Can we Debunk Disinformation by Leveraging SpeakerCredibility and Perplexity Measures?

In the present age, fighting disinformation is the main concern after pan-demic. The exponential growth of fake news and its role in deteriorating generalpublic trust and democratic standards certainly calls for counter-combat approaches.The prediction of chances of news to be fake is deemed to be a hard task since mostdeceptive news has its roots in trustworthy news. Influential fake news can be createdwith a minor fabrication in legitimate news that can be used for political, entertain-ment or business-related gains. This work provides an approach to segregate newsinto real and fake exploiting data from multiple sources. Bidirectional Encoder Rep-resentations from Transformers (BERT) is deployed to develop deep learning modelsin order to capture the contextual information present in the data. The proposed modelalso incorporate the speaker profile and the associated credibility in order to improve the performance based on a hybrid sequence encoding model. The model outper-formed the previous state-of-the-art benchmark of the LIAR dataset. This attests therole of the speaker profile and credibility in predicting the validity of news. Addition-ally, leveraging perplexity scores based on BERT and GPT-2 embeddings can help todetermine the legitimacy of the source of fake news on the NELA-GT dataset.