BERT-CA Sentinels: Adapting BERT with Attention Mechanism for Satirical News Detection in Social Media Texts
Rakhi Seth, Aakanksha Sharaff · 2024
The exponential proliferation of misinformation, particularly on social media platforms, has emerged as a formidable issue with far-reaching adverse consequences on a worldwide level. Satirical articles, unlike false news, are not meant to deceive or influence the reader. Instead, their purpose is to entertain by ridiculing or critiquing a prominent figure. The major challenge for detecting satire from fake news due to intentional use of subtle language to avoid the fake news detectors. This subtle language consists of casual approach and vagueness in language. To solve above issues, this proposed approach utilized the integrated technique that captures the statistical relationship between words in the text using conditional probability alongside BERT attention mechanism to generate a conditional attention (CA) to improve the contextual comprehension in satire detection task. In next stage, for enhancing lexicon-sentiment and to extract opinion expression with sequential patterns as sentinels this approach is using Bi-LSTM with Conditional Random Field (CRF) by computing simulating labels and transition scores. To determine the best overall CRF scores utilizing Viterbi Algorithm that makes this approach more accurate. To evade vagueness, next this proposed method uses Bi-Normal Separation (BNS) feature scaling in classification space by determining the strong and weak correlation with positive and negative class to yield the high performance. This methodology entails FAKENEWSDATA as dataset, which is then subjected to preprocessing techniques to eliminate noise and achieve normalization. The proposed approach outperforms as compared to other state-of-art algorithms like SVM, LR, DT, and KNN with 98.68% in terms of accuracy.