fake news in unstructured input data demanded the development of a new approach for their effective detection. The study investigated the BM25 algorithm, methods for selecting parameters k1 and b, and their impact on the algorithm's effectiveness in detecting fake news. It was established that precise and detailed adjustment of these parameters is crucial in achieving optimal accuracy and data processing speed. The results showed that the successful selection of BM25 parameters improves the model's accuracy by up to 14 % compared to standard term frequency – inverse document frequency (TF-IDF) calculations. These results were made possible by experimentally tuning different combinations of k1 and b parameters, in which the algorithm shows the best speed indicator or the most accurate estimate of the importance of a term in a document. Balanced values of k1 and b parameters were identified, leading to the algorithm's optimal speed and accuracy in assessing word importance considering the input data's peculiarities. The balanced setting of the BM25 algorithm parameters explains the obtained results. They can be used for automated recognition and analysis of news and information on social media based on natural language processing. However, in practice, the effectiveness of the set of parameters depends on linguistic variations, content, and the theme within new input data sets