US 11,947,914 B2
Fact checking based on semantic graphs
Duyu Tang, Beijing (CN); Nan Duan, Beijing (CN); Ming Zhou, Beijing (CN); Jiun-Hung Chen, Redmond, WA (US); Pengcheng Wang, Redmond, WA (US); and Ying Qiao, Redmond, WA (US)
Assigned to Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed by Microsoft Technology Licensing, LLC, Redmond, WA (US)
Filed on Jun. 30, 2020, as Appl. No. 16/916,599.
Prior Publication US 2021/0406475 A1, Dec. 30, 2021
Int. Cl. G06F 40/30 (2020.01); G06F 16/21 (2019.01); G06N 3/08 (2023.01)
CPC G06F 40/30 (2020.01) [G06F 16/212 (2019.01); G06N 3/08 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computer-implemented method, comprising:
obtaining, from an evidence database, a set of evidence sentences associated with a text to be fact checked, the text to be fact checked received as input by a computing device;
constructing, by the computing device, a first graph indicating semantic information of the text and a second graph indicating semantic information of the set of evidence sentences;
inputting the first graph and the second graph into a graph-based pre-trained model to generate text word representations, evidence word representations, and a joint text-evidence representation;
generating, by a graph convolutional network, node representations for the text word representations and the evidence word representations;
determining, by the computing device, a veracity of a statement in the text by evaluating the node representations and the joint text-evidence representation using a graph attention network; and
generating an output for transmission that includes the veracity of the statement.