Leveraging Contextual Embeddings: A BERT-Based Approach to Mitigating Semantic Ambiguity in Digital News

Abstract

The high rate of growth of the digital false information requires detection systems that are able to address the nuance in the language. The standard paradigms of machine learning (ML), example Support Vector Machines on TF-IDF vectorization, may be characterized by significant weaknesses when dealing with semantic ambiguity since the latter depends on the static frequencies of words, but not dynamic contextual dependencies. In this study, we have suggested a powerful detection system based on Bidirectional Encoder Representations a Transformer (BERT) used to retrieve deep contextual embeddings. The work of the proposed model is to analyze the news articles in both directions by using a multi-head self-attention mechanism and as such, the model is capable of differentiating between authentic and fake news items using complex semantic clues. The comparative analysis shows that the baseline SVM did reach 85 which is high, however, the fine-tuned BERT architecture showed a much higher level of performance with 94 and thus a high level of accuracy to identify long-range dependencies and reduce the ambiguity which is required to maintain the integrity of digital news