Abstract
Given the rising prevalence of disinformation and fake news online, the
detection of fake news in social media posts has become an essential
task in the field of natural language processing (NLP). In this paper,
we propose a fake detection model named, FakEDAMR that encodes textual
content using the Abstract Meaning Representation (AMR) graph, a
semantic representation of natural language that captures the underlying
meaning of a sentence. The graphical representation of textual content
holds longer relation dependency in very few distances. A new fake news
dataset, FauxNSA, has been created using tweets from the Twitter
platform related to ‘Nupur Sharma’ and ‘Agniveer’ political controversy.
We represent each sentence of the tweet using AMR graph and then use
this in combination with textual features to classify fake news.
Experimental results on two different sets of features show that adding
AMR graph features improves F1-score and accuracy. In the experiments,
Random Forest with AMR-encoded features outperforms other models in
Feature-set 1, achieving 88.90%, 89.48%, 87.09% accuracy and 85.92%,
88.69%, 86.70% F1-score on the FauxNSA, Covid19-FND, and KFN datasets,
respectively. However, when Feature-set 2 is used, BiLSTM with
AMR-encoded features emerges as the top-performing model. It achieves
highest accuracy and F1-score of 93.96% and 91.96% on the dataset. It
also maintains high performance on Covid19-FND and KFN datasets, with
accuracy and F1-scores of 93.26% and 93.20% on Covid19-FND, and
93.52% and 93.52% on KFN, respectively.