Hybrid method and device for transaction validation based on the comparison of two strings

Abstract

The transaction validation process compares a candidate string (Cc) to a reference string (Cr), representing names of countries, people, or organizations. It uses: a primary block (Bs) that determines an initial score (Si) based on rules or a lightweight model and then compares this score to predefined domains for immediate validation (Da), non-validation (Dna), and control (Dc). An expert block (Be) that uses complex AI to determine a similarity score (Sc) with complex rules and advanced models. The process generates a validation (A) or non-validation (NA) status based on these scores and activates the expert block if necessary to obtain similarity validation (Ds). Figure for publication: Figure 1

Description

Title of the invention: Hybrid method and device for transaction validation based on the comparison of two character strings Scope of the invention

[0001] The present invention relates to a hybrid device for validating a transaction by comparing a string of characters, called the candidate string, with a string of characters, called the reference string. It also relates to an associated method. Previous art

[0002] Verifying the correspondence between two names, of people or of organization, is an essential function in many applications ranging from database management to security protocols.

[0003] In the payments sector, there is a wide range of use cases that require robust name matching.

[0004] For example, instant verification of payment names: in several fraud scenarios, a fraudster lies about their name to a victim in order to deceive them. The ability to compare the entered name to the real name associated with an IBAN is therefore crucial to preventing this type of fraud.

[0005] Money laundering can also be more easily avoided if the names of account holders can be checked before each transaction.

[0006] Furthermore, preventing blacklisted users from making transactions requires checking a cardholder's name against public sanctions lists to prevent money laundering and criminal activity.

[0007] A name matching process must be able to handle a very diverse set of matching cases. • Match names containing typographical errors or missing characters: e.g., Sarah / Sarag, Kjel / Kjeld • Match homonyms, e.g., Kristof / Christophe • Match synonyms, e.g., Kristof / Christophe • Detect and ignore titles, for example: Mr / Dr / Ms... • detect matches with initials • detect matches with double surnames • process data relating to legal entities (BV, NV, SA,...) • etc.

[0008] A name matching process must respect constraints of time, cost and linguistic variability for each matching operation. Description of the invention

[0009] One idea that is at the basis of the invention is to observe that these matching operations are complex, long and energy-intensive to process.

[0010] A device and a method are proposed applying an original method of validating a transaction based on the comparison of a string of characters, called the candidate string, with a string of characters, called the reference string, the two strings of characters transcribing the names of countries or persons or organizations, comprising: • a so-called primary processing block, and • an expert processing block comprising an intelligence module artificial, said to be complex.

[0011] The device and method are configured for • generate a transaction validation status for the case where an initial comparison score belongs to a predetermined immediate validation domain, • generate a transaction non-validation state in the case where the initial comparison score belongs to a predetermined immediate non-validation domain, • activate the expert processing block when the initial comparison score belongs to a predetermined control domain, • generate a validation state when the similarity score belongs to a predetermined domain called similarity validation. The primary processing block is configured to: • determine the initial comparison score of the two strings, the comparison score being determined on the basis of a set of predefined rules or a lightweight machine learning model, then • compare said initial comparison score with a predetermined domain called immediate validation, a predetermined domain called non-immediate validation, and a predetermined control domain,

[0012] The expert processing block is configured to: • determine a similarity score for the two strings, the comparison similarity score being determined using a combination of complex rules and advanced machine learning models,

[0013] The device and method are configured for • generate a transaction validation status in the event that the initial comparison score falls within the predetermined immediate validation range, • generate a transaction non-validation state in the case where the initial comparison score falls within the predetermined immediate non-validation range, • activate the expert processing block when the initial comparison score falls within the predetermined control range, • generate a validation state when the similarity score belongs to a predetermined domain called similarity validation.

[0014] The hybrid device / method improves efficiency by employing the primary block for the rapid resolution of primary name matching scenarios, thus conserving computing resources.

[0015] The device / method guarantees the robustness and accuracy of the name matching by using the expert block for the most difficult cases, which the primary block cannot solve.

[0016] The two-block device / method is scalable, which allows the device to handle a wide range of complexities in multilingual name searches without compromising speed or accuracy.

[0017] The hybrid device / method offers the advantage of modularity, allowing the inclusion of a primary block and a complex expert block. This modularity enables the device to quickly resolve cases of primary and common name matching, while also handling more complex matching scenarios.

[0018] The device / method can be configured to return a suggested string of characters when the similarity score belongs to a predetermined domain called the similarity suggestion domain.

[0019] The suggested character string can be determined from an entry in a pre-existing database.

[0020] The device / method according to the invention can be configured to generate a invalidation state, when after the suggested string has been returned, the expert processing block is activated during a comparison of another candidate string with the reference string, and the similarity score of said other candidate string and the reference string does not belong to the similarity validation domain.

[0021] The suggested character string can be determined from an entry in a database.

[0022] Advantageously, the primary processing block is further configured to preprocess character strings corresponding to names by converting them in lowercase and removing character strings corresponding to common titles

[0023] Preprocessing names according to a uniform case and removing common titles normalizes the input data, which can significantly reduce the complexity of the matching process.

[0024] The initial preprocessing step can improve device performance by reducing the risk of false matches due to case sensitivity or irrelevant title information.

[0025] Simplifying input data through preprocessing can lead to faster processing times for the primary block, thereby improving the overall throughput of the device.

[0026] Preferably, the primary processing block identifies and replaces initials in character strings with corresponding full names.

[0027] By transforming initials into full names, the device can increase the accuracy of matches, particularly in cases where abbreviations can be a source of ambiguity or confusion. The transformation consists of searching the reference string to see if a word begins with the same letter as an initial present in the candidate string, and if so, performing the replacement.

[0028] This function can improve the user experience by reducing the number of potential false negatives, thereby increasing the reliability of the name matching process.

[0029] The expert processing block can use semantic embedding models to generate vector representations of names and compare the vectors to determine the complex similarity score.

[0030] The inclusion of semantic embedding models in the expert block offers the advantage of automation. These models generate vector representations of the names. It is then possible to determine a similarity score from these representations (for example, with a distance calculation), thus automating the process of determining name matches without manual intervention.

[0031] The use of semantic embedding models to create vector representations allows for a nuanced comparison of names, capturing semantic similarities that may not be apparent during a direct comparison of character strings.

[0032] The generation of similarity scores based on vector comparisons allows the device to more effectively manage orthographic and phonetic variations of names, as well as cases of transliteration (transcription of a name into a different writing system).

[0033] Vector representations facilitate the matching of names in different languages ​​and scripts, which is particularly useful in a multilingual context.

[0034] Advantageously, the expert processing block may include a language detection function for identifying and adjusting transliteration cases. The adjustment comprises a language detection step followed by transcription of the candidate character string into a target (Latin) alphabet.

[0035] This offers the advantage of heterogeneity. This function allows the device to identify and adjust transliteration cases, where names may be written differently in different languages. By taking linguistic variations into account, the device can more efficiently process multilingual name matching.

[0036] By identifying the language of the names, the device can apply language-specific processing rules, which improves the accuracy of the matching process.

[0037] The adjustment of transliteration cases ensures that the device remains effective even when dealing with names that have been converted from one writing system to another, which is common in a globalized environment.

[0038] Preferably, the primary processing block further includes a calculation component configured to return a matching score between two paired strings.

[0039] Providing a matching score indicating the determination of the match allows for a quantifiable evaluation of the results, which can be essential for applications requiring high accuracy, such as identity verification or data deduplication.

[0040] The matching score allows the device to classify potential matches, which facilitates user review and further analysis, thereby improving the efficiency of manual verification processes.

[0041] The calculation component can implement the Levenshtein distance between the two strings.

[0042] The use of Levenshtein distance ensures that the matching score reflects the actual similarity between the character strings, which improves the reliability of the matching process.

[0043] The matching score can be that returned for the identity of two strings of characters when the Levenshtein distance is less than a predetermined threshold.

[0044] The predetermined threshold for the Levenshtein distance allows the device to avoid false positives that can occur in the case of minor spelling variations or common typographical errors in a string of characters, thereby improving the accuracy of the device.

[0045] The processing block system can be configured to rearrange the components of a name in order to minimize the Levenshtein distance and facilitate matching

[0046] Reorganizing the components of a name to minimize the Levenshtein distance can significantly improve matching accuracy for multicultural or multilingual datasets where the order of names may vary, thus ensuring better quality data normalization.

[0047] The ability to facilitate matching by minimizing the Levenshtein distance can reduce the need to manually correct or normalize data, thereby saving time and resources in data processing workflows. Brief description of the figures

[0048] Other features and advantages of the invention will become apparent upon reading the detailed description that follows, for an understanding of which reference should be made to the accompanying drawings, in which: • [Fig.1] is a schematic view of a hybrid device according to the invention. • [Fig.2] is another schematic view of the hybrid device. Detailed description of the invention

[0049] The present invention relates to a hybrid device (D) designed to validate a transaction by comparing a string of characters Sc, called candidate, with a string of characters Sr, called reference, the two strings of characters transcribing names of countries, people, organizations or others.

[0050] The device comprises two parts: a primary processing block Bs equipped to handle primary matching tasks, and an expert processing block Be configured to handle the most complex name matching scenarios.

[0051] The first element of the invention is the primary processing block B s which can be a carefully crafted set of rules or a lightweight machine learning model.

[0052] The objective of the primary block B is to accelerate the name matching process by quickly resolving the most common and basic cases, such as exact matches or variations due to typographical errors. This first line of defense in the name matching sequence makes it possible to determine and immediately return the matching results when possible, which significantly improves the efficiency of the device.

[0053] The so-called primary processing block is configured to: • determine an initial If score for comparing the two strings, the comparison score being determined based on a set of predefined rules or a lightweight machine learning model, then • compare said initial score Si of comparison with a predetermined domain Da said of immediate validation, a predetermined domain Dna said of non-immediate validation and a predetermined domain De said of control.

[0054] The so-called expert processing block Be includes a complex artificial intelligence module Mi, configured to: • determine a similarity score (Sc) of the two strings, the comparison similarity score being determined using a combination of complex rules and advanced machine learning models.

[0055] The hybrid device D is configured to: • generate a validation state A of the transaction for the case where said initial comparison score belongs to the predetermined immediate validation domain Da, • generate a non-validation state Na of the transaction in the case where said initial comparison score belongs to the predetermined immediate non-validation domain Dna, • activate the expert processing block Be when the initial comparison score Si belongs to the predetermined control range, • generate a validation state A when the similarity score belongs to a predetermined domain Ds called similarity validation domain.

[0056] When the primary block B fails to determine a match, the task is transferred to the second block: the expert block Be. This sophisticated level consists of a combination of complex rules and high-level machine learning models. Semantic embedding models can be used to compare homonyms and multilingual names, using the distance between vectors as a similarity score. Language detection and transcription to a common Latin alphabet can also be used to handle transliteration cases. The expert block considers linguistic nuances and cultural contexts to distinguish similar names with high accuracy.

[0057] The cornerstone of this invention is the collaboration between the primary and complex blocks of artificial intelligence. The former acts as a filter, resolving the primary cases and only passing the complex cases on to the latter, thus saving computing resources and time. This multi-layered approach levels allows the hybrid D device to remain lightweight and fast without compromising its ability to handle complex matching tasks.

[0058] The diagram shown in [Fig. 1] comprises two primary blocks representing the main blocks of the device: • Reference Bs. “Primary Processing Block” — represented by a rectangular block at the top, this block is responsible for the initial filtering and processing of names based on predefined rules or a lightweight machine learning model. This block corresponds to the first part of the system, which handles cases of primary and common name matching. • Reference Be. “Expert processing block” — represented by a rectangular block below the primary block, this sophisticated block undertakes additional name processing using complex rules and advanced machine learning models when the primary AI block is insufficient to find a match.

[0059] In addition, three result states are described in the diagram: • Reference A — “Validation” — an oval shape suggests a successful matching result when a match is determined, • NA reference — “Non-Validation” — an oval shape indicates a negative result when it is determined that there is no match, • Reference SC — “Suggest a correction” — another oval shape indicates a negative result when a match is determined.

[0060] Thus, the diagram makes it possible to visualize the overall process and the decision flow of the device for matching multilingual names.

[0061] Of course, the invention is not limited to the examples just described, and many modifications can be made to these examples without departing from the scope of the invention. Furthermore, the various features, forms, variants, and embodiments of the invention can be combined in various ways, provided they are not incompatible or mutually exclusive.

Claims

1. Demands Method (P) for validating a transaction by comparing a string of characters (Ce), called the candidate string, with a string of characters (Cr), called the reference string, the two strings of characters transcribing the names of countries, people or organizations, the method implementing: • a primary processing block (B s) configured to: i. determine an initial comparison score (Si) between the two strings, the comparison score being determined based on a set of predefined rules or a lightweight machine learning model, then ii. compare said initial comparison score with a predetermined domain (Da) referred to as immediate validation, a predetermined domain (Dna) referred to as non-immediate validation, and a predetermined domain (De) referred to as control, • an expert processing block (Be) comprising a complex artificial intelligence module (Mi), configured to: i. determine a score (Sc), called a similarity score, for the two strings of characters, the comparison similarity score being determined using a combination of complex rules and advanced machine learning models, the process comprising: • a step for generating a validation state (A) of the transaction in the case where said initial comparison score belongs to the predetermined domain of immediate validation, • a step to generate a non-validation (NA) state for the transaction in the case where said initial comparison score belongs to the predetermined domain of immediate non-validation, • an activation step of the expert processing block when said initial comparison score belongs to the predetermined control domain, • a generation step of a validation state (A) when the similarity score belongs to a predetermined domain (Ds) called similarity validation.

2. Method according to the preceding claim, comprising a step of returning a string of characters (Sc), said to be suggested, when the similarity score belongs to a predetermined domain said to be similarity suggestion.

3. A method according to the preceding claim, wherein the suggested character string (Sc) is determined from an entry in a database (Bd).

4. A method according to either of the two preceding claims, further configured to generate an invalidation state (Na), where after the suggested string has been returned, the expert processing block (Be) is activated when comparing another candidate string with the reference string, and the similarity score of said other candidate string and the reference string does not fall within the similarity validation domain.

5. A method according to any one of the preceding claims in which the primary processing block (Bs) identifies and replaces, in character strings, the initials with corresponding full names.

6. A method according to any one of the preceding claims, wherein the expert processing block uses semantic embedding models to generate vector representations of names and compare the vectors to determine the complex similarity score.

7. A method according to any one of the preceding claims, wherein the expert processing block (Be) includes a language detection function for identifying and adjusting transliteration cases.

8. A method according to any one of the preceding claims, wherein the primary processing block further comprises a computation component (C) configured to return a matching score between two paired strings.

9.

10. A method according to the preceding claim, wherein the primary processing block (B s) is configured to rearrange the components of a name in order to minimize the Levenshtein distance and facilitate matching. Device (D) for validating a transaction based on the comparison of a string of characters (Ce), called the candidate, with a string of characters (Cr), called the reference, the two strings of characters transcribing names of countries or people, said device comprising: • a primary processing block (B s) configured to: i. determine an initial comparison score (Si) between the two strings, the comparison score being determined based on a set of predefined rules or a lightweight machine learning model, then ii. compare said initial comparison score with a predetermined domain (Da) referred to as immediate validation, a predetermined domain (Dna) referred to as non-immediate validation, and a predetermined domain (De) referred to as control, • an expert processing block (Be) comprising a complex artificial intelligence module (Mi), configured to: i. determine a score (Sc), called a similarity score, for the two strings of characters, the comparison similarity score being determined using a combination of complex rules and advanced machine learning models, the device being configured for • generate a validation state (A) of the transaction in the case where said initial comparison score belongs to the predetermined immediate validation domain, • generate a non-validation (NA) status for the transaction in the case where the initial comparison score belongs to the predetermined immediate non-validation domain, activate the expert processing block when said initial comparison score belongs to the predetermined control domain, generate a validation state (A) when the similarity score belongs to a predetermined domain (Ds) called similarity validation.

Images