FarFetched: An Entity-centric Approach for Reasoning on Textually Represented Environments

Anonymous

Abstract: We address the problem of automatically acquiring knowledge from news articles and leverage it to estimate the veracity of a user's claim based on the supporting or refuting content within the accumulated evidence. We present FarFetched, an entity-centric approach for reasoning based on news, where latent connections between events, actions or statements are discovered via their identified entity mentions and are represented with the help of a knowledge graph. We propose a way of selecting specific subsets from the accumulated wealth of information based on the user hypothesis and construct relevant premises relying on the semantic similarity between them. We leverage textual entailment recognition to provide a measurable way for assessing whether the user claim is plausible based on the selected evidence. Our work is demonstrated on the less-resourced Greek language and supported by the training of state-of-the-art models for STS and NLI that are evaluated on benchmark datasets.