CARES: A commonsense knowledge-enriched and graph-based contextual learning approach for rumor detection on social media

With the popularity of social media, rumors have become more prevalent than ever. The impact of rumors is adverse, making their detection a growing concern and an indispensable requirement. It necessitates a refined rumor detection approach that ensures credible information to users and maintains a trustworthy social media ecosystem. This paper presents a rumor detection approach, CARES, as a fusion of contextual learning-based rumor-relevant information, commonsense-based background knowledge, and embedding models. It utilizes the knowledge associated with the social media posts and reactions to leverage the textual and latent information and capture their inherent semantic affinity, the dynamics of social interaction such as responsive and reactive behavior of social media users, and commonsense-based background knowledge. Considering that social media entities are intertwined and rumors correlate with topics and contextual entities, CARES identifies and employs them to encode underlying rumor-based patterns and provide commonsense-based understanding. It uses a graph-based representation for contextual learning that recognizes two prevalent categories of words that form a building block of rumor-based patterns. The utilization of external lexicon-based knowledge in pattern selection leverages the reactive behavior of social media users in the expression of emotion, sentiment, skepticism and inquisitiveness, which makes the patterns more enriched and generic towards rumors. These patterns are finally ranked, and only the top-k check-worthy patterns are used for detecting rumors. CARES leverages commonsense understanding using ConceptNet, which provides facts as semantic evidence and enriches the posts with structural and background knowledge surrounding the contextual utterances and related topics. In order to preserve the semantic relations and commonsense understanding, it uses two pre-trained embedding models, GloVe and ConceptNet Numberbatch. The proposed approach is evaluated over three publicly available X (formerly known as Twitter) datasets. The experimental results are promising, and they remarkably outperform the nine baselines and state-of-the-art approaches, demonstrating substantial effectiveness for detecting rumors on social media. •CARES: Rumor detection using contextual learning & intuitive knowledge enrichment.•Integrates word embedding, prior knowledge and contextualized rumor-based patterns.•Generates rich structural descriptors for r