IP Library › Granted Patent US 10,936,863
Granted Patent B2
US 10,936,863 · App. 16/500,102 · Granted Mar 2, 2021

Systems and methods for neuronal visual-linguistic data retrieval from an imaged document

Inventors: Amos Simantov (Hod Hasharon, IL); Roy Shilkrot (Kibbutz Negba, IL); Nimrod Morag (Hod Hasharon, IL); Rinon Gal (Tel-Aviv, IL)
Assignee: WAY2VAT LTD.
G06K9/00449G06N3/08G06N20/00G06Q30/04
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 10,936,863
App. No.
16/500,102
Granted
Mar 2, 2021
Kind
B2
Abstract

Systems and methods for automatic information retrieval from imaged documents. Deep network architectures retrieve information from imaged documents using a neuronal visual-linguistic mechanism including a geometrically trained neuronal network. An expense management platform uses the neuronal visual-linguistic mechanism to determine geometric-semantic information of the imaged document.

Claims (26)

1. A method for use in an expense management platform operable to perform content analysis of an imaged invoice document comprising at least one invoice, in an improved manner,

said expense management platform comprising:

an automatic invoice analyzer (AIA) comprising a dedicated invoice optical character recognition (OCR) engine, said automatic invoice analyzer (AIA) operable to perform automated analysis of the at least one invoice;

a machine learning engine comprising a knowledge repository; and

a neuronal visual-linguistic mechanism trained to perform visual-linguistic semantic analysis, wherein said neuronal visual-linguistic mechanism comprises a geometrically trained neuronal network,

said method comprising the steps of:

receiving, via a communication interface, the imaged invoice document;

pre-processing, by said automatic invoice analyzer (AIA), said at least one invoice;

extracting, by said dedicated invoice optical character recognition (OCR) engine, a set of OCR results associated with said at least one invoice;

generating, by said automatic invoice analyzer (AIA), an OCR-enhanced image of the at least one invoice;

applying, by said automatic invoice analyzer (AIA), linguistic semantic analysis geometrically to determine a semantic information of at least one element of the at least one invoice; and

producing, by said automatic invoice analyzer (AIA), one or more analysis results using said neuronal visual-linguistic mechanism;

wherein the step of applying further comprises:

creating an embedding model operable to generate an encoding vector to find an embedding that encodes a pair relationship of a word and a target neighbor;

training a network using the encoding vector; and

creating a working dataset based an initial dataset and further performing augmentation using synthetic image transformations and simulated noise to receive said working dataset.

2. An expense management platform operable to perform automatic data retrieval from an imaged invoice document, said expense management platform comprising:

an automatic invoice analyzer (AIA) operable to perform automated analysis of the imaged invoice document;

a machine learning engine comprising a knowledge repository, said machine learning engine operable to update the knowledge repository with data pertaining to said invoice imaged document; and

a neuronal visual-linguistic mechanism trained to perform visual-linguistic semantic analysis;

wherein said neuronal visual-linguistic mechanism comprises a geometrically trained neuronal network, such that said expense management platform is operable to produce one or more analysis results using said neuronal visual-linguistic mechanism to determine geometric-semantic information;

wherein said visual-linguistic mechanism comprises a full tagging sub-system operable to provide word matching; and

wherein said full tagging sub-system, comprises:

a Char2Vec network operable to generate at least one word-level embedding from an associated character level content and an associated geometric context;

a labeling network operable to generate a heat-map of at least one label probability using a base image overlaid with data of the at least one word-level embedding; and

an inverted-softmax linear classifier configured to determine a best candidate word to match the at least one label probability using said heat-map and a set of hand-crafted features.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Oct 26, 2020
From: SIMANTOV, AMOS; SHILKROT, ROY; MORAG, NIMROD; GAL, RINON
To: WAY2VAT LTD.
Reel/Frame 054162/0500 →
Continuity (3)
Provisional Application 62642686 · Mar 14, 2018
Provisional Application 62585116 · Nov 13, 2017
Related Publication 20200110930A1 · Apr 9, 2020
Cited By (1)
US 12,573,225