Paragraph Level Embedding for Detecting Persuasive Contents of Electronic News With Native Ads

Persuasive content in online media, including native ads, significantly influences political attitudes, viewpoints, and financial interests. To differentiate between native ads and authentic news, a system is necessary for readers to distinguish between these two content types quickly. This study proposes using various paragraph-level embeddings to identify which paragraphs contain persuasive content in native ads. The paragraphs are separated, manually labelled, and processed using embedding techniques. State-of-the-art word embedding models, including GloVe, FastText, BERT, and RoBERTa, are selected from static and contextual embeddings. The study compares two machine learning (ML) models (SVM and Random Forest) with two deep learning (DL) models (BiLSTM and CNN) using these four embeddings. The findings demonstrate that BERT-Bi-LSTM attains the highest accuracy and precision at 95%, albeit with a slight risk of overfitting. Furthermore, static embeddings improved performance when combined with machine learning classification models. Contextual embeddings yielded superior performance in deep learning classification models.