WO2022191769A1

WO2022191769A1 - Kyc method and kyc platform for correction of raw text without use of text regions

Info

Publication number: WO2022191769A1
Application number: PCT/SG2021/050124
Authority: WO
Inventors: Mary Grace VILLANUEVA; Jun Arvin MANALANG; Joseph Christian NOEL
Original assignee: Voyager Innovations Holdings Pte. Ltd.
Priority date: 2021-03-10
Filing date: 2021-03-10
Publication date: 2022-09-15

Abstract

A KYC method (10) or KYC platform (24) for receiving and processing a plurality of validation requests (11) which include ID images (11B). Validation requests (11) with fraudulent ID images (11B) are red flagged using facial matching. Document types for each ID image (11B). ID images (11B) undergo OCR processing without the use of text regions to produce raw text (12). Raw text (12) from the ID images (11B) is post-edited to create corrected text (14). Post-editing includes use of tailored error correction models (13) associated with the document type of the ID image (11B).

Description

KYC METHOD AND KYC PLATFORM FOR CORRECTION OF RAW TEXT WITHOUT USE OF TEXT REGIONS

TECHNICAL CONTRIBUTION

The present disclosure relates to know your customer review of validation requests. More particularly, the present disclosure relates to a know your customer computer-implemented method or know your customer platform for correction of raw text generated by OCR from images of customer identification documents having printed text.

BACKGROUND

Know your customer or know your client (KYC) is typically a computerized validation process for customer biographical text and identification documents collected from online websites. Customers or staff of a business enter biographical text into the data entry fields of a webpage and upload an image of a form of identification (hereafter, “the ID image”) to create a validation request.

The biographical text typically includes names, identification numbers, dates of birth, addresses, effective dates, and/or addresses. The biographical text can be structured using field restrictions in the data entry fields of the data entry webpage, for instance by use of pop up calendars for the input of dates. For computerized analysis of the ID images, the printed text of the ID images must be captured conducting OCR upon the ID image. The ID images must undergo OCR during the validation process because the biographical data printed on the ID image must be compared against the biographical text submitted with the validation request. The biographical data printed on the ID image may also be compared against other proprietary or public information collected separately from the incoming validation request.

Processing the images of customer identification documents poses various challenges. While each document type may include the same basic biographical data (such as name and date of birth), the template for the placement and presentation of this information varies between each document type. Different document types, for instance, may list the basic biographical data in a different order with different font sizes, and different security features. The readability of the images may also vary due to image resolution, lighting conditions, font size, security features, and damage to the documents themselves.

The formatting and content of identification documents varies between government agencies and may also vary depending on the year of issuance of the identification documents. Philippines identification document types, for instance, can include a driver’s license, Government Service Insurance System (GSID) card, a National Bureau of Investigation (NBI) card, a passport, a Philippine Health Insurance Corporation (PhilHealth) card, a postal identity card, a Professional Regulation Commission (PRC) card, a Social Security System (SSS) card, a Tax Identification Number (TIN) card, a Unified Multi-Purpose ID (UMID) card, and a voter ID card. KYC review is adapted to various types of identification or documentation based on the jurisdiction (e.g., country or state) and/or the business needs (e.g., financial, government, etc).

An overview of how such a KYC method or KYC platform can be constructed is provided in the article entitled “How we built a modern, state of the art OCR pipeline”, AB Saravanan, https://blog.signzy.com/how-we-built-a-modern-state-of-the-art-ocr-pipeline-preciousdory- dc3a4ae0e87, 06 October 2018. In this Saravanan article, the author discusses the partitioning of the total ID image area into a series of cropped images, one cropped image for each of a series of text regions. Each text region is identified as a specific area on the ID image set aside for the printing of biographical data for a specific data field. The specific areas can be set aside, for instance, for the printing of the biographical data into a name data field, an identification number data field, an address data field, or a date of birth data field. Each text region is separately processed by the OCR engine and error correction module. The supervised training data for the error correction model includes paired sets of cropped images of specific text regions (from a larger ID image) and the corresponding corrected text for that cropped image. A similar text region approach is also taken in the published patent application numbered W02020141890 entitled “Method and Apparatus for Document Management” dated 09 July 2020.

The driving business demands for a KYC method or KYC platform are its accuracy, speed, throughput, and cost. If the validation requests are not performed accurately, a fraudulent account could be established. If the validation request are not performed quickly or the queue of pending validation requests backs up, potential customers submitting validation requests could be lost to competitors. If the costs are high, particularly the cost per validation request, running cost may balloon during customer growth periods. Though many software vendors provide turnkey software products or services to perform KYC, the fees for such software or service typically scale with the number of validation requests processed. Especially for financial industries with a large potential customer base, the number of validation requests can be in the hundreds of thousands or more each week, making cost a key concern.

While text regions can be used to further refine and specialize the OCR and error correction process used on each specific text region, use of text regions adds an additional complexity and marries the KYC method or KYC platform to a limited selection of OCR engine choices. What is needed is a KYC architecture that decouples the error correcting step from the OCR conversion step such that a low cost and robust OCR engine can be selected without impacting the error correction step.

SUMMARY

A general embodiment of the invention is a KYC method or KYC platform for receiving and processing a plurality of validation requests which include ID images. Validation requests with fraudulent ID images are red flagged using facial matching. Document types for each ID image are identified without the use of text regions to produce raw text. Raw text from the ID images is post-edited to create corrected text. Post-editing includes use of tailored error correction models associated with the document type of the ID image.

A first embodiment of the invention is a computer- implemented KYC method, the method comprising: (a) receiving at a request database in data communication with a network a plurality of validation requests through a network; (b) storing each validation request in one of a plurality of request records of the request database; (c) facial matching to compare each photograph to each of a plurality of library images to determine a duplication score for the validation request associated with the photograph; (d) identifying a document type of each ID image; (e) conducting OCR to convert the one or more ID images of each validation request to raw text; (f) post-editing the raw text to execute sequence-to- sequence error correction upon the raw text of each ID image of each validation request to create corrected text for each ID image of the validation request; and (g) evaluating each validation request to compare the biographical text delineated into the first plurality of data fields of the validation request against the corrected text delineated into the second plurality of data fields of the validation request to determine a deviation score for the validation request. Each validation request includes biographical text and one or more ID images purportedly identifying a customer associated with the validation request. The biographical text for each validation request is delineated into a first plurality of data fields within the request record for the validation request. At least one of the one or more ID images of each validation request includes a photograph of the customer. The duplication score of each validation request is evaluated according to a duplication threshold. The facial matching step includes red flagging each validation request exceeding the duplication threshold. The conducting OCR step does not identify a plurality of text regions in each ID image. The raw text created in the conducting OCR step does not include fonts or formatting. The post-editing step includes accessing a plurality of tailored error correction models, each tailored error correction model created by supervised learning specifically for one of the document types without the use of text regions. For each ID image of the validation request, the post-editing step includes selecting the tailored error correction model associated with the document type of the ID image, as identified in the identifying step, to: (1) execute the sequence-to- sequence error correction upon the raw text of the ID image of the validation request; and (2) delineate the corrected text of the ID image into a second plurality of data fields for the validation request. The deviation score of each validation request is evaluated according to a deviation threshold. Each validation request exceeding the deviation threshold is red flagged.

A second embodiment of the invention is a KYC platform for receiving and processing a plurality of validation requests, the platform comprising: (a) a request database in data communication with a network; (b) a facial matching module configured to compare each photograph to each of a plurality of library images to determine a duplication score for the validation request associated with the photograph; (c) a document type identifier configured to determine a document type of each ID image; (d) an OCR engine configured to convert the one or more ID images of each validation request to raw text; (e) a post-editor configured to execute sequence-to- sequence error correction upon the raw text of each ID image of each validation request to create corrected text for each ID image of the validation request; and (f) an evaluation module configured, for each validation request, to compare the biographical text delineated into the first plurality of data fields of the validation request against the corrected text delineated into the second plurality of data fields of the validation request to determine a deviation score for the validation request. The request database is configured to receive the validation requests through the network and store each validation request in one of a plurality of request records of the request database. Each validation request includes biographical text and one or more ID images purportedly identifying a customer associated with the validation request. The biographical text for each validation request is delineated into a first plurality of data fields within the request record for the validation request. At least one of the one or more ID images of each validation request includes a photograph of the customer. The library images are stored in the KYC platform or are accessible to the KYC platform through the network. The duplication score of each validation request is evaluated according to a duplication threshold. The facial matching module red flags each validation request exceeding the duplication threshold. The OCR engine is not configured to identify a plurality of text regions in each ID image. The raw text created by the OCR engine does not include fonts or formatting. The post-editor includes a plurality of tailored error correction models, each tailored error correction model created by supervised learning specifically for one of the document types without the use of text regions. For each ID image of the validation request, the post-editor selects the tailored error correction model associated with the document type of the ID image, as identified by the document type identifier, to: (1) execute the sequence-to- sequence error correction upon the raw text of the ID image of the validation request; and (2) delineate the corrected text of the ID image into a second plurality of data fields for the validation request. The deviation score of each validation request is evaluated according to a deviation threshold. Each validation request exceeding the deviation threshold is red flagged.

A third embodiment of the invention is a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out any one of the methods of the first embodiment and the methods of the alternatives of the first embodiment. A fourth embodiment of the invention is a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out any one of the methods of the first embodiment and the methods of the alternatives of the first embodiment.

In contrast to the background art, the invention does not employ text regions during the OCR conversion step by the OCR engine or the post-editor. The OCR engine of the invention performs OCR on each ID image as whole (rather than separating the ID image into a series of text regions). The post-editor of the invention similarly performs error correction upon the raw text of each ID image as a whole (rather than separating the ID image into a series of text regions). The tailored error correction model used by the post-editor of the invention are optimized to each document type through supervised learning. The document type for each ID image is determined prior to the post-editing step and this document type information is fed into the post-editor so that it may select the appropriate tailored error correction model.

Without the need for OCR engines with text region capability, the OCR engine can be selected for use with ID images of various photographic quality at a low cost and with high image-processing throughput. Hence, optimization of the OCR engine’s creation of raw text and the post-editor’s creation of corrected text can be independently optimized.

The architecture of the invention also enables easy swapping of OCR engines, as the required output of the OCR engine is limited to raw text. In the event an OCR engine with improved performance on ID images of various photographic quality or an OCR engine provided at a lower cost becomes available, the invention’s architecture enables the previous OCR engine to be swapped out and replaced with the new OCR engine.

Another technical benefit of the invention is that it does not strive to output a final decision for every validation request. To the extent that the invention is unable to meet a final decision on the validation request, the invention sends the validation request to a second level review. BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the present disclosure are described herein with reference to the drawings in which:

FIG. 1 is a flowchart representation of a computer- implemented KYC method in an embodiment of the invention.

FIG. 2 is a block diagram representing a KYC platform in a KYC system in an embodiment of the invention.

FIG. 3 is a textual flowchart representation of a in an embodiment of the invention.

FIG. 4 is a block diagram representing an identifying step for a document type in an embodiment of the invention.

FIG. 5 is a chart detailing the accuracy rates of document type identification in an embodiment of the invention.

FIG. 6 is a chart detailing the accuracy rates of post-editing error correction in an embodiment of the invention.

DETAILED DESCRIPTION

In the following detailed description, reference is made to the accompanying drawings, which form a part hereof. The illustrative embodiments described in the detailed description, drawings and claims are not meant to be limiting. Other embodiments can be utilized, and other changes can be made, without departing from the spirit or scope of the subject matter presented herein. Unless specified otherwise, the terms “comprising,” “comprise,” “including” and “include” used herein, and grammatical variants thereof, are intended to represent “open” or “inclusive” language such that they include recited elements but also permit inclusion of additional, un-recited elements. As used herein, a computer or server may include a single stand-alone computer, a server, multiple dedicated servers, and/or a virtual server running on a larger network of servers and/or cloud-based service. As used herein, a data processing system may reside within a single stand-alone computer, a server, multiple dedicated servers, a cloud-based service, and/or a virtual server running on a larger network of servers. As used herein, a database may store data to and access data from a single stand alone computer, a data server, multiple dedicated data servers, a cloud-based service, and/or a virtual server running on a larger network of servers.

FIG. 1 is a flowchart representation of a computer-implemented KYC method 10 in an embodiment of the invention. Validation requests 11 include biographical text 11A and at least one ID image 11B. The ID image 11B is the input for a facial matching step 1-01, a document type identifying step 1-02, and a conducting OCR step 1-03.

In a facial matching step 1-01, the ID images 11B are compared against a plurality of library images (not illustrated) to determine a duplication score for the validation request 11. Validation requests 11 with a duplication score exceeding a duplication threshold are red flagged and either rejected outright or fed into a second level review step 1-07.

Validation requests 11 with a duplication score not exceeding the duplication threshold are further processed in the document type identifying step 1-02 to determine a document type (identified in FIG. 1 as document types 01 . . . n).

The conducting OCR step 1-03 is performed upon each ID image 11B to create raw text 12. Raw text 12 does not include fonts or formatting. Raw text 12 could be represented using standard text codes such as ASCII, CCCII, or Unicode. The conducting OCR step 1-03 is intentionally limited to use of a simple OCR engine 24E, avoiding the capture of distinct font or formatting information. The conducting OCR step 1-03 does not delineate text fields as in the background art.

The post-editing step 1-04 has two inputs. A first input is the document type determined in the document type identifying step 1-02 (identified in FIG. 1 as document types 01 . . . n). The document type is used in the post-editing step 1-04 by the post-editor 24F to select a tailored error correction model 13 (identified in FIG. 1 as MODELoi _{. . . n}) for execution of a sequence-to-sequence error correction upon the raw text 12 to produce corrected text 14. Corrected text 14, like raw text 12 does not include fonts or formatting. The corrected text 14 is compared against the biographical text 11A submitted by the customer with the validation request 11 in an evaluation step 1-05 that determines a deviation score. The corrected text 14 can also be compared against third party data records 15 (from one of a plurality of third party data servers, as depicted in FIG. 2 item 21). Validation requests 11 with a deviation score exceeding a deviation threshold are red flagged and either rejected outright or fed into a second level review step 1-07. Validation requests 11 with a deviation score not exceeding the duplication threshold are deemed to have entered a passing step 1-06. Third parties could be governmental agencies, non-governmental agencies, credit agencies, partner financial institutions, and/or any other type entity with potentially relevant data for the KYC process.

The passing step 1-06 typically instigates further processing of the application for which the validation request 11 was submitted by the customer. For example, once a potential banking customer’s validation request 11 makes it through the KYC requirements, the banking customer bank account request application moves forward toward activation.

FIG. 2 is a block diagram representing a KYC platform 24 in a KYC system 20 in an embodiment of the invention. FIG. 2 depicts the entire KYC system 20. The KYC system 20 includes a KYC platform 24 in data communication with an off-platform 25 infrastructure 25. The off-platform 25 infrastructure 25, as illustrated, includes a laptop computer 22 accessing an example banking website for upload of biographical text 11A and at least one ID image 11B. The laptop computer 22 and a plurality of third party data servers 21 are connected to a web server 24A of the KYC platform 24 through the network 23. Validation requests 11 include the biographical text 11A and ID images 11B. The validation requests 11 are delivered from the customer’s laptop computer 22 to the KYC platform 24 via the network 23.

The KYC platform 24 includes a web server 24A, a request database 24B, a facial matching module 24C, a document type identifier 24D, an OCR engine 24E, a post-editor 24F, and an evaluation module 24G. The request database 24B can be a separate database or a portion of another database or another platform module. E.g., the request database 24B can be a queue or buffer in the web server 24A, the facial matching module 24C, the document type identifier 24D, the OCR engine 24E, the post-editor 24F, and/or the evaluation module 24G of the KYC platform 24.

FIG. 3 is a textual flowchart representation of a in an embodiment of the invention. The flowchart 3-00 illustrates the steps 3-01 to 3-04 (see below) for implementing an embodiment of the invention.

3-01 Customer enters biographical text 11A into a first plurality of data fields via a banking website using a laptop computer 22 and uploads an ID image 11B of the customer’s passport. The banking website sends the customer’s biographical text 11A and ID image 11B to the KYC platform 24 as a validation request 11 through the network 23.

3-02 A facial matching module 24C compares the customer’s passport photograph with images in a library of images. Passport photographs matching an image from the library results in either rejection or a second level review of the validation request 11.

3-03 The document type of the ID image 11B is determined first and used to select a tailored error correction model 13 associated with that document type to correct raw text 12 from the OCR engine 24E.

3-04 Corrected text 14 is compared to the biographical text 11A entered by the customer and possibly also third party data records 15. If the deviation score exceeds a deviation threshold, the validation request 11 is red flagged for rejection or a second level review.

FIG. 4 is a block diagram representing an identifying step for a document type in an embodiment of the invention. FIG. 4 illustrates details of the document type identifying step 1-02 of FIG. 1. In the document type identifying step 1-02, ID images 11B undergo CNN analysis 40. The CNN analysis 40 does not involve use of any biographical text 11A of the validation request 11. A single CNN model can be used on ID images 11B of all document types. Alternatively, multiple CNN models can be used by the KYC method 10, such as where each individual CNN model is used on just a subset of identification documents types, such as just the identification documents from a specific country or region. The CNN analysis 40 can determine a probability score for each document type. As illustrated in FIG. 4, the CNN analysis 40 has resulted in a 95% likelihood of the specific ID image 11B being a driver’s license.

The document type identifying step 1-02 depicted in FIG. 4 is performed prior to the post editing step 1-04 because the document type conclusion is used to select the appropriate tailored error correction model 13. E.g., while there may be only one CNN model to determine the document type, there are multiple tailored error correction models 13, each tailored error correction model 13 specifically created by supervised learning for just one of the document types.

FIG. 5 is a chart 5-00 detailing the accuracy rates of document type identification in an embodiment of the invention. The accuracy rates are determined across a set of sample ID images 11B including various document types commonly submitted in the Philippines for KYC validation requests 11. The overall accuracy in identifying the document type using CNN analysis 40 in an embodiment of the invention was over 93%.

FIG. 6 is a chart 6-00 detailing the accuracy rates of post-editing error correction in an embodiment of the invention. The accuracy rates in the chart 6-00 are detailed for a set of sample ID images 11B including driver’s licenses and UMID cards document types. The post-editing step 1-04 leverages the previous document type identifying step 1-02. With the knowledge of the document type, the KYC method 10 and KYC platform 24 are enabled to select the appropriate tailored error correction model 13 to correct the raw text 12. The overall accuracy in error correcting the raw text 12 using an embodiment of the invention was over 94%. Note that the less standardized data from the ID images 11B (such as the customer’s address) has a lower accuracy rating after error correction than the more standardized fields.

A general embodiment of the invention is a KYC method 10 or KYC platform 24 for receiving and processing a plurality of validation requests 11 which include ID images 11B. Validation requests 11 with fraudulent ID images 11B are red flagged using facial matching. Document types for each ID image 11B are identified. ID images 11B undergo OCR processing without the use of text regions to produce raw text 12. Raw text 12 from the ID images 11B is post-edited to create corrected text 14. Post-editing includes use of tailored error correction models 13 associated with the document type of the ID image 11B.

A first embodiment of the invention is a computer- implemented KYC method 10, the method comprising: (a) receiving at a request database 24B in data communication with a network 23 a plurality of validation requests 11 through a network 23; (b) storing each validation request 11 in one of a plurality of request records of the request database 24B; (c) facial matching to compare each photograph to each of a plurality of library images to determine a duplication score for the validation request 11 associated with the photograph; (d) identifying a document type of each ID image 11B (e) conducting OCR to convert the one or more ID images 11B of each validation request 11 to raw text 12; (f) post-editing the raw text 12 to execute sequence-to- sequence error correction upon the raw text 12 of each ID image 11B of each validation request 11 to create corrected text 14 for each ID image 11B of the validation request 11; and (g) evaluating each validation request 11 to compare the biographical text 11A delineated into the first plurality of data fields of the validation request 11 against the corrected text 14 delineated into the second plurality of data fields of the validation request 11 to determine a deviation score for the validation request 11. Each validation request 11 includes biographical text 11A and one or more ID images 11B purportedly identifying a customer associated with the validation request 11. The biographical text 11A for each validation request 11 is delineated into a first plurality of data fields within the request record for the validation request 11. At least one of the one or more ID images 11B of each validation request 11 includes a photograph of the customer. The duplication score of each validation request 11 is evaluated according to a duplication threshold. The facial matching step includes red flagging each validation request 11 exceeding the duplication threshold. The conducting OCR step does not identify a plurality of text regions in each ID image 11B. The raw text 12 created in the conducting OCR step does not include fonts or formatting. The post-editing step includes accessing a plurality of tailored error correction models 13, each tailored error correction model 13 created by supervised learning specifically for one of the document types without the use of text regions. For each ID image 11B of the validation request 11, the post-editing step includes selecting the tailored error correction model 13 associated with the document type of the ID image 11B, as identified in the identifying step, to: (1) execute the sequence-to-sequence error correction upon the raw text 12 of the ID image 11B of the validation request 11; and (2) delineate the corrected text 14 of the ID image 11B into a second plurality of data fields for the validation request 11. The deviation score of each validation request 11 is evaluated according to a deviation threshold. Each validation request 11 exceeding the deviation threshold is red flagged.

In an alternative of the first embodiment: (i) each validation request 11 red flagged for exceeding the duplication threshold is given a rejection status; and (ii) each rejection status includes a first log file detailing a rejection rationale.

In an alternative of the first embodiment: (a) each validation request 11 red flagged, in the evaluating step, for exceeding the deviation threshold is given a second level review recommendation; and (b) each second level review recommendation includes a second log file detailing a second level review rationale.

In an alternative of the first embodiment, the evaluating step further includes: (a) linking, through the network 23, to one or more third party data servers 21; (b) submitting a data request to at least one of the third party data servers 21 for each validation request 11, each data request associated with the customer purportedly identified in the validation request 11; (c) receiving a third party data record 15 for the customer in response to the data request submission, the third party data record 15 including additional biographical data for the customer; and (d) comparing the additional biographical data for the customer against the second plurality of data fields associated with the validation request 11 of the customer. In this embodiment: (a) to identify the customer in the data request, each data request can include a subset of the bibliographic text included in the validation request 11 associated with the data request; and (b) the subset of the bibliographic text can be at least one of a driver’s license number, a passport number, a tax identification number, a formal name, and an address.

In an alternative of the first embodiment, the post-editing step includes natural language processing with part-of-speech tagging. In an alternative of the first embodiment: (a) the step of identifying the document type of each ID image 11B is performed by a convolutional neural network analysis upon the ID image 11B; and (b) the identifying step includes determining the document type of each ID image 11B without the use of either the raw text 12 or the corrected text 14 associated with the ID image 11B.

A second embodiment of the invention is a KYC platform 24 for receiving and processing a plurality of validation requests 11, the platform comprising: (a) a request database 24B in data communication with a network 23; (b) a facial matching module 24C configured to compare each photograph to each of a plurality of library images to determine a duplication score for the validation request 11 associated with the photograph; (c) a document type identifier 24D configured to determine a document type of each ID image 11B; (d) an OCR engine 24E configured to convert the one or more ID images 11B of each validation request 11 to raw text 12; (e) a post-editor 24F configured to execute sequence-to-sequence error correction upon the raw text 12 of each ID image 11B of each validation request 11 to create corrected text 14 for each ID image 11B of the validation request 11; and (f) an evaluation module 24G configured, for each validation request 11, to compare the biographical text 11A delineated into the first plurality of data fields of the validation request 11 against the corrected text 14 delineated into the second plurality of data fields of the validation request 11 to determine a deviation score for the validation request 11. The request database 24B is configured to receive the validation requests 11 through the network 23 and store each validation request 11 in one of a plurality of request records of the request database 24B. Each validation request 11 includes biographical text 11A and one or more ID images 11B purportedly identifying a customer associated with the validation request 11. The biographical text 11A for each validation request 11 is delineated into a first plurality of data fields within the request record for the validation request 11. At least one of the one or more ID images 11B of each validation request 11 includes a photograph of the customer. The library images are stored in the KYC platform 24 or are accessible to the KYC platform 24 through the network 23. The duplication score of each validation request 11 is evaluated according to a duplication threshold. The facial matching module 24C red flags each validation request 11 exceeding the duplication threshold. The OCR engine 24E is not configured to identify a plurality of text regions in each ID image 11B. The raw text 12 created by the OCR engine 24E does not include fonts or formatting. The post-editor 24F includes a plurality of tailored error correction models 13, each tailored error correction model 13 created by supervised learning specifically for one of the document types without the use of text regions. For each ID image 11B of the validation request 11, the post-editor 24F selects the tailored error correction model 13 associated with the document type of the ID image 11B, as identified by the document type identifier 24D, to: (1) execute the sequence-to-sequence error correction upon the raw text 12 of the ID image 11B of the validation request 11; and (2) delineate the corrected text 14 of the ID image 11B into a second plurality of data fields for the validation request 11. The deviation score of each validation request 11 is evaluated according to a deviation threshold. Each validation request 11 exceeding the deviation threshold is red flagged.

In an alternative of the second embodiment: (a) each validation request 11 red flagged for exceeding the duplication threshold is given a rejection status; and (b) each rejection status includes a first log file detailing a rejection rationale.

In an alternative of the second embodiment: (a) each validation request 11 red flagged by the evaluation module 24G for exceeding the deviation threshold is given a second level review recommendation; and (b) each second level review recommendation includes a second log file detailing a second level review rationale.

In an alternative of the second embodiment, the evaluation module 24G is further configured to: (a) link, through the network 23, to one or more third party data servers 21; (b) submit a data request to at least one of the third party data servers 21 for each validation request 11, each data request associated with the customer purportedly identified in the validation request 11; (c) receive a third party data record 15 for the customer in response to the data request submission, the third party data record 15 including additional biographical data for the customer; and (d) compare the additional biographical data for the customer against the second plurality of data fields associated with the validation request 11 of the customer. In this embodiment: (a) to identify the customer in the data request, each data request can include a subset of the bibliographic text included in the validation request 11 associated with the data request; and (b) the subset of the bibliographic text can be at least one of a driver’s license number, a passport number, a tax identification number, a formal name, and an address. In an alternative embodiment of the second embodiment, the post-editor 24F is configured to execute natural language processing with part-of- speech tagging.

In an alternative of the second embodiment: (a) the document type identifier 24B employs a convolutional neural network analysis upon the ID image 11B to determine the document type; and (b) the document type identifier 24D is configured to determine the document type of each ID image 11B without the use of either the raw text 12 or the corrected text 14 associated with the ID image 11B.

A third embodiment of the invention is a computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out any one of the methods of the first embodiment and the methods of the alternatives of the first embodiment.

A fourth embodiment of the invention is a computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out any one of the methods of the first embodiment and the methods of the alternatives of the first embodiment.

The KYC method 10 and KYC platform 24 provide a technical solution directed at a cost- effective but accurate validation request 11 review process for large volumes of incoming validation requests 11 where the identification document images have varying readability, format, and content. Assisted by AI / ML techniques, the KYC method 10 and KYC platform 24 perform a first level automated review of each validation request 11. Validation requests 11 that cannot be validated by the KYC platform 24 can be either rejected outright or routed to a second level review.

The second level review can be performed by third party software and/or personnel, making the second level review more costly (per validation request 11) than the first level review. If personnel are involved in the second level review, this human review can also introduce a significant time delay for completion of the review. The KYC method 10 and KYC platform 24 is used to reduce the total number of validation requests 11 that are passed through to the second level review, thereby reducing the overall processing costs for KYC validation. There are two initial steps performed by the KYC method 10 and KYC platform 24 which are performed prior to the conducting OCR step 1-03. A first initial step, employed to detect duplicate or fraudulent accounts, is a facial matching step 1-01 comparing the faces pictured on the incoming ID images 11B with the faces captured in a plurality of library images. The library images can be photographs of existing customers, people in a proprietary database, and/or people in a public database. Matching facial features between the photographs within the ID images 11B submitted by a potential new customer and an existing customer can, for instance, indicate a duplicate or fraudulent KYC validation request 11.

A second initial step taken by the KYC method 10 and KYC platform 24 is the document type identifying step 1-02. This second initial step is performed on the ID image 11B, without the use of raw text 12 or corrected text 14. This second initial step is vital as it allows the KYC method 10 and KYC platform 24 to select a tailored error correction model 13 for the post-editing sequence-to-sequence error correction performed in the post-editing step 1-04. Use of the tailored error correction models 13 enables the post-editor 24F to consider the “context” of the raw text 12 without the need for the complexity and cost resulting from the use of text regions.

The tailored error correction models 13 can be trained with supervision to be familiar with the vocabulary and format of the typical biographical data found within each document type. The post-editing step 1-04 is approached as a “sequence-to-sequence” problem, where the raw text 12 is the source sequence and the corrected text 14 is the target sequence. The tailored error correction models 13 are each created using supervised learning. Each model is limited to a single document type. The training set for each model can include thousands of sample ID images (each including printed text), each sample ID images true text (e.g., the actual text printed in onto the identification card), each sample ID’s raw text 12 outputted from a given OCR engine 24E, and each sample ID image’s corrected text 14 from a sample post-editor 24F. In this manner the post-editor 24F (with its set of tailored error correction models 13) has an understanding of how the given OCR engine 24E misreads the printed text of the incoming ID images 11B and has learned how to correct these mistakes for each document type. To address readability variance for the printed text of the incoming ID images 11B, the KYC method 10 and KYC platform 24 have the flexibility to select a robust (and potentially low- cost) OCR engine 24E. An effective OCR engine 24E typically exhibits good performance with varying lighting conditions of the ID images 11B and also has the capability of reading longer texts. Because the KYC method 10 and KYC platform 24 only require raw text 12 of the entire ID image 11B, it is very simple to swap out and replace earlier OCR engines 24E with new revisions of the same OCR engine 24E, OCR engines 24E licensed by different third parties, or other home-grown OCR engines 24E. The invention is also flexible enough to accommodate different OCR engines 24E for use in different countries, for use with different languages, for use in different business application, and for use with different throughput volumes. In short, the invention is directed at optimizing flexibility, efficiency, accuracy, and costs. As a safety valve, to the extent the invention is unable to process specific validation requests 11, the invention directs these validation requests 11 to second level review. The second level review can be a third party KYC software product, a third party KYC service, or review by personnel.

While various aspects and embodiments have been disclosed herein, it will be apparent that various other modifications and adaptations of the invention will be apparent to the person skilled in the art after reading the foregoing disclosure without departing from the spirit and scope of the invention and it is intended that all such modifications and adaptations come within the scope of the appended claims. The various aspects and embodiments disclosed herein are for purposes of illustration and are not intended to be limiting, with the true scope and spirit of the invention being indicated by the appended claims.

Claims

1. A computer- implemented KYC method, the method comprising:

(a) receiving at a request database in data communication with a network a plurality of validation requests through a network;

(b) storing each validation request in one of a plurality of request records of the request database;

(i) wherein each validation request includes biographical text and one or more ID images purportedly identifying a customer associated with the validation request;

(ii) wherein the biographical text for each validation request is delineated into a first plurality of data fields within the request record for the validation request; and

(iii) wherein at least one of the one or more ID images of each validation request includes a photograph of the customer;

(c) facial matching to compare each photograph to each of a plurality of library images to determine a duplication score for the validation request associated with the photograph,

(i) wherein the duplication score of each validation request is evaluated according to a duplication threshold; and

(ii) wherein the facial matching step includes red flagging each validation request exceeding the duplication threshold;

(d) identifying a document type of each ID image;

(e) conducting OCR to convert the one or more ID images of each validation request to raw text,

(i) wherein the conducting OCR step does not identify a plurality of text regions in each ID image; and

(ii) wherein the raw text created in the conducting OCR step does not include fonts or formatting;

(f) post-editing the raw text to execute sequence-to- sequence error correction upon the raw text of each ID image of each validation request to create corrected text for each ID image of the validation request,

(i) wherein the post-editing step includes accessing a plurality of tailored error correction models, each tailored error correction model created by supervised learning specifically for one of the document types without the use of text regions; and

(ii) wherein, for each ID image of the validation request, the post-editing step includes selecting the tailored error correction model associated with the document type of the ID image, as identified in the identifying step, to:

(1) execute the sequence-to- sequence error correction upon the raw text of the ID image of the validation request; and

(2) delineate the corrected text of the ID image into a second plurality of data fields for the validation request; and

(g) evaluating each validation request to compare the biographical text delineated into the first plurality of data fields of the validation request against the corrected text delineated into the second plurality of data fields of the validation request to determine a deviation score for the validation request,

(i) wherein the deviation score of each validation request is evaluated according to a deviation threshold; and

(ii) wherein each validation request exceeding the deviation threshold is red flagged.

2. The method of claim 1,

(a) wherein each validation request red flagged for exceeding the duplication threshold is given a rejection status; and

(b) wherein each rejection status includes a first log file detailing a rejection rationale.

3. The method of claim 1,

(a) wherein each validation request red flagged, in the evaluating step, for exceeding the deviation threshold is given a second level review recommendation; and

(b) wherein each second level review recommendation includes a second log file detailing a second level review rationale.

4. The method of claim 1, wherein the evaluating step further includes:

(a) linking, through the network, to one or more third party data servers;

(b) submitting a data request to at least one of the third party data servers for each validation request, each data request associated with the customer purportedly identified in the validation request;

(c) receiving a third party data record for the customer in response to the data request submission, the third party data record including additional biographical data for the customer; and

(d) comparing the additional biographical data for the customer against the second plurality of data fields associated with the validation request of the customer.

5. The method of claim 4,

(a) wherein, to identify the customer in the data request, each data request includes a subset of the bibliographic text included in the validation request associated with the data request; and

(b) wherein the subset of the bibliographic text is at least one of a driver’s license number, a passport number, a tax identification number, a formal name, and an address.

6. The method of claim 1, wherein the post-editing step includes natural language processing with part-of-speech tagging.

7. The method of claim 1,

(a) wherein the step of identifying the document type of each ID image is performed by a convolutional neural network analysis upon the ID image; and

(b) wherein the identifying step includes determining the document type of each ID image without the use of either the raw text or the corrected text associated with the ID image.

8. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any one of claims 1-7.

9. A computer-readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the method of any one of claims 1-7.

10. A KYC platform for receiving and processing a plurality of validation requests, the platform comprising:

(a) a request database in data communication with a network,

(i) wherein the request database is configured to receive the validation requests through the network and store each validation request in one of a plurality of request records of the request database;

(ii) wherein each validation request includes biographical text and one or more ID images purportedly identifying a customer associated with the validation request;

(iii) wherein the biographical text for each validation request is delineated into a first plurality of data fields within the request record for the validation request; and

(iv) wherein at least one of the one or more ID images of each validation request includes a photograph of the customer;

(b) a facial matching module configured to compare each photograph to each of a plurality of library images to determine a duplication score for the validation request associated with the photograph,

(i) wherein the library images are stored in the KYC platform or are accessible to the KYC platform through the network;

(ii) wherein the duplication score of each validation request is evaluated according to a duplication threshold; and

(iii) wherein the facial matching module red flags each validation request exceeding the duplication threshold;

(c) a document type identifier configured to determine a document type of each ID image;

(d) an OCR engine configured to convert the one or more ID images of each validation request to raw text,

(i) wherein the OCR engine is not configured to identify a plurality of text regions in each ID image; and

(ii) wherein the raw text created by the OCR engine does not include fonts or formatting; (e) a post-editor configured to execute sequence-to- sequence error correction upon the raw text of each ID image of each validation request to create corrected text for each ID image of the validation request,

(i) wherein the post-editor includes a plurality of tailored error correction models, each tailored error correction model created by supervised learning specifically for one of the document types without the use of text regions; and

(ii) wherein, for each ID image of the validation request, the post-editor selects the tailored error correction model associated with the document type of the ID image, as identified by the document type identifier, to:

(f) an evaluation module configured, for each validation request, to compare the biographical text delineated into the first plurality of data fields of the validation request against the corrected text delineated into the second plurality of data fields of the validation request to determine a deviation score for the validation request,

11. The platform of claim 10,

12. The platform of claim 10,

(a) wherein each validation request red flagged by the evaluation module for exceeding the deviation threshold is given a second level review recommendation; and

13. The platform of claim 10, wherein the evaluation module is further configured to:

(a) link, through the network, to one or more third party data servers;

(b) submit a data request to at least one of the third party data servers for each validation request, each data request associated with the customer purportedly identified in the validation request;

(c) receive a third party data record for the customer in response to the data request submission, the third party data record including additional biographical data for the customer; and

(d) compare the additional biographical data for the customer against the second plurality of data fields associated with the validation request of the customer.

14. The platform of claim 13,

15. The platform of claim 10, wherein the post-editor is configured to execute natural language processing with part-of- speech tagging.

16. The platform of claim 10,

(a) wherein the document type identifier employs a convolutional neural network analysis upon the ID image to determine the document type; and

(b) wherein the document type identifier is configured to determine the document type of each ID image without the use of either the raw text or the corrected text associated with the ID image.