Fraud risk prediction method and device, computing equipment and computer program product

By using triple information in the knowledge graph in the insurance fraud risk prediction and combining text data, the problem of low prediction accuracy in the prior art is solved, and the accuracy of insurance fraud risk prediction is improved.

CN118916493APending Publication Date: 2024-11-08PEOPLE'S INSURANCE COMPANY OF CHINA
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202410954174.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

In the prior art, in the prediction of insurance fraud risk, relying solely on numerical and category characteristics leads to low prediction accuracy.

Method used

Use named entities in the text data to obtain triple information from the pre-constructed knowledge graph and input them into the trained fraud risk prediction model to improve prediction accuracy.

Benefits of technology

By combining text data and triple information of named entities, the characteristics of fraud risk prediction are enriched, the accuracy of prediction results is improved, and the prediction results are more credible and relevant.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916493A_ABST
    Figure CN118916493A_ABST
Patent Text Reader

Abstract

The invention discloses a fraud risk prediction method and device, computing equipment and a computer program product, and aims to solve the problem of low prediction accuracy of fraud risk prediction in the prior art. The prediction method comprises the steps of obtaining text data used for describing a target case; on the basis of the named entities in the text data, triple information containing the named entities is obtained from a pre-constructed knowledge graph; and inputting the text data and the triple information containing the named entity into the trained fraud risk prediction model to obtain a fraud risk prediction result for the target case output by the fraud risk prediction model. The input of the fraud risk prediction model contains the triple information of the named entity in the text data, and the knowledge graph is constructed through the credible information capable of being used for insurance fraud identification, so that when the fraud risk is predicted based on the triple information and the text data for describing the target case, the accuracy of the prediction result can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence, and in particular to a fraud risk prediction method, apparatus, computing device and computer program product. Background Art

[0002] The goal of predicting insurance fraud risk is to identify potential insurance fraud behaviors in order to reduce the losses incurred by insurance companies due to fraudulent behaviors.

[0003] At present, the prediction methods of fraud risk usually include: obtaining numerical characteristics (such as insurance time, reporting time, case damage assessment amount, claim amount, etc.) and category characteristics (such as life insurance cases, vehicle insurance cases, property insurance cases, etc.) of the target case; predicting the fraud risk of the target case based on the obtained characteristics.

[0004] Practice has shown that if fraud risk prediction is based only on numerical and categorical features, the accuracy of the prediction results will be low because the features are too one-sided. Summary of the invention

[0005] The embodiments of the present application provide a fraud risk prediction method, apparatus, computing device and computer program product to solve the problem of low prediction accuracy in fraud risk prediction using existing technologies.

[0006] The present application embodiment adopts the following technical solutions:

[0007] A fraud risk prediction method, comprising:

[0008] Obtain text data for describing the target case;

[0009] Based on the named entities in the text data, obtaining triple information containing the named entities from a pre-built knowledge graph;

[0010] Inputting the text data and the triple information containing the named entity into a trained fraud risk prediction model to obtain a fraud risk prediction result for the target case output by the fraud risk prediction model;

[0011] The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0012] A model training method, comprising:

[0013] Obtain text data for describing sample cases;

[0014] Based on the named entities in the text data, obtaining triple information containing the named entities from a pre-built knowledge graph;

[0015] Using the text data and the triple information containing the named entity as input of the fraud risk prediction model to be trained, and using the fraud results of the sample case as the true results of the fraud risk prediction model;

[0016] In each round of training iteration, the text data describing the sample case and the triple information containing the named entities in the text data are input into the fraud risk prediction model to be trained, and the value of the loss function is determined based on the output result of the fraud risk prediction model to be trained in each round of iteration and the actual result;

[0017] Based on the principle of minimizing the value of the loss function, the parameters of the fraud risk prediction model are adjusted to obtain a trained fraud risk prediction model;

[0018] The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0019] A fraud risk prediction device, comprising:

[0020] A text data acquisition unit, which acquires text data used to describe a target case;

[0021] A triple information acquisition unit, based on the named entities in the text data, acquires triple information containing the named entities from a pre-constructed knowledge graph;

[0022] A risk prediction unit, configured to input the text data and the triplet information containing the named entity into a trained fraud risk prediction model to obtain a fraud risk prediction result output by the fraud risk prediction model for the target case;

[0023] The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0024] A model training device, comprising:

[0025] A text data acquisition unit, which acquires text data used to describe the sample case;

[0026] A triple information acquisition unit, based on the named entities in the text data, acquires triple information containing the named entities from a pre-constructed knowledge graph;

[0027] An information setting unit, used to use the text data and the triple information containing the named entity as inputs of the fraud risk prediction model to be trained, and use the fraud results of the sample cases as the actual results of the fraud risk prediction model;

[0028] The iterative training unit inputs the text data describing the sample case and the triple information containing the named entities in the text data into the fraud risk prediction model to be trained, and determines the value of the loss function based on the output result of the fraud risk prediction model to be trained in each round of iteration and the actual result;

[0029] The parameter adjustment unit adjusts the parameters of the fraud risk prediction model based on the principle of minimizing the value of the loss function to obtain a trained fraud risk prediction model;

[0030] The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0031] A computing device, comprising: a memory and a processor, wherein:

[0032] The memory is used to store computer programs;

[0033] The processor is coupled to the memory and is used to execute the computer program stored in the memory to perform the above method.

[0034] A computer-readable storage medium storing a computer program, wherein the computer program can implement the above method when executed by a computer.

[0035] At least one of the above technical solutions adopted in the embodiments of the present application can achieve the following beneficial effects:

[0036] The text data describing the target case and the triple information containing the named entities in the text data are jointly input into the fraud risk prediction model to predict the fraud risk of the target case, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed using credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model is credible information related to the target case, when predicting the fraud risk based on the triple information and the text data describing the target case, the accuracy of the prediction result can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 A specific implementation flow chart of a fraud risk prediction method provided in an embodiment of the present application;

[0039] Figure 2 A specific implementation flow chart of a model training method provided in an embodiment of the present application;

[0040] Figure 3 An example diagram of a sentence tree formed in the model training method provided in an embodiment of the present application;

[0041] Figure 4 For the embodiment of this application, Figure 3 Schematic diagram of soft position encoding of sentence trees in ;

[0042] Figure 5 A schematic diagram of the specific structure of a fraud risk prediction device provided in an embodiment of the present application;

[0043] Figure 6 A schematic diagram of the specific structure of a model training device provided in an embodiment of the present application;

[0044] Figure 7 A schematic diagram of the specific structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0046] Those skilled in the art may appreciate that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of the present application are also applicable to similar technical problems.

[0047] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and need not be used to describe a specific order or sequential order. It should be understood that the terms used in this way can be interchangeable under appropriate circumstances, which is only to describe the distinction mode adopted by the objects of the same attributes when describing in the embodiments of the present application. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, so that the process, method, system, product or equipment comprising a series of units need not be limited to those units, but may include other units that are not clearly listed or inherent to these processes, methods, products or equipment.

[0048] In order to solve the problem of low prediction accuracy when using existing technologies to predict fraud risks, Example 1 of the present application provides a fraud risk prediction method.

[0049] The executor of the method may be any computing device that can implement the method, such as a server, a mobile phone, a personal computer, a smart wearable device, a smart robot, and the like.

[0050] In addition, the embodiments of the present application do not limit the execution order of different steps. When using the method provided in the embodiments of the present application, the execution order of different steps can be adjusted according to actual needs.

[0051] For ease of description, the method provided in the embodiment of the present application is described in detail below by taking a fraud risk prediction device as the executor of the method as an example.

[0052] like Figure 1 As shown, it is a specific implementation flow chart of a fraud risk prediction method provided in an embodiment of the present application, including the following steps 11 to 13:

[0053] Step 11, obtaining text data for describing the target case.

[0054] The target case in this embodiment may be an insurance claim case. An insurance claim case refers to a case in which an insurance company compensates the insured for a risk accident or loss that occurs during the insurance period according to the terms of the insurance contract.

[0055] The target case may involve various types of insurance, such as vehicle insurance, property insurance, life insurance, etc.

[0056] Text data refers to data stored in text form. Text data usually includes a series of characters (including letters, numbers, punctuation marks and other special characters) that can express complete meaning. Text data usually exists in the form of files, such as txt, csv, xml and other formats.

[0057] The text data describing the target case can be the materials required for applying for insurance claims. The text data at least includes the claim application form (including the insured's personal information, accident description, loss situation, etc.), the insurance contract (to confirm the scope and conditions of the claim), and the identity document (to confirm the identity of the insured); the text data can also include the corresponding accident report (if it is a vehicle insurance claim, the accident process, liability determination, etc.), medical certificate (if it is a life insurance claim, medical certificate is required, such as doctor's diagnosis, treatment plan, etc.) or property loss list (if it is a property insurance claim, a property loss list and the value of each lost item in the list are required) based on the type of the target case.

[0058] Step 12: Based on the named entities in the text data, triple information containing the named entities is obtained from a pre-constructed knowledge graph.

[0059] Named entities refer to information in text data that has a specific meaning or refers to a specific object. This information is usually very important for understanding the semantics of text data. Usually, named entities include three major categories (entity, time, and number) and seven subcategories (personal names, organization names, place names, time, date, currency, and percentage).

[0060] In this embodiment, the named entities in the text data may be people, cars and related information, such as the policyholder, the insured, the parties to the target case, the target car, the license plate number of the target car, etc.

[0061] A knowledge graph is a structured semantic knowledge base that is used to describe entities (or concepts) in the physical world and the relationships between entities (usually semantic relationships).

[0062] The knowledge graph is a semantic network composed of nodes and edges. Nodes represent entities (or concepts) in the physical world, and edges represent the attributes of entities or semantic relationships between entities. Among them, entities in the physical world, such as people, places, objects, etc.; concepts are information that abstractly describes entities, such as "animal" is a concept, and "dog" is an entity under this concept; attributes are usually used to describe the characteristics of entities, such as "age", "color", etc.; attribute values ​​are the values ​​of attributes, such as "12 years old", "red"; semantic relationships, such as "father-son relationship", "insurance relationship", etc.

[0063] The basic data unit of the knowledge graph is a triple, which consists of a subject (entity), a predicate (attribute or relationship), and an object (attribute value or another entity). Usually, a triple expresses facts in the form of "entity 1 (such as "Zhang San")-relationship (such as "friend")-entity 2 (such as "Li Si")", "entity (such as "Zhang San")-attribute (such as "age")-attribute value (such as "20 years old")", connecting the subject and the object through the predicate to form a network knowledge structure. One representation of a triple can be: head entity, relationship, tail entity.

[0064] The knowledge graph in this embodiment refers to a knowledge graph constructed based on relevant data in the insurance field. Specifically, it can be constructed based on trusted information that can be used to identify insurance fraud, and two named entities and the semantic relationship between the two named entities are extracted from the trusted information to form a triple, wherein the semantic relationship can be determined by performing semantic analysis on the trusted information.

[0065] Among them, the relevant data used to construct the specified knowledge graph includes at least insurance policy data, accident data, claims data, industry fraud data and company blacklist data.

[0066] Policy data refers to information and records related to a specific policy (insurance contract), including at least: basic policy information (policy number, policy start and end time, insurance product type, insurance amount, etc.), policyholder information (name or title, contact information, address, etc.), insured person information (name, contact information, address, etc.), premium payment information (premium amount, payment method, payment record, etc.), policy change information (such as change time, change content, etc.), etc.

[0067] Accident data usually includes information related to vehicle accidents or insurance claims, including at least: basic information about the accident (such as the time, address, accident vehicle information, etc.), accident description information (such as accident type, liability type and detailed description of the accident, etc.), claim information (such as assessed damage amount, claim amount, etc.) and vehicle owner information (name, contact information), etc.

[0068] Claims data usually includes various information related to the insurance claims process, including at least: claim application information (such as applicant information, policy information and accident information), claim amount (including claim amount, approved amount and payment amount), claim type determined based on insurance products (such as medical insurance claims, critical illness insurance claims, accident insurance claims, etc.), claim time (including application submission time, insurance company review time, actual payment time), claim reason (specific reason for claim, such as illness, accident, etc.); it may also include reasons for rejection (such as the applicant did not provide sufficient information for underwriting, it is a disclaimer clause). Usually, in a claim case, there is only one reason for claim and one reason for rejection, or only one of them is filled with specific content.

[0069] Based on the named entities in the text data, triple information containing the named entities is obtained from the pre-built knowledge graph, specifically including:

[0070] Identify named entities in text data, query the pre-built knowledge graph based on the identified named entities, and obtain triple information containing the identified named entities from the knowledge graph.

[0071] For example, the text data describing the target case (the actual text data may contain multiple sentences, which is only an example and does not constitute a limitation on the text data) is "The vehicle with license plate number Beijing A11111 driven by Zhang San collided with the vehicle with license plate number Beijing B22222 driven by Li Si on Chang'an Avenue", and the identified named entities include "Zhang San", "Li Si", "Beijing A11111", "Beijing B22222" and "Chang'an Avenue"; based on the identified named entities, the pre-constructed knowledge graph is queried respectively to obtain triple information containing named entities, such as: (Wang Wu, insured, Zhang San), (Zhang San, insured person, policy number 123), (Li Si, insured, Wang Wu), (Wang Wu, car owner, Beijing A11111) and (Li Si, car owner, Beijing B22222). It should be noted that for a named entity, the number of corresponding triples is not certain, that is, the corresponding triple may not be found in the knowledge graph. For example, the named entity "Chang'an Street" does not have a corresponding triple in the knowledge graph. For another example, the named entity "Zhang San" has two corresponding triples in the knowledge graph - (Wang Wu, insurance, Zhang San) and (Zhang San, insured person, policy number 123).

[0072] Step 13: input the text data and the triplet information containing the named entity into a trained fraud risk prediction model to obtain a fraud risk prediction result for the target case output by the fraud risk prediction model.

[0073] The fraud risk prediction model of this embodiment is an anti-fraud prediction model that integrates a large language model and knowledge graph embedding. Specifically, the fraud risk prediction model may include a knowledge graph enhanced language sub-model and a classification sub-model.

[0074] Among them, the knowledge graph enhanced language sub-model determines the triple vector representing the semantics of the target case based on text data and triple information; the classification sub-model determines the fraud risk prediction result of the target case based on the triple vector.

[0075] In one implementation, the features input to the classification sub-model also include numerical features of the target case and categorical features of the target case.

[0076] Among them, the numerical features and categorical features of the target case can be obtained from the text data describing the target case.

[0077] The text data, triple information containing named entities in the text data, numerical features of the target case and categorical features of the target case are all used to predict fraud risks, enriching the feature dimensions.

[0078] In this embodiment, the text data describing the target case and the triple information containing the named entities in the text data are input into the fraud risk prediction model to predict the fraud risk of the target case, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed by credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model is credible information related to the target case, when the fraud risk is predicted based on the triple information and the text data describing the target case, the accuracy of the prediction result can be improved.

[0079] Embodiment 2 of the present application provides a model training method based on the fraud risk prediction method provided in the above embodiments.

[0080] The executor of the method may be any computing device that can implement the method, such as a server, a mobile phone, a personal computer, a smart wearable device, a smart robot, and the like.

[0081] The different steps of the method may be implemented by the same execution subject or by different execution subjects. The embodiment of the present application does not limit the execution subject used to implement the method.

[0082] In addition, the embodiments of the present application do not limit the execution order of different steps. When using the method provided in the embodiments of the present application, the execution order of different steps can be adjusted according to actual needs.

[0083] For ease of description, the method provided in the embodiment of the present application is described in detail below by taking a model training device as the executor of the method as an example.

[0084] The model training method may specifically include the following: Figure 2 The following steps 21 to 25 are shown:

[0085] Step 21, obtaining text data for describing sample cases.

[0086] The sample case may be a historical insurance claim case. In addition to the text data content of the target case in the first embodiment, the text data of the sample case usually also includes the claim result. The claim result includes whether the claim is settled: if the claim is settled, the claim result also includes the corresponding liability determination result, underwriting information, and claim result; if the claim is not settled, the claim result also includes the reason for not settling the claim (such as this case is an insurance fraud case, does not meet the claim conditions, falls under the exemption clause, etc.), supporting documents and underwriting information supporting non-claim.

[0087] Step 22: Based on the named entities in the text data, triple information containing the named entities is obtained from a pre-constructed knowledge graph.

[0088] Identify named entities in text data describing sample cases, query a pre-built knowledge graph based on the identified named entities, and obtain triple information containing the identified named entities from the knowledge graph.

[0089] The knowledge graph is constructed using trusted information that can be used to identify insurance fraud, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0090] The method of identifying named entities in text data describing sample cases can be achieved through existing rule matching methods, statistical model-based recognition methods, and deep learning-based methods.

[0091] The rule matching method can be implemented by matching named entities in the text through predefined rules or patterns and labeling or classifying them. It is usually applicable to named entity types whose entity types or their expressions are relatively regular and recognizable, such as date, time, etc.

[0092] The implementation of recognition based on statistical models can be to regard the recognition problem of named entities as a sequence labeling problem, use statistical models such as hidden Markov models or conditional random field models for supervised training, and recognize named entities in the target text data based on the trained statistical models.

[0093] The implementation of the deep learning-based method can be to label the named entities in the sample text data, use a deep neural network (such as a convolutional neural network, a recurrent neural network) for supervised training, and recognize the named entities in the target text data based on the trained deep neural network.

[0094] Step 23, using the text data and the triple information containing the named entity as input to the fraud risk prediction model to be trained, and using the fraud results of the sample case as the actual results of the fraud risk prediction model.

[0095] The fraud risk prediction model is an anti-fraud prediction model that integrates a large language model and knowledge graph embedding. Specifically, the fraud risk prediction model may include a knowledge graph enhanced language sub-model and a classification sub-model.

[0096] Among them, the knowledge graph enhanced language sub-model determines the triple vector representing the semantics of the target case based on text data and triple information; the classification sub-model determines the fraud risk prediction result of the target case based on the triple vector.

[0097] When training the fraud risk prediction model, the knowledge graph enhanced language sub-model and the classification sub-model are trained jointly, that is, the text data and the triple information containing the named entities in the text data are input into the knowledge graph enhanced language sub-model, and the triple vector output by the knowledge graph enhanced language sub-model is input into the classification sub-model. The output results of the classification sub-model are compared with the actual results to adjust the parameters of the knowledge graph enhanced language sub-model and the classification sub-model.

[0098] The knowledge graph enhanced language sub-model can be obtained by combining the knowledge graph embedding model (KGE) with a pre-trained LM (Language Model) model. The LM model can be a BERT (Bidirectional Encoder Representations from Transformers) model, which is a pre-trained language model based on the Transformer architecture. The combination of the knowledge graph embedding model and the LM model can be collectively referred to as a K-BERT-type model.

[0099] The classification sub-model may be a DNN (Deep Neural Network) model, such as an MLP (Multilayer Perceptron) model.

[0100] Step 24, in each round of training iteration, the text data describing the sample case and the triple information containing the named entities in the text data are input into the fraud risk prediction model to be trained, and the value of the loss function is determined based on the output results of the fraud risk prediction model to be trained in each round of iteration and the actual results.

[0101] like Figure 3 The figure shows an example of a sentence tree formed based on the text data describing the case and the triple information containing the named entities in the text data. The knowledge graph enhanced language sub-model can form a sentence tree based on the input text data and the triple information containing the named entities in the text data. Figure 3 The form of the sentence tree shown.

[0102] In this sentence tree, "The vehicle with license plate number Beijing A11111 driven by Zhang San collided with the vehicle with license plate number Beijing B22222 driven by Li Si on Chang'an Avenue" is the text data describing the case, which serves as the trunk of the sentence tree. The triple information obtained from the knowledge graph (such as: (Wang Wu, insured, Zhang San), (Zhang San, insured person, policy number 123), (Li Si, insured, Wang Wu), (Wang Wu, car owner, Beijing A11111) and (Li Si, car owner, Beijing B22222) etc.) serves as the branches of the sentence tree.

[0103] The knowledge graph enhanced language sub-model converts the above sentence tree into the corresponding embedding representation. Specifically, each word segment or fragment in the sentence tree is tokenized (the smallest unit into which text data is segmented, which can be a word, punctuation mark, Chinese word segment, etc. Figure 4 Each token in the sentence tree is represented by a dotted box) embedding, that is, the trunk and branches of the sentence tree are soft-position encoded. Soft position encoding means that the position encoding of the relationship and tail entity in the branch is the same as the position encoding of the two tokens after the head entity in the trunk. It should be noted that the above-mentioned head entity, relationship, and tail entity are a representation of a triple, that is, a triple is composed of a head entity, a relationship, and a tail entity, such as Figure 4 In the example, (Zhang San, insured person, policy number) correspond to (head entity, relationship, tail entity) respectively. Figure 4 As shown, the position code of "Zhang San" in the trunk of the sentence tree is 1, and the position codes of the two tokens after "Zhang San" in the trunk are 2 and 3 respectively. The relationship in the branch corresponding to the triplet (Zhang San, insured person, policy number) - the insured person, and the tail entity - the policy number, have corresponding position codes 2 and 3.

[0104] in addition, Figure 4 The token encoded as 0 in the text box, [CLK], indicates the beginning of a sentence.

[0105] After performing embedding representation on the sentence tree, in order to enable the model to more accurately learn the relationships between tokens in the sentence tree, facilitate better learning of the semantics of the embedded knowledge, and not affect the semantics of the main sentence, the knowledge graph enhanced language sub-model introduces a visibility matrix to represent the visibility between tokens. The definition of the visibility matrix is shown in the following formula (1):

[0106]

[0107] In formula (1), indicates that two tokens are on the same branch (or main trunk), indicates that two tokens are not on the same branch (or main trunk), where i and j represent the labels of the tokens in the sentence tree, and this label is unique, that is, there are no tokens with the same label in the entire sentence tree. For Figure 4 the sentence tree shown above, for the token "driving" with a soft position number of 2 and the token "policy number" with a soft position number of 3, which are not on the same main trunk (or branch), the corresponding value in the visibility matrix is -∞; for the token "driving" with a soft position number of 2 and the token "of" with a soft position label of 3, which are on the same main trunk, the corresponding values of the two in the visibility matrix are 0.

[0108] Based on the constructed visibility matrix, the structural information of the sentence tree is obtained. When performing vector representation of the input text data according to this visibility matrix, it is possible to learn the embedded knowledge (triple information) without interfering with the original text data (the main trunk in the sentence tree).

[0109] The knowledge graph enhanced sub-language model, based on the constructed visibility matrix, uses a Mask-Transformer encoder to generate triple vectors. The Mask-Transformer encoder can limit the visible area of the self-attention mechanism according to the visibility matrix, enabling the tokens in the branch to act more as the existence enriching the corresponding tokens on the main trunk without directly affecting other tokens on the main trunk, thus alleviating the knowledge noise problem.

[0110] So far, the principle of knowledge embedding of the knowledge graph enhanced language sub-model has been introduced.

[0111] In one implementation, the determination of the loss function of the fraud risk prediction model specifically includes the following steps 241 to step 242:[[]]END]

[0112] Step 241, based on the triple vector, determine the scoring function of the fraud risk prediction model.

[0113] The triple vector output by the knowledge graph enhanced language sub-model is represented as (h, r, t), where h, r, and t represent the head entity vector, relation vector, and tail entity vector, respectively.

[0114] The scoring function f(h, r, t) is expressed as shown in the following formula (2):

[0115]

[0116] In formula (2), b is a constant and can be set manually.

[0117] Step 242: Determine the loss function according to the scoring function.

[0118] According to the scoring function, the probability model shown in the following formula (3) is determined:

[0119]

[0120] Among them, Prh|r, t) represents the probability that the head entity vector is h in the obtained triple vector when the relation vector and the tail entity vector are r and t respectively; Indicates the different head entities randomly sampled in the obtained triple vector (there are multiple elements in the head entity vector h, represents a randomly selected element). Based on the above formula (3), the corresponding Prr|h, t) and Prt|h, r) can also be obtained.

[0121] According to the above probability model, the expression of the loss function is determined as shown in the following formula (4):

[0122] L=-∑ (h,r,t)∈G (logPr h|r, t)+logPr r|h, t) +logPr t|h, r)) (4)

[0123] In formula (4), G is the set of triple vectors output by the knowledge graph enhanced language sub-model.

[0124] The output of the classification sub-model can be in the form of a prediction score.

[0125] In one implementation, the triplet vector is input into the classification sub-model to obtain a prediction result of the fraud risk in this iteration, including:

[0126] The triplet vector, the numerical features of the sample cases and the categorical features of the sample cases are input into the classification sub-model to obtain the prediction result of the fraud risk in this round of iteration.

[0127] Among them, the numerical features of the sample cases and the categorical features of the sample cases can be extracted from the text data and encoded to form the corresponding numerical features and categorical features. The encoding of numerical data and categorical data can be achieved through the encoding layer of the knowledge graph enhanced language sub-model.

[0128] Step 25, based on the principle of minimizing the value of the loss function, adjust the parameters of the fraud risk prediction model to obtain a trained fraud risk prediction model.

[0129] The model training method provided in this embodiment inputs the text data describing the sample case and the triple information containing the named entities in the text data into the fraud risk prediction model to be trained, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model to be trained is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed by credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model to be trained is credible information related to the sample case, so when the fraud risk is predicted based on the triple information and the text data describing the sample case, the accuracy of the prediction result can be improved.

[0130] In order to solve the problem of low prediction accuracy in fraud risk prediction using existing technologies, based on the same inventive concept as the above-mentioned embodiments of the present application, embodiment three of the present application provides a fraud risk prediction device.

[0131] The specific structural diagram of the device is shown in Figure 5 As shown, it includes the following functional units 51-53:

[0132] A text data acquisition unit 51, which acquires text data for describing a target case;

[0133] A triple information acquisition unit 52, based on the named entity in the text data, acquires triple information containing the named entity from a pre-constructed knowledge graph;

[0134] The risk prediction unit 53 is used to input the text data and the triple information containing the named entity into the trained fraud risk prediction model to obtain the fraud risk prediction result for the target case output by the fraud risk prediction model.

[0135] The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0136] The trusted information used in constructing the knowledge graph includes at least historical policy information, historical accident information, and historical claims information.

[0137] The risk prediction unit 53 is also used to input the text data, the triple information of the named entities in the text data, the numerical features of the target case and the categorical features of the target case into a trained fraud risk prediction model to obtain the fraud risk prediction result for the target case output by the fraud risk prediction model.

[0138] The above-mentioned fraud risk prediction device provided by this embodiment inputs the text data describing the target case and the triple information containing the named entities in the text data into the fraud risk prediction model to predict the fraud risk of the target case, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed by credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model is credible information related to the target case, when the fraud risk is predicted based on the triple information and the text data describing the target case, the accuracy of the prediction result can be improved.

[0139] Embodiment 4 of the present application provides a model training device based on the model training method provided in the above embodiment 2.

[0140] like Figure 6 As shown, the model training device may specifically include the following functional units 61-65:

[0141] A text data acquisition unit 61, which acquires text data for describing a sample case;

[0142] A triple information acquisition unit 62, based on the named entity in the text data, acquires triple information containing the named entity from a pre-constructed knowledge graph;

[0143] An information setting unit 63, configured to use the text data and the triple information containing the named entity as inputs of a fraud risk prediction model to be trained, and use the fraud results of the sample cases as actual results of the fraud risk prediction model;

[0144] The iterative training unit 64 inputs the text data describing the sample case and the triple information containing the named entities in the text data to the fraud risk prediction model to be trained, and determines the value of the loss function based on the output result of the fraud risk prediction model to be trained in each round of iteration and the actual result;

[0145] The parameter adjustment unit 65 adjusts the parameters of the fraud risk prediction model based on the principle of minimizing the value of the loss function to obtain a trained fraud risk prediction model;

[0146] The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

[0147] The fraud risk prediction model includes a knowledge graph enhanced language sub-model and a classification sub-model.

[0148] The information setting unit 63 is further used to input the triplet vector, the numerical features of the sample case and the categorical features of the sample case into the classification sub-model to obtain the prediction result of the fraud risk in this round of iteration.

[0149] In each round of training iteration, the iterative training unit 64 inputs the text data describing the sample case and the triple information containing the named entities in the text data into the fraud risk prediction model to be trained, and is specifically used to: input the text data and the triple information containing the named entities in the text data into the language sub-model to obtain a triple vector representing the sample case; input the triple vector into the classification sub-model to obtain the prediction result of the fraud risk in this round of iteration.

[0150] The iterative training unit 64 determines the loss function, specifically including:

[0151] Based on the triple vector, a scoring function of the fraud risk prediction model is determined; and according to the scoring function, the loss function is determined.

[0152] The above-mentioned model training device provided in this embodiment inputs the text data describing the sample case and the triple information containing the named entities in the text data into the fraud risk prediction model to be trained, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model to be trained is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed by credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model to be trained is credible information related to the sample case, so when predicting the fraud risk based on the triple information and the text data describing the sample case, the accuracy of the prediction result can be improved.

[0153] Based on the same inventive concept as the aforementioned embodiments of the present application, the embodiments of the present application also provide a computing device.

[0154] like Figure 7As shown, the computing device includes: a memory 71 and a processor 72. The memory 71 can be configured to store various other data to support operations on the electronic device. Examples of these data include instructions for any application or method for operating on the electronic device. The memory 71 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

[0155] The processor 72 is coupled to the memory 71 and is used to execute the computer program stored in the memory 71 to execute the fraud risk prediction method described in the aforementioned embodiment 1, or the model training method described in embodiment 2 of the present application.

[0156] When the processor 72 executes the computer program to perform the fraud risk prediction method, the text data describing the target case and the triple information containing the named entities in the text data are input into the fraud risk prediction model to predict the fraud risk of the target case, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed by credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model is credible information related to the target case, so when the fraud risk is predicted based on the triple information and the text data describing the target case, the accuracy of the prediction result can be improved.

[0157] When the processor 72 executes the computer program to execute the model training method, the text data describing the sample case and the triple information containing the named entities in the text data are input into the fraud risk prediction model to be trained, thereby enriching the features used for fraud risk prediction. Since the triple information input into the fraud risk prediction model to be trained is in the knowledge graph, containing the triple information of the named entities in the text data, and the knowledge graph is constructed by credible information that can be used for insurance fraud identification, that is, the triple information input into the fraud risk prediction model to be trained is credible information related to the sample case, so when the fraud risk is predicted based on the triple information and the text data describing the sample case, the accuracy of the prediction result can be improved.

[0158] When the processor 72 executes the computer program in the memory 71, in addition to the above functions, it can also realize other functions, and the details can be found in the description of the previous embodiments.

[0159] Further, such as Figure 7 As shown, the computing device also includes: a display 74, a communication component 73, a power component 75, an audio component 76 and other components. Figure 7 Only some components are shown schematically, and it does not mean that the computing device only includes Figure 7 Components shown.

[0160] Accordingly, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, and when the computer program is executed by a computer, the methods provided in the above embodiments can be implemented.

[0161] An embodiment of the present application also provides a computer program product, including a computer program, which implements the methods provided in the above embodiments when executed by a processor.

[0162] The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a processor, the methods provided in the above embodiments are implemented.

[0163] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0164] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A fraud risk prediction method, characterized in that: include: Obtain text data for describing the target case; Based on the named entities in the text data, obtaining triple information containing the named entities from a pre-built knowledge graph; Inputting the text data and the triple information containing the named entity into a trained fraud risk prediction model to obtain a fraud risk prediction result for the target case output by the fraud risk prediction model; The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

2. The method according to claim 1, characterized in that The credible information used in constructing the knowledge graph includes at least historical policy information, historical accident information, and historical claims information.

3. The method according to claim 1, characterized in that The method further comprises: The text data, triple information containing named entities in the text data, numerical features of the target case and categorical features of the target case are input into a trained fraud risk prediction model to obtain a fraud risk prediction result for the target case output by the fraud risk prediction model.

4. A model training method, characterized in that: include: Obtain text data for describing sample cases; Based on the named entities in the text data, obtaining triple information containing the named entities from a pre-built knowledge graph; Using the text data and the triple information containing the named entity as input of the fraud risk prediction model to be trained, and using the fraud results of the sample case as the true results of the fraud risk prediction model; In each round of training iteration, the text data describing the sample case and the triple information containing the named entities in the text data are input into the fraud risk prediction model to be trained, and the value of the loss function is determined based on the output result of the fraud risk prediction model to be trained in each round of iteration and the actual result; Based on the principle of minimizing the value of the loss function, the parameters of the fraud risk prediction model are adjusted to obtain a trained fraud risk prediction model; The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

5. The method according to claim 4, characterized in that The fraud risk prediction model includes a knowledge graph enhanced language sub-model and a classification sub-model; In each iteration of the training, the text data describing the sample case and triple information containing the named entities in the text data are input into the fraud risk prediction model to be trained, including: Inputting the text data and triple information containing named entities in the text data into the language sub-model to obtain a triple vector representing the sample case; The triplet vector is input into the classification sub-model to obtain the prediction result of the fraud risk in this round of iteration.

6. The method according to claim 5, characterized in that The triplet vector is input into the classification sub-model to obtain the prediction result of the fraud risk in this iteration, including: The triplet vector, the numerical features of the sample cases and the categorical features of the sample cases are input into the classification sub-model to obtain the prediction result of the fraud risk in this round of iteration.

7. The method according to claim 5, characterized in that The loss function is determined in the following way: Based on the triplet vector, determining a scoring function of a fraud risk prediction model; The loss function is determined according to the scoring function.

8. A fraud risk prediction device, characterized in that: include: A text data acquisition unit, which acquires text data used to describe a target case; A triple information acquisition unit, based on the named entities in the text data, acquires triple information containing the named entities from a pre-constructed knowledge graph; A risk prediction unit, configured to input the text data and the triplet information containing the named entity into a trained fraud risk prediction model to obtain a fraud risk prediction result output by the fraud risk prediction model for the target case; The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

9. A model training device, characterized in that: include: A text data acquisition unit, which acquires text data used to describe the sample case; A triple information acquisition unit, based on the named entities in the text data, acquires triple information containing the named entities from a pre-constructed knowledge graph; An information setting unit, used to use the text data and the triple information containing the named entity as inputs of the fraud risk prediction model to be trained, and use the fraud results of the sample cases as the actual results of the fraud risk prediction model; The iterative training unit inputs the text data describing the sample case and the triple information containing the named entities in the text data into the fraud risk prediction model to be trained, and determines the value of the loss function based on the output result of the fraud risk prediction model to be trained in each round of iteration and the actual result; The parameter adjustment unit adjusts the parameters of the fraud risk prediction model based on the principle of minimizing the value of the loss function to obtain a trained fraud risk prediction model; The knowledge graph is constructed using trusted information that can be used for insurance fraud identification, and the triple information of the knowledge graph includes: two named entities in the trusted information and a predicate representing the semantic relationship between the two named entities.

10. A computing device, characterized in that include: Memory and processor, wherein: The memory is used to store computer programs; The processor is coupled to the memory and is used to execute the computer program stored in the memory to execute the method according to any one of claims 1 to 7.

11. A computer-readable storage medium storing a computer program, wherein the computer program can implement the method according to any one of claims 1 to 7 when executed by a computer.

12. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 7 when being executed by a processor.

Citation Information

Patent Citations

  • Vehicle insurance claim settlement fraud risk identification method and system based on knowledge graph

    CN113837886A

  • Text classification method and device based on semantics, computer equipment and storage medium

    CN117874234A

  • Anti-fraud risk assessment method and apparatus, training method and apparatus, and readable storage medium

    WO2023124204A1