US 12,718,063 B2 · Deep learning architecture for adverse media screening

Abstract

Systems, methods, and apparatus are provided for deep learning architecture for adverse media screening. A dynamic data stream may be generated from news reports acquired from a variety of media channels in a variety of formats. Reports may be separated by format. A graphing module for each type of media may use geometric clustering algorithms to group reports into a coherence graph. Fact-checking sources may be incorporated into the graph. A graph neural network may determine coherence between news claims and fact-checking sources, between news claims and supporting report details, between similar reports, and/or between video and audio within the report itself. The network may output a legitimacy score for a report. A credible report may be flagged and retrieved in the course of a customer screening. Materiality may be determined for a credible report and may trigger a customer account closure protocol.

A method for using an optoelectronic deep learning architecture to detect fabricated report material regarding a client, the method comprising, using a photonic processor:

generating a dynamic data stream comprising news reports from a plurality of media channels, the data stream comprising text-based media, video-based media, and audio-based media;
separating text-based media from the data stream;
using a geometric clustering algorithm, grouping the text-based media into a first tree comprising a plurality of nodes, an arrangement of the nodes based at least in part on a similarity between node attributes, each node comprising a news report, the attributes for each node comprising title, source, reporter, date of publication, and topic;
at a first graph neural network (GNN), screening a node from the first tree to determine credibility of the news report, the screening comprising:
concatenating features for the node, the features comprising:
word embeddings from text extracted from the node, the text comprising a claim and supporting details associated with the claim;
word embeddings from text extracted from a fact-checking source, the text comprising a claim verified through credentialed reporting;
and
an affinity matrix representing an arrangement of nodes in the first tree having a threshold similarity;
and
inputting the concatenated features to one or more fully connected layers in the first GNN to determine:
a coherence between the claim and the supporting details in the news report;
a coherence between the news report and the fact-checking source;
and
a legitimacy score associated with the news report;
separating video-based media from the data stream;
using the geometric clustering algorithm, grouping the video-based media into a second tree comprising a plurality of nodes, each node comprising a video news report, an arrangement of the nodes based at least in part on a similarity between node attributes;
splitting a video news report into a set of time-based frames;
at a second GNN, screening a node from the second tree to determine credibility of the video news report, the screening comprising:
concatenating features from a frame, the features comprising:
an event extracted from video associated with the frame;
an audio clip extracted from video associated with the frame;
and
an affinity matrix associated with the arrangement of nodes in the second tree;
and
inputting the concatenated features to one or more fully connected layers in the second GNN to determine:
a coherence between the event and the audio clip;
and
a legitimacy score associated with the video news report;
screening output from the first GNN and the second GNN for an adverse news report regarding the client;
determining whether the adverse news report regarding the client is credible and material, based at least in part on comparison of a legitimacy score associated with the adverse news report to a threshold legitimacy score;
and
in response to a determination that the adverse news report regarding the client is credible and material:
flagging the adverse news report and storing the flagged adverse news report in a data repository;
initiating closure of an account associated with the client;
notifying the client of the adverse news report and the closure of the account;
and
closing the account.

35 more elements

Deep learning architecture for adverse media screening

Life of this patent

Deep learning architecture for adverse media screening

Life of this patent

Abstract

Claims17