Time-aware evidence ranking for fact-checking

Implicit Temporal Reasoning for Evidence-Based Fact-Checking

“…Metadata m is represented as a one-hot vector and encoded by a CNN (filter size = 3, kernel size = 3) with ReLU activation and 1D max pooling. Allein et al (2021), who explicitly modeled temporal relations between a claim and its evidence by constraining model parameters on evidence rankings following various assumptions on temporal relevance. This could be attributed to the expressive power of large pretrained Transformers-based language models and the orders of magnitude of their pretraining set size.…”

“…Zhou et al ( 2020) constructed (entity, value, time)-tuples representing supposedly temporal facts and verified their correctness using probabilistic graphical models. Allein et al (2021) constrained the evidence ranking in fact-checking models on time using silver-standard evidence rankings respecting four assumptions on temporal relevance. Instead of verifying the temporal correctness of claim tuples or explicitly enforcing time-dependent evidence rankings, we let fact-checking models reason implicitly over temporal aspects of claims and evidence in natural language when checking the claims.…”

“…This way, we lay bare the temporal relation between c and e i at the document level. We adopt the approach of Allein et al (2021) and compute the distance in days ∆pub ∈ Z between the publication date of c and that of e i , where ∆pub < 0 indicates that e i was published before c, ∆pub = 0 indicates that e i and c were published on the same day, and ∆pub > 0 indicates that e i was published after c. The publication date of e i is then assigned to a time bucket b pub ∈ T doc given ∆pub. V means that a publication date was found (only for claims) and X that no publication date was found.…”

“…The publication date of an evidence document, however, is not given in the metadata. Since its publication date is often communicated before the ellipsis ('...') at the beginning of its text, we can automatically extract the date from the text (Allein et al, 2021).…”

Preventing profiling for ethical fake news detection

“… Chen, et al (2022) considered both text and image as input and quantified their ambiguity by computing cross-modal correlations. Others moved beyond text and image, and leveraged contextual features on propagation ( Bian, et al, 2020 , Zhou and Zafarani, 2019 ), source ( Yuan, Ma, Zhou, Han, & Hu, 2020 ), and time ( Allein et al, 2021 , Song, Shu, and Wu, 2021 ). Shu, Mahudeswaran, Wang, and Liu (2020) , for example, combined multiple contextual features in multimodal, hierarchical propagation networks using linguistic, structural, and temporal features from micro-level and macro-level propagation networks to detect fake news on Twitter.…”