Recall method and system incorporating textual representations

Abstract

The present disclosure relates to a recall method and system combining text representation. The method comprises: obtaining text information related to a merchant, the text information comprising at least one of a merchant name, a merchant address, and a WiFi text name related to the merchant; converting the text information into a first vector; determining candidate merchant pairs belonging to the same entity based on the first vector; and scoring and ranking the candidate merchant pairs to output merchant pairs belonging to the same entity.

Description

Technical Field

[0001] This disclosure relates to the field of risk control, and in particular to recall methods and systems that combine textual representations. Background Technology

[0002] Having the same entity control different merchants is important information for risk control, such as loan limits, internal risk control mechanisms for identifying and cracking down on related businesses, etc.

[0003] Existing recall methods targeting different merchants belonging to the same entity only directly link them based on information such as identification number, email address, and mobile phone number. They do not make full use of weak media such as text messages and Wi-Fi information, and cannot fully identify related merchant pairs.

[0004] This disclosure addresses, but is not limited to, the many factors mentioned above. Summary of the Invention

[0005] Based on this, this disclosure proposes a method and system for recalling merchants by comprehensively utilizing information from various dimensions, thereby enabling a more thorough exploration of relationships between merchants. Specifically, as a replacement or supplement to existing association-based recall (i.e., recall through phone numbers, identification numbers, business licenses, etc.), the method and system of this disclosure innovatively vectorize the merchants' textual information (such as merchant names, merchant addresses, WiFi text information, etc.), resulting in higher similarity of representation vectors for positive samples (belonging to the same entity) and lower similarity of representation vectors for negative samples (not belonging to the same entity). Therefore, this textual information can be used to recall merchants belonging to the same entity, thereby identifying merchants belonging to the same entity that were previously undetectable.

[0006] The method and system disclosed herein comprehensively consider association recall and vector recall, applying the multi-path recall concept from recommendation systems to merchant relationship recognition. In addition to strong media such as identity identifiers, phone numbers, and email addresses, Wi-Fi association is added. Simultaneously, text vector retrieval recall is performed based on merchant names, merchant addresses, and Wi-Fi text names, uncovering previously uncovered relationship pairs through multi-path recall. Furthermore, the method and system consider cross-domain text relationship pairs in the selection of training samples, and employ multi-task learning methods that comprehensively consider cosine sentences and unsupervised contrastive learning, improving the model's generalization ability and universal expressive power, and achieving better results on existing datasets. Additionally, the method and system employ whitening dimensionality reduction, compressing high-dimensional (e.g., 768 dimensions) text vectors to low-dimensional (e.g., 32 dimensions), significantly improving the efficiency of vector retrieval.

[0007] According to a first aspect of this disclosure, a recall method combining text representation is provided, comprising: acquiring text information related to a merchant, the text information including at least one of a merchant name, a merchant address, and a WiFi text name related to the merchant; converting the text information into a first vector; determining candidate merchant pairs belonging to the same entity based on the first vector; and scoring and ranking the candidate merchant pairs to output merchant pairs belonging to the same entity.

[0008] According to one embodiment, the method further includes obtaining at least one of a telephone number, identity number, email address, social network account, and WiFi address associated with a merchant, and the method further includes: when the at least one of the telephone number, identity number, email address, social network account, and WiFi address of two merchants is the same, determining that the two merchants belong to the same entity and selecting the two merchants as a candidate merchant pair.

[0009] According to another embodiment, the text information also includes a sequence of IP addresses from the merchant's login history, and the method further includes: converting the IP address sequence into a second vector; and determining candidate merchant pairs belonging to the same entity based on the second vector.

[0010] According to yet another embodiment, the transformation is achieved through a text representation model, which is obtained by training a preliminary model.

[0011] According to yet another embodiment, the preliminary model is the Mencius model.

[0012] According to yet another embodiment, the training uses a Cosine Sentence scheme, and the Cosine Sentence scheme uses the following loss function: Where λ is a hyperparameter, (u i ,u j ) and (u k ,u l ) represents a sample pair, where i, j, k, l are positive integers, and Ω is a positive integer. pos It is a set of positive samples, and Ω neg It is the set of negative samples, wherein the training makes it possible for any positive sample pair (u i ,u j )∈Ω pos and negative sample pairs (u k ,u l )∈Ω neg Both have cos(u) i ,u j ) greater than cos(u k ,u l ).

[0013] According to yet another embodiment, the training also uses contrastive learning to train the initial model, wherein the contrastive learning uses the SimCSE scheme.

[0014] According to yet another embodiment, the training further weights the loss function of the Cosine Sentence scheme with the loss function of the SimCSE scheme to obtain a weighted loss function.

[0015] According to yet another embodiment, the method further includes using whitening dimensionality reduction to reduce the dimensionality of the vectors output by the text representation model.

[0016] According to yet another embodiment, the method further includes applying risk control to merchants belonging to the same entity.

[0017] According to a second aspect of this disclosure, a recall system incorporating text representation is provided, comprising: an acquisition module configured to acquire merchant-related text information, the text information including at least one of a merchant name, a merchant address, and a WiFi text name; a text representation module configured to convert the text information into a first vector; an identification module configured to determine candidate merchant pairs belonging to the same entity based on the first vector; and an output module configured to score and sort the candidate merchant pairs to output merchant pairs belonging to the same entity.

[0018] According to one embodiment, the acquisition module is further configured to acquire at least one of the following associated with a merchant: a phone number, an identity identifier, an email address, a social network account, and a WiFi address. The identification module is further configured to determine that two merchants belong to the same entity and designate them as a candidate merchant pair when at least one of the following two merchants is the same: a phone number, an identity identifier, an email address, a social network account, and a WiFi address.

[0019] According to another embodiment, the text information also includes a sequence of IP addresses from the merchant's login history, and the text representation module is configured to convert the IP address sequence into a second vector, and the identification module is further configured to determine candidate merchant pairs belonging to the same entity based on the second vector.

[0020] According to yet another embodiment, the transformation is achieved through a text representation model, which is obtained by training a preliminary model.

[0021] According to yet another embodiment, the preliminary model is the Mencius model.

[0022] According to yet another embodiment, the training uses a Cosine Sentence scheme, and the Cosine Sentence scheme uses the following loss function: Where λ is a hyperparameter, (u i ,u j ) and (u k ,u l ) represents a sample pair, where i, j, k, l are positive integers, and Ω is a positive integer. pos It is a set of positive samples, and Ω neg It is the set of negative samples, wherein the training makes it possible for any positive sample pair (u i ,u j )∈Ω pos and negative sample pairs (u k ,u l )∈Ω neg Both have cos(u) i ,u j ) greater than cos(u k ,u l ).

[0023] According to yet another embodiment, the training also uses contrastive learning to train the initial model, wherein the contrastive learning uses the SimCSE scheme.

[0024] According to yet another embodiment, the training further weights the loss function of the Cosine Sentence scheme with the loss function of the SimCSE scheme to obtain a weighted loss function.

[0025] According to yet another embodiment, the text representation module is further configured to use whitening dimensionality reduction to reduce the dimensionality of the output vector.

[0026] According to yet another embodiment, the system further includes a risk control module configured to perform risk control on merchant pairs belonging to the same entity.

[0027] According to a third aspect of this disclosure, a computing system for recall incorporating textual representations is provided, comprising: a processor; and a memory storing computer-executable instructions that, when executed by the processor, perform the method according to a first aspect of this disclosure.

[0028] The aspects generally include, as substantially as described herein with reference to the accompanying drawings and as explained by the drawings, methods, apparatus, systems, computer program products, and processing systems.

[0029] The foregoing has broadly outlined the features and technical advantages of the examples according to this disclosure so that the following detailed description may be better understood. Additional features and advantages will be described thereafter. The disclosed concepts and specific examples can be readily used as the basis for modifying or designing other structures for implementing the same purposes as this disclosure. Such equivalent exercises do not depart from the scope of the appended claims. The characteristics of the concepts disclosed herein, in both their organization and manner of operation, and their associated advantages, will be better understood by considering the following description in conjunction with the accompanying drawings. Each drawing is provided for illustrative and descriptive purposes and does not define any limitation on the claims. Attached Figure Description

[0030] To gain a more detailed understanding of the features described above in this disclosure, reference can be made to various aspects of the above-briefly summarized content, some of which are illustrated in the accompanying drawings. However, it should be noted that the drawings illustrate only certain typical aspects of this disclosure and should not be considered as limiting its scope, as other equivalent aspects are permissible in this description. Identical reference numerals in different drawings may identify the same or similar elements.

[0031] Figure 1 A flowchart of a recall method incorporating text representations according to an example embodiment of the present disclosure is shown;

[0032] Figure 2 A schematic block diagram of a recall system incorporating text representation according to an example embodiment of the present disclosure is shown; and

[0033] Figure 3 A schematic block diagram of a computational system for recall incorporating textual representations is shown according to an example embodiment of the present disclosure. Detailed Implementation

[0034] Terminology Explanation:

[0035] Recall: Using certain strategies, features, or models to determine a candidate set, and then using the candidate set for subsequent scoring and ranking models to obtain search results.

[0036] Multi-path recall: Using different strategies, features or models to recall a portion of the candidate set, then mixing the candidate sets together for use by subsequent ranking models, and finally making them available for various downstream tasks.

[0037] Text Representation: Simply put, text representation is not treating text as a string, but as a vector that is more convenient to process mathematically (that is, text feature extraction).

[0038] The inventors recognized that different merchants belonging to the same entity is a crucial component of merchant profiling. The aim is to identify instances where multiple merchants are controlled and operated by the same entity in direct-to-direct, direct-to-indirect, and indirect-to-indirect connection scenarios. This information can ultimately be used to facilitate direct connections between indirect merchants, level out merchant loan limits, identify and eliminate potential fraudulent connections within risk control systems, and perform various other tasks. However, existing recall methods only directly associate merchants based on identification numbers, email addresses, and phone numbers, failing to fully utilize weaker media such as text messages and Wi-Fi information. This existing approach can only find strongly associated merchant pairs, but it cannot effectively cover weakly associated merchant pairs, and cannot fully uncover all related merchants.

[0039] The method and system disclosed herein can comprehensively utilize information from various dimensions for recall, thereby more fully uncovering the relationships between merchants. Specifically, as a replacement or supplement to the original association recall (i.e., recall through phone numbers, identification numbers, business licenses, etc.), the method and system disclosed herein innovatively vectorize the merchant's textual information (such as the merchant's name, address, WiFi text information, etc.), thereby making the representation vector similarity of positive samples (belonging to the same entity) higher and the representation vector similarity of negative samples (not belonging to the same entity) lower. Therefore, this textual information can be used to recall merchants belonging to the same entity, thus identifying merchants belonging to the same entity that were previously undetectable.

[0040] The method and system disclosed herein comprehensively consider association recall and vector recall, applying the multi-path recall concept from recommendation systems to merchant relationship recognition. In addition to strong media such as identity identifiers, phone numbers, and email addresses, Wi-Fi association is added. Simultaneously, text vector retrieval recall is performed based on merchant names, merchant addresses, and Wi-Fi text names, uncovering previously uncovered relationship pairs through multi-path recall. Furthermore, the method and system consider cross-domain text relationship pairs in the selection of training samples, and employ multi-task learning methods that comprehensively consider cosine sentences and unsupervised contrastive learning, improving the model's generalization ability and universal expressive power, and achieving better results on existing datasets. Additionally, the method and system employ whitening dimensionality reduction, compressing high-dimensional (e.g., 768 dimensions) text vectors to low-dimensional (e.g., 32 dimensions), significantly improving the efficiency of vector retrieval.

[0041] The detailed description that follows, taken in conjunction with the accompanying drawings, is intended as a description of various configurations and is not intended to represent the only configurations in which the concepts described herein can be practiced. This detailed description includes specific details to provide a thorough understanding of the various concepts. However, it will be apparent to those skilled in the art that these concepts can be practiced without these specific details.

[0042] refer to Figure 1 The diagram illustrates a flowchart of a recall method 100 incorporating textual representation according to an example embodiment of the present disclosure.

[0043] refer to Figure 1 Method 100 may include obtaining merchant-related text information at block 110. Those skilled in the art will understand that merchant-related text information can be obtained in any suitable manner. In one example, method 100 may receive input merchant text information. In another example, method 100 may read merchant text information from a file, database, form, etc. Furthermore, the merchant's text information can be any suitable text information. In one embodiment of this disclosure, the text information may include at least one of the following: merchant name, merchant address, and WiFi text name.

[0044] Next, in box 120, method 100 may include converting text information into a first vector. Here, the purpose of text information conversion is to vectorize the merchant's text information (such as merchant name, merchant address, etc.), so that the representation vectors of positive samples (merchant pairs belonging to the same entity) have higher similarity and the representation vectors of negative samples (merchant pairs not belonging to the same entity) have lower similarity.

[0045] In one embodiment of this disclosure, this transformation is achieved through a text representation model. In this embodiment, the text representation model can be obtained by training various suitable preliminary models. For example, the preliminary model can be an application of transfer learning, utilizing virtually unlimited text to learn context-dependent representations of each member of the input sentence. It implicitly learns general grammatical and semantic knowledge, and it can transfer knowledge learned from the open domain to downstream tasks. In a further embodiment, method 100 may employ the Mencius model, a pre-trained Chinese model based on the Transformer architecture. Compared to other Chinese language models, the Mencius model's greatest advantage is its small size and high precision, achieving excellent results with only 1 billion parameters. It adopts a lightweight training strategy, aiming to build a small model with a billion parameters, fully leveraging the potential of the model with existing parameters, which is conducive to rapid and low-cost deployment in real-world business scenarios. Table 1 below provides a performance evaluation of the Mencius model used in method 100 compared to other models.

[0046]

[0047] Table 1

[0048] As can be seen from Table 1, the Mengzi model used in this disclosure is small in size (10³M or 1B, referring to 10³ trillion parameters and 1 billion parameters, respectively), while the accuracy is relatively high (for example, in AFQMC, Mengzi has an accuracy of 74.58% compared to BERT's 74.16%, and so on).

[0049] In another embodiment of this disclosure, the initial model is trained using a Cosine Sentence scheme. According to this embodiment, the training data used is text-matched, labeled data. This labeled data is used to construct a sentence vector model, where the labeled data utilized are common sentence pair samples, i.e., each sample is in the format "(sentence 1, sentence 2, label)". In a preferred embodiment of this disclosure, the labels may include positive sample labels and negative sample labels.

[0050] The inventors recognized that existing loss functions in the Cosine Sentence scheme, such as t(1-cos(u,v)), are inadequate. 2 +(1-t)cos 2 The loss function (u,v) is essentially based on cosine similarity, aiming to maximize the similarity of positive samples and minimize the similarity of negative samples. However, directly optimizing cosine similarity using these loss functions doesn't yield ideal results. If optimization is performed directly based on cosine similarity (as shown in the loss function above), the target for negative sample pairs is too low, leading to overlearning and loss of generalization ability. Even setting a higher threshold for negative sample pairs, such as using a loss function t(1-cos(u,v))+(1-t)max(cos(u,v),0.7), so that optimization stops if the similarity of negative sample pairs is below 0.7, thus reducing the likelihood of overlearning, only partially alleviates the problem, and selecting the appropriate threshold remains challenging.

[0051] Therefore, in the preferred embodiment of this disclosure, the Cosine Sentence scheme used can employ the following loss function:

[0052] Where λ is a hyperparameter, (u i ,u j ) and (u k ,u l ) represents a sample pair, where i, j, k, l are positive integers, and Ω is a positive integer. pos It is a set of positive samples, and Ω neg It is the set of negative samples. Therefore, this training makes it possible for any positive sample pair (u... i ,u j )∈Ω pos and negative sample pairs (uk ,u l )∈Ω neg Both have cos(u) i ,u j ) greater than cos(u k ,u l This allows the similarity of positive sample pairs to be greater than that of negative sample pairs, and the degree of similarity is determined by the model itself. This makes the model's loss function only related to the relative order of the prediction results, rather than depending on specific values, thus eliminating the difficulty of setting thresholds.

[0053] In yet another embodiment of this disclosure, contrastive learning is used to train the initial model as an alternative to or supplement to the Cosine Sentence scheme. In this embodiment, the purpose of contrastive learning is to bring similar samples closer together and push away dissimilar samples, thereby improving the model's vector representation capability. According to this embodiment, contrastive learning can use the SimCSE scheme. SimCSE utilizes unsupervised learning to improve the vector representation capability of sentences, using Dropout as a data augmentation method to construct positive sample pairs for contrastive learning.

[0054] Considering that single-task training objectives may overlook potential information in related tasks that could improve the target task, adding other related tasks may lead to better final results. In another embodiment of this disclosure, preferably, the loss function of the Cosine Sentence scheme and the loss function of the SimCSE scheme can be weighted and combined to form a joint training scheme based on Cosine Sentence and contrastive learning.

[0055] Therefore, during training, the model input consists of merchant-related text, including merchant names, merchant addresses, and WiFi text names. Positive samples are text pairs that have been identified as belonging to the same entity. In a preferred embodiment, to specifically optimize the cross-domain text relationship matching problem, text pairs can include four categories: merchant name-merchant name, merchant name-merchant address, merchant name-WiFi text name, and merchant address-merchant address. After training, each text corresponds to a vector. For each batch of vectors, the model of this disclosure calculates two loss functions: a supervised CosineSentence loss function and an unsupervised SimCSE loss function, and finally weights the two loss functions.

[0056] Experimental results show that multi-task joint training can reduce the risk of overfitting and has better generalization ability. The model will try to obtain a representation suitable for all tasks, rather than being limited to a single task. In this embodiment, the weights of the two loss functions mentioned above can be arbitrarily chosen when weighting the combination.

[0057] Considering the high dimensionality of the output vectors of the text representation model (e.g., 768 dimensions), such high-dimensional data consumes excessive resources in the process of retrieving hundreds of millions of text vectors. In another embodiment of this disclosure, method 100 may further include using whitening dimensionality reduction to reduce the dimensionality of the vectors output by the text representation model. Thus, with minimal loss of model accuracy, high-dimensional (e.g., 768 dimensions) text vectors are compressed to low dimensions (e.g., 32 dimensions), thereby significantly improving the efficiency of vector retrieval.

[0058] In this embodiment, it is assumed that the sentence vectors already contain sufficient semantics. If using cosine similarity to compare the similarity of sentence pairs is ineffective, a reasonable guess is that the coordinate basis of the generated sentence vectors is not an orthonormal basis, i.e., it exhibits anisotropy. To solve the problem of sentence embedding vectors, it is necessary to transform the sentence vectors into an orthonormal basis. Therefore, the whitening operation in this disclosure is to transform all sentence vectors into vectors with a mean of 0 and a covariance matrix that is the identity matrix.

[0059] Continue to refer to Figure 1 In box 130, method 100 may include determining candidate merchant pairs belonging to the same entity based on a first vector and scoring and sorting the candidate merchant pairs in box 140 to output merchant pairs belonging to the same entity.

[0060] For example, method 100 can employ a dual-tower model. A first vector (which may include a merchant name vector, address vector, merchant address latitude and longitude vector, etc.) is input into the dual-tower model. Data fusion is performed through an attention mechanism to form a query vector. This query vector is then used to retrieve candidate vectors with sufficiently high similarity (this retrieval can also use the cosine loss function mentioned above), thereby obtaining an accurate score for the merchant relationship pair. Subsequently, if the score exceeds a predetermined threshold, it is determined that the merchant and the candidate merchant belong to the same entity.

[0061] Those skilled in the art will understand that the operations in boxes 130 and 140 are any other suitable scoring and sorting methods in the art, which will not be described further here.

[0062] In one embodiment of this disclosure, method 100 may acquire text information of a merchant, convert it into a vector, and then compare this vector with the vectors of stored merchants to determine candidate merchants belonging to the same entity as the acquired merchant, and form candidate merchant pairs. Subsequently, method 100 may score and sort these candidate merchant pairs to finally output merchant pairs belonging to the same entity that include the acquired merchant.

[0063] In another embodiment of this disclosure, method 100 may obtain text information of two merchants and determine whether the two merchants belong to the same entity, or method 100 may obtain text information of multiple merchants and determine which of these merchants belong to the same entity, which will not be elaborated here.

[0064] It will be understood that the merchant's text information may include different types of text such as merchant name, merchant address, WiFi text name, etc. Thus, in the comparison and determination process, method 100 compares text of the same type, such as comparing the merchant names of different merchants, comparing the merchant addresses of different merchants, and so on.

[0065] Subsequently, the output merchant pairs can be applied to various risk control scenarios, such as blacklist / whitelist management, related transactions, etc., which will not be elaborated here.

[0066] In another embodiment of this disclosure, as a supplement to text representation, method 100 may further include obtaining at least one of a telephone number, identity number, email address, social network account, and WiFi address associated with a merchant. Thus, method 100 may further include determining that two merchants belong to the same entity and designating them as a candidate merchant pair when at least one of the telephone number, identity number, email address, social network account, and WiFi address is the same for two merchants. In this way, the candidate merchant pairs obtained based on text representation (as described in step 130) can be supplemented to obtain more complete candidate information.

[0067] In another embodiment of this disclosure, the obtained text information may further include a sequence of IP addresses from the merchant's login history, and method 100 may further include converting the IP address sequence into a second vector and determining candidate merchant pairs belonging to the same entity based on the second vector.

[0068] Figure 2 A schematic block diagram of a recall system 200 incorporating textual representation according to an example embodiment of the present disclosure is shown.

[0069] like Figure 2 As shown, system 200 may include an acquisition module 201, a text representation module 203, a recognition module 205, and an output module 207. In one embodiment of this disclosure, the acquisition module 201 may be configured to acquire text information related to merchants, the text representation module 203 may be configured to convert the text information into a first vector, the recognition module 205 may be configured to determine candidate merchant pairs belonging to the same entity based on the first vector, and the output module 207 may be configured to score and sort the candidate merchant pairs to output merchant pairs belonging to the same entity. In another embodiment of this disclosure, the merchant's text information may include at least one of the merchant's name, merchant address, and WiFi text name.

[0070] Those skilled in the art will understand that the acquisition module 201 can be configured to acquire merchant-related text information in any suitable manner. In one example, the acquisition module 201 can be configured to receive input merchant text information. In another example, the acquisition module 201 can be configured to read merchant text information from files, databases, forms, etc.

[0071] In another embodiment of this disclosure, the acquisition module 201 may further be configured to acquire at least one of the following associated with a merchant: a telephone number, an identification number, an email address, a social network account, and a WiFi address. In this embodiment, the identification module 305 may further be configured to determine that two merchants belong to the same entity and designate them as a candidate merchant pair when at least one of the following is identical: a telephone number, an identification number, an email address, a social network account, and a WiFi address.

[0072] In yet another embodiment of this disclosure, the merchant's text information may further include a sequence of IP addresses representing the merchant's login history. In this embodiment, the text representation module 203 may also be configured to convert the IP address sequence into a second vector, and the identification module 205 may also be configured to determine candidate merchant pairs belonging to the same entity based on the second vector.

[0073] In another embodiment of this disclosure, the transformation is achieved through a text representation model. In this embodiment, the text representation model can be obtained by training a preliminary model, such as a Mencius model. Further according to this embodiment, the model can be trained using a Cosine Sentence scheme. In this embodiment, the Cosine Sentence scheme uses the following loss function:

[0074] Where λ is a hyperparameter, (u i ,u j ) and (u k ,u l ) represents a sample pair, where i, j, k, l are positive integers, and Ω is a positive integer. pos It is a set of positive samples, and Ω neg It is a set of negative samples.

[0075] This training ensures that for any positive sample pair (u) i ,u j )∈Ω pos and negative sample pairs (u k ,u l )∈Ω neg Both have cos(u) i ,u j ) greater than cos(uk ,u l ).

[0076] In yet another embodiment of this disclosure, training may also utilize contrastive learning to train the initial model. In this embodiment, the contrastive learning may employ the SimCSE scheme.

[0077] In a preferred embodiment of this disclosure, the model training can further involve weighting the loss function of the Cosine Sentence scheme with the loss function of the SimCSE scheme to obtain a weighted loss function. This allows for a joint training scheme based on Cosine Sentence and contrastive learning, thereby reducing the risk of overfitting, improving generalization ability, and enabling the model to attempt representations suitable for all tasks, rather than being limited to a single task.

[0078] In yet another embodiment of this disclosure, the text representation module 203 may also be configured to use whitening dimensionality reduction to reduce the dimensionality of the output vector.

[0079] In yet another embodiment of this disclosure, system 200 may optionally include a risk control module 209. In this embodiment, risk control module 209 may be configured to apply merchant pairs belonging to the same entity for risk control. For example, merchant pairs belonging to the same entity may be used for blacklist / whitelist settings, related transaction alerts, etc.

[0080] Figure 3 A schematic block diagram of a computing system 300 for recalling text representations according to an embodiment of the present disclosure is shown. As shown, the computing system 300 includes a processor 305 and a memory 310. The memory 310 stores computer-executable instructions 315 that can be executed by the processor 305 to implement the above-described combination. Figure 1 The method described.

[0081] Therefore, the method and system disclosed herein divide the identification of merchant relationships into two stages: the first stage is the multi-path recall stage of the candidate relationship set, which adds vector-based recall on the basis of association recall; the second stage is the downstream merchant relationship identification stage, which, after constructing the overall multi-path recall relationship pool, can perform corresponding merchant relationship modeling according to actual business needs, and can also be used for employee relationship identification, general acquaintance relationship identification, etc.

[0082] It will be understood that although various embodiments have been described in this disclosure by way of merchants, the methods and systems of this disclosure can also be applied to a variety of other scenarios, such as the affiliation of different personnel, etc.

[0083] The above detailed description includes references to the accompanying drawings, which form part of the detailed description. The drawings illustrate specific embodiments that can be practiced by way of illustration. These embodiments are also referred to herein as “examples.” Such examples may include elements other than those shown or described. However, examples including the shown or described elements are also contemplated. Furthermore, examples of any combination or arrangement of those elements shown or described are contemplated, or with reference to specific examples (or one or more aspects thereof) shown or described herein, or with reference to other examples (or one or more aspects thereof) shown or described herein.

[0084] In the appended claims, the terms “comprising” and “including” are open-ended, meaning that a system, apparatus, article of manufacture, or process containing elements other than those listed after such terms in a claim is still considered to fall within the scope of that claim. Furthermore, in the appended claims, the terms “first,” “second,” and “third,” etc., are used merely as designations and are not intended to indicate a numerical order of their contents.

[0085] Furthermore, the order of operations described in this specification is exemplary. In alternative embodiments, the operations may be performed in a different order than that shown in the accompanying drawings, and the operations may be combined into a single operation or broken down into more operations.

[0086] The above description is intended to be illustrative and not restrictive. For example, the examples described above (or one or more aspects thereof) may be used in conjunction with other embodiments. Other embodiments may be used by those skilled in the art after reviewing the above description. The abstract allows the reader to quickly determine the nature of this technical disclosure. This abstract is submitted and it is understood that it is not intended to interpret or limit the scope or meaning of the claims. Furthermore, in the above detailed description, various features may be grouped together to make this disclosure flow smoothly. However, the claims may not state every feature disclosed herein, as embodiments may characterize a subset of said features. Furthermore, embodiments may include fewer features than those disclosed in a particular example. Therefore, the appended claims are thus incorporated into the detailed description, with each claim existing independently as a separate embodiment. The scope of the embodiments disclosed herein should be determined by reference to the full scope of the appended claims and equivalents of such claims.

Claims

1. A recall method combining text representation, comprising: Obtain text information related to the merchant, including at least one of the merchant name, merchant address, and WiFi text name; The text information is converted into a first vector; Based on the first vector, candidate merchant pairs belonging to the same entity are determined; and The candidate merchant pairs are scored and sorted to output merchant pairs belonging to the same entity. The text information also includes a sequence of IP addresses from the merchant's login history, and the method further includes: Convert the IP address sequence into a second vector; as well as Candidate merchant pairs belonging to the same entity are determined based on the second vector, wherein the transformation is achieved through a text representation model, which is obtained by training a preliminary model, wherein the training uses a Cosine Sentence scheme, and wherein the Cosine Sentence scheme uses the following loss function: , where λ is a hyperparameter, (u 𝑖 ,u 𝑗 ) and (u 𝑘 ,u l ) represents a sample pair, where i, j, k, l are positive integers, and Ω is a positive integer. 𝑝𝑜𝑠 It is a set of positive samples, and Ω neg It is a set of negative samples. The training described therein makes it possible for any positive sample pair (u) 𝑖 ,u 𝑗 )∈ Ω 𝑝𝑜𝑠 and negative sample pairs (u 𝑘 ,u l ) ∈ Ω 𝑛𝑒𝑔 They all have cosplay (𝑢) 𝑖 , 𝑢 𝑗 ) is greater than cos(𝑢 𝑘 , 𝑢 l ).

2. The method according to claim 1, further comprising obtaining at least one of a telephone number, identity number, email address, social network account, and WiFi address associated with the merchant, and the method further comprising: If at least one of the following is the same for two merchants: their phone number, identification number, email address, social network account, or WiFi address, then the two merchants are identified as belonging to the same entity and are designated as a candidate merchant pair.

3. The method according to claim 1, wherein the preliminary model is the Mencius model.

4. The method of claim 1, wherein the training further uses contrastive learning to train the initial model, wherein the contrastive learning uses the SimCSE scheme.

5. The method of claim 4, wherein the training further weights the loss function of the Cosine Sentence scheme with the loss function of the SimCSE scheme to obtain a weighted loss function.

6. The method according to claim 5 further includes using whitening dimensionality reduction to reduce the dimensionality of the vector output by the text representation model.

7. The method according to claim 1 further includes applying merchant pairs belonging to the same entity for risk control.

8. A recall system incorporating textual representation, comprising: The acquisition module is configured to acquire text information related to the merchant, the text information including at least one of the merchant name, merchant address, and WiFi text name related to the merchant; A text representation module, configured to convert the text information into a first vector; The identification module is configured to determine candidate merchant pairs belonging to the same entity based on the first vector; as well as An output module, configured to score and sort candidate merchant pairs to output merchant pairs belonging to the same entity. The text information also includes IP address sequences of merchant login history, and the text representation module is configured to convert the IP address sequences into a second vector. The recognition module is further configured to determine candidate merchant pairs belonging to the same entity based on the second vector. This conversion is achieved through a text representation model, which is obtained by training a preliminary model. The training uses a Cosine Sentence scheme, and the Cosine Sentence scheme uses the following loss function: , where λ is a hyperparameter, (u 𝑖 ,u 𝑗 ) and (u 𝑘 ,u l ) represents a sample pair, where i, j, k, l are positive integers, and Ω is a positive integer. 𝑝𝑜𝑠 It is a set of positive samples, and Ω neg It is a set of negative samples. The training described therein makes it possible for any positive sample pair (u) 𝑖 ,u 𝑗 )∈ Ω 𝑝𝑜𝑠 and negative sample pairs (u 𝑘 ,u l ) ∈ Ω 𝑛𝑒𝑔 They all have cosplay (𝑢) 𝑖 , 𝑢 𝑗 ) is greater than cos(𝑢 𝑘 , 𝑢 l ).

9. The system according to claim 8, wherein the acquisition module is further configured to acquire at least one of a telephone number, identity number, email address, social network account, and WiFi address associated with a merchant, and the identification module is further configured to determine that two merchants belong to the same entity and to designate the two merchants as a candidate merchant pair when at least one of the telephone number, identity number, email address, social network account, and WiFi address of two merchants is the same.

10. The system according to claim 8, wherein the preliminary model is the Mencius model.

11. The system of claim 8, wherein the training further uses contrastive learning to train the initial model, wherein the contrastive learning uses the SimCSE scheme.

12. The system of claim 11, wherein the training further weights the loss function of the Cosine Sentence scheme with the loss function of the SimCSE scheme to obtain a weighted loss function.

13. The system of claim 12, further comprising the text representation module being configured to use whitening dimensionality reduction to reduce the dimensionality of the output vector.

14. The system according to claim 13 further includes a risk control module, the risk control module being configured to apply risk control to merchants belonging to the same entity.

15. A computational system for recall incorporating textual representations, comprising: processor; as well as A memory storing computer-executable instructions that, when executed by the processor, perform the method according to any one of claims 1-7.

Images