In general, a fact verification attempts to obtain supported evidence from text in order to verify claims. The labels can contain "supported", "refuted", or "not enough info" to classify a claim. In many respects, this is a natural language interpretation process of entailments. Some methods in this process may incorporate evidence concatenation or individual evidence-claim pairs. Unfortunately, such methods are limited in sufficiently identifying relational and logical attributes of information from the evidence. In order to integrate and reason over several evidences, one has to utilize a graph network for aggregation and reasoning to enable a connected evidence graph with a means of identifying information propagation. A deep workflow process using deep learning with graphs is one approach. The first step in the process is to use a sentence encoder with Bert. The second step is to combine evidence reasoning with aggregation in a modified graph attention network. DAGs can further be utilized for relation and event extraction representations and linkage.
