Method and device for mining insurance responsibility judgment rule

Abstract

The invention provides a method and device for mining insurance responsibility judgment rules. The method relates to the technical field of data processing, and comprises the following steps: a large responsibility judgment model performs responsibility judgment on cost items of a case included in a responsibility judgment data set of a target batch based on a responsibility judgment rule set to obtain a responsibility judgment result and compares the responsibility judgment result with a responsibility judgment label; if the comparison is different, error feedback is generated; the rule combing large model corrects the responsibility judgment rule set based on error feedback, randomly samples output, and determines and updates an optimal responsibility judgment rule set according to a reward function; taking the next batch of responsibility judgment data set as a new target batch of responsibility judgment data set, and repeating the above steps until iteration training of the target number of batches of responsibility judgment data sets is completed; and updating the rule combing large model based on construction of the preference data set and a direct preference optimization algorithm, and repeating the steps until a target number of cycles are completed to obtain a final responsibility judgment rule set. The responsibility judgment rule formed in this way can be understood by the large responsibility judgment model and can be conveniently used for responsibility judgment of actual cases.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and device for mining insurance liability judgment rules. Background Art

[0002] A big model refers to a machine learning model with a large number of parameters and complex structures, which can understand human instructions, process massive amounts of data, and complete complex tasks, such as natural language processing, computer vision, and speech recognition. However, there are still some problems in applying big models to insurance liability judgments. For example, the description of the exemption clauses in insurance policies is often not clear enough and has a coarse granularity, and there is a lack of clear liability judgment rules for the big model. The liability judgment rules formed in actual liability judgments have the recognized experience of audit experts, but they have not been sorted out and written down due to the large number of insurance policies. Summary of the invention

[0003] In view of this, the present application provides a method and device for mining insurance liability judgment rules, aiming to mine the judgment standard rules that form a large model understanding.

[0004] In a first aspect, the present application provides a method for mining insurance liability rules, the method comprising:

[0005] The large accountability model, based on the accountability rule set, conducts accountability on the expense items of the cases included in the accountability data set of the target batch, obtains the accountability results and compares them with the accountability labels in the accountability data set;

[0006] If there is a discrepancy in the alignment, an error feedback is generated;

[0007] The rule combing big model executes the rule combing instruction to modify the judgment rule set based on the error feedback, and randomly samples the output of the rule combing big model, calculates the reward value of each output in combination with the reward function, determines the optimal judgment rule set and updates it;

[0008] The next batch of judgment data sets are used as the new target batch of judgment data sets, and the judgment model is executed based on the judgment rule set and subsequent steps until the iterative training of the target number of batches of judgment data sets is completed;

[0009] A preference data set for this round of iterative training is constructed based on the randomly sampled outputs and reward values ​​corresponding to each rule combing instruction in this round of iterative training. Based on the preference data set and the direct preference optimization algorithm, the rule combing large model is trained and updated, and the accountability large model is executed based on the accountability rule set and subsequent steps until the target number of cycles are completed to obtain the final accountability rule set.

[0010] Optionally, the method further includes:

[0011] Initializing and constructing the liability judgment rule set according to the exemption clause of the insurance policy, wherein the liability judgment rule set includes a plurality of liability judgment rules and a number of each liability judgment rule;

[0012] Obtaining a judgment data set, wherein the judgment data set includes case information and judgment labels, wherein the judgment labels are judgment information marked for each expense item of each case, and the judgment information includes the expense item, a judgment on whether to judge, and a judgment reason for violating the number of the judgment rule in the judgment rule set;

[0013] The judgment data set is divided into a plurality of batches.

[0014] Optionally, if there is a difference in the comparison, an error feedback is generated, including:

[0015] When the accountability model judges the target cost item as violating the accountability rule No. X, and the accountability label of the target cost item is that the target cost item is not to be judged, generating feedback that the target cost item does not violate the accountability rule No. X and should not be judged;

[0016] When the accountability model does not judge the target cost item, but the accountability label of the target cost item is that the target cost violates the Xth judgment rule, feedback is generated that the target cost item violates the Xth judgment rule and should be judged, and the Xth judgment rule is one of the judgment rules included in the judgment rule set.

[0017] Optionally, the method further includes:

[0018] If there is no difference in the comparison, the step of using the next batch of judgment data sets as the new target batch of judgment data sets and subsequent steps are performed.

[0019] Optionally, the rule combing model modifies the accountability rule set based on the error feedback, and determines the optimal accountability rule set by randomly sampling the output of the rule combing model and calculating the reward value of each output in combination with the reward function, including:

[0020] Setting the hyperparameters of the rule combing model, and constructing rule combing instructions in combination with the target batch of judgment data sets, the judgment rule set, the judgment results and the error feedback;

[0021] Input the rule combing instruction M times into the rule combing big model to obtain M outputs of the rule combing big model;

[0022] The reward function is used to calculate the reward value of each output in the M outputs, and the output with the highest reward value is taken as the optimal judgment rule set.

[0023] Optionally, the reward value of the Nth output among the M outputs is calculated using a reward function, including:

[0024] Based on the Nth output, the accuracy of the judgment of the judgment data validation set is used as the task reward R(task) for the Nth output. N , the judgment data verification set includes case information and judgment labels;

[0025] The inverse of the ratio of the length of the Nth output to the longest output length among the M outputs is used as the generated length reward R (token) for the Nth output. N ;

[0026] The inverse of the language model loss of the large judgment model for the Nth output is used as the language fluency reward R(loss) for the Nth output. N ;

[0027] According to the task reward R(task) N , Generate length reward R (token) N and language fluency reward R(loss) N , calculate the reward value of the Nth output through the reward function;

[0028] The reward function of the Nth output is:

[0029] Reward function (Nth output) = R(task) N +λ(token)R(token) N +λ(loss)R(loss) N ,

[0030] Among them, λ(token) and λ(loss) are both hyperparameters.

[0031] Optionally, constructing a preference data set for this round of iterative training according to the randomly sampled output and reward value corresponding to each rule combing instruction in this round of iterative training includes:

[0032] Calculate the difference between the output with the highest reward value and the output with the lowest reward value in the randomly sampled output corresponding to each rule combing instruction in this round of iterative training;

[0033] If the difference is greater than a threshold, the output with the highest reward value, the output with the lowest reward value, and the rule combing instruction are taken as a piece of preference data in the preference data set.

[0034] Optionally, the method further includes:

[0035] The case information to be judged and the final judgment rule set are used to construct a judgment task instruction, which is input into the judgment big model to obtain the judgment result of the case information to be judged.

[0036] In a second aspect, the present application provides a device for mining insurance liability rules, the device comprising:

[0037] A judgment unit, which is used for judging the cost items of the cases included in the judgment data set of the target batch based on the judgment rule set by the judgment big model, obtaining the judgment result and comparing it with the judgment label in the judgment data set;

[0038] a generating unit, configured to generate an error feedback if a difference is found in the comparison;

[0039] An updating unit, used for the rule combing big model to modify the judgment rule set based on the error feedback, by randomly sampling the output of the rule combing big model, calculating the reward value of each output in combination with the reward function, and determining and updating the optimal judgment rule set;

[0040] A processing unit, used to use the next batch of judgment data sets as the new target batch of judgment data sets, and execute the judgment model based on the judgment rule set and subsequent steps until the iterative training of the target number of batches of judgment data sets is completed;

[0041] The optimization unit is used to construct a preference data set for this round of iterative training according to the output and reward value of random sampling corresponding to each batch of the judgment data set in this round of iterative training, train and update the rule combing model based on the preference data set and the direct preference optimization algorithm, and execute the judgment model based on the judgment rule set and subsequent steps until the target number of cycles is completed to obtain the final judgment rule set.

[0042] Optionally, the generating unit is further configured to, if there is no difference in the comparison, execute the step of using the next batch of judgment data sets as the new target batch of judgment data sets and the subsequent steps.

[0043] In a third aspect, the present application provides a device comprising a memory and a processor, wherein the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes a method for mining insurance liability rules as described in any one of the first aspects.

[0044] In a fourth aspect, the present application provides a computer storage medium having codes stored therein. When the codes are executed, a device executing the codes implements a method for mining insurance liability rules as described in any one of the first aspects above.

[0045] The present application provides a method for mining insurance liability rules. When executing the method, the liability model, based on the liability rule set, makes a judgment on the expense items of the cases included in the liability data set of the target batch, obtains the judgment result and compares it with the liability label in the liability data set; if there is a difference in the comparison, an error feedback is generated; the rule combing model modifies the liability rule set based on the error feedback, randomly samples the output of the rule combing model, calculates the reward value of each output in combination with the reward function, determines the optimal liability rule set and updates it; takes the next batch of liability data sets as the new target batch of liability data sets, and executes the liability model based on the liability rule set and subsequent steps until the iterative training of the target number of batches of liability data sets is completed; constructs the preference data set of this round of iterative training according to the output and reward value of the random sampling corresponding to each batch of liability data sets in this round of iterative training, trains and updates the rule combing model based on the preference data set and the direct preference optimization algorithm, and executes the liability model based on the liability rule set and subsequent steps until the target number of rounds of cycles are completed to obtain the final liability rule set. The expense items of the cases in the accountability data set are judged through the accountability rule set based on the accountability big model, and the accountability rule set is corrected through the rule combing big model to achieve multiple rounds and batches of iterative training of the accountability data set. At the same time, each round of rule combing big model will be trained and updated through the preference data set and the direct preference algorithm. In this way, a set of accountability rules that is consistent with the actual case accountability can be combed out, and the accountability rules are mined and combed into text through the big model. The formed accountability rules can be understood by the accountability big model and used for the judgment of actual cases, while greatly saving the workload of manual combing. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 A flowchart of a method for mining insurance liability rules provided in an embodiment of the present application;

[0048] Figure 2 A schematic diagram of a specific application process of a method for mining insurance liability judgment rules provided in an embodiment of the present application;

[0049] Figure 3 A schematic diagram of the structure of a device for mining insurance liability rules provided in an embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to be able to understand the features and technical contents of the embodiments of the present disclosure in more detail, the implementation of the embodiments of the present disclosure is described in detail below in conjunction with the accompanying drawings. The attached drawings are for reference only and are not used to limit the embodiments of the present disclosure. In the following technical description, for the convenience of explanation, a full understanding of the disclosed embodiments is provided through multiple details. However, one or more embodiments can still be implemented without these details. In other cases, to simplify the drawings, well-known structures and devices can be simplified for display.

[0051] The terms "first", "second", etc. in the specification and claims of the embodiments of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged where appropriate, so as to describe the embodiments of the embodiments of the present disclosure described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions.

[0052] Unless otherwise stated, the term "plurality" means two or more.

[0053] In the embodiment of the present disclosure, the character " / " indicates that the preceding and following objects are in an "or" relationship. For example, A / B indicates: A or B.

[0054] The term "and / or" is a description of the association relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or, A and B.

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. 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 creative work are within the scope of protection of this application.

[0056] See also Figure 1 , Figure 1 A flowchart of a method for mining insurance liability rules provided in an embodiment of the present application, a method for mining insurance liability rules, the method comprising:

[0057] S101. The large accountability model judges the expense items of the cases included in the accountability data set of the target batch based on the accountability rule set, obtains the accountability results and compares them with the accountability labels in the accountability data set.

[0058] The liability rule set includes multiple liability rules and the number of each liability rule. For example, the format of the liability rule set can be: ["liability rule 1": "xxx", "liability rule 2": "xxx", ...]. The initial liability rule set can be set according to the original text of the exemption clause of the insurance policy.

[0059] The accountability data set includes historical case information (which may include case information of multiple cases) and manually annotated accountability labels. The accountability labels are accountability information annotated for each expense item of each case. The accountability information includes the expense item, the judgment on whether to be accountable, and the reason for the violation of the accountability rule in the accountability rule set. The format of the accountability label can be: [{"expense item":"xxx","whether to be accountable":"yes","reason for accountability":"violation of the i-th accountability rule"}], where the value range of i is the number of the accountability rule in the accountability rule set.

[0060] The judgment result is the result of judging the cost item of each case in the judgment data, including whether a judgment is made and which judgment rule in the judgment rule set is violated when the judgment is made.

[0061] S102: If there is a difference in the comparison, an error feedback is generated.

[0062] Error feedback refers to the difference between the accountability results of the accountability model for each cost item in each case and the accountability information of the accountability label corresponding to the cost item in the accountability data set.

[0063] In addition, if there is no difference in the comparison, the following step S104 is directly executed.

[0064] S103, the rule combing big model executes the rule combing instruction to modify the accountability rule set based on the error feedback, and randomly samples the output of the rule combing big model, calculates the reward value of each output in combination with the reward function, determines the optimal accountability rule set and updates it.

[0065] The rule combing instruction is an instruction constructed by combining the judgment data set of the target batch, the judgment rule set, the judgment result and the error feedback. The rule combing big model executes the rule combing instruction and modifies the judgment rule set according to the error feedback.

[0066] Exemplarily, the above-mentioned rule combing instructions are constructed by setting the hyperparameters of the rule combing big model, combining the judgment data set of the target batch, the judgment rule set, the judgment result and the error feedback. Then, the rule combing instructions are input into the rule combing big model M times to obtain M outputs of the rule combing big model, that is, the output of the rule combing big model is randomly sampled M times. Finally, the reward value of each output in the M outputs is calculated by the reward function, and the output with the highest reward value can be used as the optimal judgment rule set.

[0067] It should be noted that the above-mentioned accountability model and rule combing model are two independent instances of the same model, which reduces the differences between the two models in analysis content and language habits, and the parameters of the rule combing model will be trained and iterated in the subsequent step S105. Therefore, the rule combing model and the accountability model cannot share the same instance.

[0068] The above-mentioned judgment big model is used to complete the judgment task of case information in the judgment data set of the target batch based on the judgment rule set, and the above-mentioned combing big model is used to complete the combing and correction of the judgment rules in the judgment data set according to the judgment results and error feedback.

[0069] S104: Use the next batch of judgment data sets as the new target batch of judgment data sets, and execute the judgment model based on the judgment rule set and subsequent steps until the iterative training of the target number of batches of judgment data sets is completed.

[0070] The above steps S101-S104 are executed on the accountability data set of each batch in the multiple batches, so as to analyze the inaccuracies and imperfections of the accountability rules in the accountability rule set according to the error feedback through the rule combing large model and correct the accountability rules.

[0071] S105. Construct a preference data set for this round of iterative training based on the randomly sampled output and reward value corresponding to each rule combing instruction in this round of iterative training. Based on the preference data set and the direct preference optimization algorithm, train and update the rule combing big model, and execute the judgment big model based on the judgment rule set and subsequent steps until the target number of cycles are completed to obtain the final judgment rule set.

[0072] This application executes the above steps S101-S104 on each batch of judgment data sets in multiple batches to complete a round of training iterations of multiple batches of judgment data sets. Then, according to the multiple outputs and reward values ​​of each rule combing instruction executed by the rule combing big model in this round of iteration, the preference data is determined to form a preference data set. The rule combing big model is trained by the preference data set and the direct preference algorithm in each round, so that the output of the rule combing big model is gradually further aligned with our expectations.

[0073] Based on the above steps S101-S105, it can be known that the present application judges the expense items of the cases in the accountability data set through the accountability rule set based on the accountability big model, and corrects the accountability rule set through the rule combing big model, so as to realize multiple rounds and batches of iterative training of the accountability data set. At the same time, each round of the rule combing big model will be trained and updated through the preference data set and the direct preference algorithm. In this way, the accountability rules can be mined and combed into a document through the big model, and a set of accountability rules that are consistent with the actual case accountability can be combed out. The formed accountability rules can be understood by the accountability big model and can be used for the actual case accountability, while greatly saving the workload of manual combing.

[0074] In a specific implementation, if there is a difference in the comparison, an error feedback is generated, and the specific steps may include:

[0075] When the accountability model judges the target cost item as violating the accountability rule No. X, and the accountability label of the target cost item is that the target cost item is not to be judged, generating feedback that the target cost item does not violate the accountability rule No. X and should not be judged;

[0076] When the accountability model does not judge the target cost item, but the accountability label of the target cost item is that the target cost violates the Xth judgment rule, feedback is generated that the target cost item violates the Xth judgment rule and should be judged, and the Xth judgment rule is one of the judgment rules included in the judgment rule set.

[0077] See also Figure 2 , Figure 2 A specific application flow chart of a method for mining insurance liability rules is provided. Figure 1 In the embodiment shown, let the total number of iterations be T and the current round be t. Let there be N pieces of judgment data, which are divided into n batches to form a judgment data set of n batches, and the current batch is i. We record the state of the i-th batch in the t-th round as (t, i). In a specific execution example of a method for mining insurance judgment rules, the iterative steps from data batch i to data batch i+1 in the t-th round can be:

[0078] S201. Use the judgment data set (t, i) and the judgment rule set (t, i) to construct a judgment task instruction, recorded as judgment task prompt (t, i), and input it into the judgment model.

[0079] The judgment task instruction requires the big model to judge each cost item of each case in the judgment data set (t,i) and output the judgment result (t,i).

[0080] S202. Compare the judgment result (t, i) output by the judgment model with the judgment label in the judgment data set (t, i).

[0081] S203: If there is a difference in the comparison, an error feedback (t, i) is generated.

[0082] The first type of comparison error is: the judgment model judges the expense item A, which violates the judgment rule No. x, while the actual expense item A in the judgment label should not be judged. Generate error feedback: Expense item A does not violate the judgment rule No. x and should not be judged.

[0083] The second type of comparison error is: the accountability model does not judge the cost item A, but the actual cost item A in the accountability label should be judged, which violates the xth accountability rule. Generate error feedback: Cost item A violates the xth accountability rule and should be judged.

[0084] S204. If there is no difference in the comparison, the judgment result (t, i) is correct, and the process returns to S201 to enter the iteration of the next batch judgment data set (t, i+1).

[0085] It should be noted that the accountability rule set (t, i+1) is consistent with the accountability rule set (t, i), and the accountability rules do not need to be updated.

[0086] S205. Construct a rule combing instruction with the judgment data set (t,i), judgment rule set (t,i), judgment result (t,i), and error feedback (t,i), recorded as rule combing prompt (t,i), and input it into the rule combing model (t).

[0087] The rule combing instruction requires the large model to analyze the inaccuracies and imperfections of the judgment rules based on the error feedback information and correct the judgment rules.

[0088] S206. Randomly sample the output of the rule combing model M times, sort the outputs M times by the reward function, and update the judgment rule set (t, i) with the output with the highest reward value as the judgment rule set (t, i+1).

[0089] S207. Determine whether the iteration of all batches of the judgment data set in the tth round is completed. If not, return to step S201 and iterate in combination with the next batch of judgment data sets (t, i+1) and the judgment rule set (t, i+1) updated in step S206; if so, execute step S208.

[0090] The above random sampling of the output of the rule combing large model is a random sampling of the autoregressive generation of the large model. When the rule combing large model generates the next token (the basic unit of text, such as a word, a punctuation mark, a number or other meaningful character combination), it will first obtain the probability distribution of the next token. By setting hyperparameters such as a higher temperature coefficient, this probability distribution can be changed to enhance randomness, and the next token is obtained based on the probability distribution sampling, and then the entire text is generated by autoregression. Therefore, by setting certain hyperparameters and giving the same input to the rule combing large model M times, M different results of the large model output can be obtained.

[0091] After completing the iteration of all batches of the judgment data set in round t, the rule combing model is trained and updated from round t to round t+1. The specific steps include:

[0092] S208, calculating the difference between the output with the highest reward value and the output with the lowest reward value in the multiple outputs corresponding to each rule combing instruction in the t-th round of iterative training; if the difference is greater than a threshold, constructing a piece of preference data with the combing rule instruction, the output with the highest reward value, and the output with the lowest reward value. Summarizing all the preference data corresponding to the t-th round of iteration to form a preference data set.

[0093] For example, the format of preference data corresponding to the rule combing prompt(t,i) may be:

[0094] (Rule combing prompt(t,i), the sample with the highest reward value, the sample with the lowest reward value)

[0095] The multiple outputs corresponding to each rule combing instruction in the tth round of iterative training refer to the combing rule instructions (denoted as rule combing prompt(t,i)) input to the rule combing large model each time steps S205-S206 are executed in the tth round and the multiple outputs of the rule combing large model.

[0096] S209, use the preference data set to train the rule combing model (t) of the tth round, the training algorithm is the direct preference optimization algorithm (DPO), obtain the trained rule combing model (t+1) of the t+1th round, and re-execute the above step S201 to obtain the final judgment rule set.

[0097] In addition, based on the final updated judgment rules, the judgment big model can be used to judge the actual cases to be processed. Specifically, the case information of the cases to be judged and the final judgment rule set are used to construct judgment service instructions, and the instructions are input into the judgment big model to obtain the judgment results of the case information to be judged.

[0098] Based on the above Figure 1 and Figure 2 In the example of , the reward value of the output will be calculated. When calculating the reward value of the Nth output (N is a positive integer less than or equal to M) among M outputs, the specific steps can be:

[0099] First, the accuracy of the judgment model on the judgment data validation set is used as the task reward R(task) for the Nth output. N .

[0100] Task reward R(task) N =Accuracy(judgment model, judgment data validation set, Nth output).

[0101] The accountability data verification set includes case information and accountability labels. The accountability data verification set can be configured simultaneously when the accountability data set is obtained, or a portion of the accountability data in the accountability data set is divided out as the accountability data verification set.

[0102] The above task reward is to expect that the output judgment rules can enable the judgment model to have a higher judgment accuracy on the judgment data verification set.

[0103] Secondly, the inverse of the ratio of the length of the Nth output to the longest output length among the M outputs is used as the generated length reward R (token) for the Nth output. N .

[0104] R(token) N =-length k / max{length1,length2,...,length M}.

[0105] The above generation length reward is expected to compress the length of the judgment rule to improve the information density. At the same time, shorter judgment rule descriptions have better generalization to a certain extent.

[0106] Then, the inverse of the language model loss of the N-th output of the judgment model is used as the language fluency reward R(loss) of the N-th output. N .

[0107] R(loss) N =1 / loss(judgment model, Nth output).

[0108] The above language fluency reward is to make the expected output content easier for the judgment model to understand.

[0109] Finally, according to the task reward R(task) N , Generate length reward R (token) N and language fluency reward R(loss) N , calculate the reward value of the Nth output through the reward function;

[0110] The reward function of the Nth output is:

[0111] Reward function (Nth output) = R(task) N +λ(token)R(token) N +λ(loss)R(loss) N ,

[0112] Among them, λ(token) and λ(loss) are both hyperparameters.

[0113] The above are some specific implementations of a method for mining insurance liability rules provided by the embodiment of the present application. Based on this, the present application also provides a corresponding device. The device provided by the embodiment of the present application will be introduced from the perspective of functional modularization.

[0114] See also Figure 3 A schematic diagram of a structure of a device for mining insurance liability rules is shown, and the device for mining insurance liability rules includes:

[0115] The judgment unit 301 is used for judging the cost items of the cases included in the judgment data set of the target batch based on the judgment rule set by the judgment model, obtaining the judgment result and comparing it with the judgment label in the judgment data set;

[0116] A generating unit 302, configured to generate error feedback if there is a difference in the comparison;

[0117] An updating unit 303 is used for the rule combing model to modify the judgment rule set based on the error feedback, by randomly sampling the output of the rule combing model, and calculating the reward value of each output in combination with the reward function, to determine the optimal judgment rule set and update it;

[0118] The processing unit 304 is used to use the next batch of judgment data sets as the new target batch of judgment data sets, and execute the judgment model based on the judgment rule set and subsequent steps until the iterative training of the target number of batches of judgment data sets is completed;

[0119] The optimization unit 305 is used to construct a preference data set for this round of iterative training based on the output and reward value of random sampling corresponding to each batch of the judgment data set in this round of iterative training, train and update the rule combing model based on the preference data set and the direct preference optimization algorithm, and execute the judgment model based on the judgment rule set and subsequent steps until the target number of cycles is completed to obtain the final judgment rule set.

[0120] Based on the above device, the cost items of the cases in the accountability data set are judged based on the accountability rule set of the accountability big model of the judgment unit, and the accountability rule set is corrected through the rule combing big model of the update unit, so as to realize multiple rounds and batches of iterative training of the accountability data set. At the same time, after each round, the rule combing big model will be trained and updated through the preference data set and the direct preference algorithm, so that a set of accountability rules that is consistent with the actual case judgment can be combed out, and the accountability rules are mined and combed into text through the big model. The formed accountability rules can be understood by the accountability big model and used for the judgment of actual cases, and the workload of manual combing can be greatly saved.

[0121] In a possible implementation, the device further includes:

[0122] The initialization unit is used to initialize and construct the liability judgment rule set according to the exemption clause of the insurance policy, and the liability judgment rule set includes multiple liability judgment rules and the number of each liability judgment rule.

[0123] A preprocessing unit is used to obtain a accountability data set, wherein the accountability data set includes case information and accountability labels, wherein the accountability labels are accountability information marked for each expense item of each case, wherein the accountability information includes the expense item, a judgment on whether to be accountable, and a reason for violating the number of the accountability rule in the accountability rule set; and the accountability data set is divided into multiple batches.

[0124] Exemplarily, the case information includes the basic information of the insured and the materials uploaded for claim settlement.

[0125] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant laws, regulations and standards of relevant countries and regions.

[0126] In a specific implementation method, the generation unit is specifically used to generate feedback that the target cost item does not violate Article X of the judgment rule and should not be judged when the judgment model judges that the target cost item violates the Xth judgment rule, and the judgment label of the target cost item is that the target cost item is not judged; and to generate feedback that the target cost item violates Article X of the judgment rule and should be judged when the judgment model does not judge the target cost item, and the judgment label of the target cost item is that the target cost violates Article X of the judgment rule, and the Article X judgment rule is one of the judgment rules included in the judgment rule set.

[0127] The generating unit is further configured to, if no difference exists in the comparison, execute the step of using the next batch of judgment data sets as the new target batch of judgment data sets and subsequent steps.

[0128] In a specific implementation, the update unit is specifically used to set the hyperparameters of the rule combing large model, and construct rule combing instructions in combination with the target batch of judgment data sets, the judgment rule set, the judgment results and the error feedback; the rule combing instructions are input into the rule combing large model M times to obtain M outputs of the rule combing large model; the reward value of each output in the M outputs is calculated through a reward function, and the output with the highest reward value is used as the optimal judgment rule set.

[0129] In a specific implementation, the updating unit is specifically used to use the accuracy of the judgment of the judgment data verification set by the judgment model based on the Nth output as the task reward R(task) for the Nth output. N , the judgment data verification set includes case information and judgment labels; the inverse of the ratio of the length of the N-th output to the longest output length among the M-th output is used as the generation length reward R (token) of the N-th output N ; The inverse of the language model loss of the large judgment model for the Nth output is used as the language fluency reward R(loss) for the Nth output N ; According to the task reward R(task) N , Generate length reward R (token) N and language fluency reward R(loss) N , calculate the reward value of the Nth output through the reward function; the reward function of the Nth output is:

[0130] Reward function (Nth output) = R(task) N +λ(token)R(token) N +λ(loss)R(loss) N .

[0131] Among them, λ(token) and λ(loss) are both hyperparameters.

[0132] In a specific implementation, the optimization unit is specifically used to calculate the difference between the output with the highest reward value and the output with the lowest reward value in the randomly sampled outputs corresponding to each rule combing instruction in this round of iterative training; if the difference is greater than a threshold, the output with the highest reward value, the output with the lowest reward value and the rule combing instruction are used as a preference data in the preference data set.

[0133] In a specific implementation, the judgment unit is specifically used to construct a judgment task instruction based on the case information to be judged and the final judgment rule set, and input it into the judgment big model to obtain the judgment result of the case information to be judged.

[0134] The embodiments of the present application also provide corresponding devices and computer storage media for implementing the solutions provided by the embodiments of the present application.

[0135] The device includes a memory and a processor, the memory is used to store instructions or codes, and the processor is used to execute the instructions or codes so that the device executes a method for mining insurance liability rules as described in any embodiment of the present application.

[0136] The computer storage medium stores codes, and when the codes are executed, the device executing the codes implements a method for mining insurance liability rules as described in any embodiment of the present application.

[0137] The "first" and "second" in the names such as "first" and "second" (if any) mentioned in the embodiments of the present application are only used as name identifiers and do not represent the first or second in order.

[0138] Through the description of the above implementation methods, it can be known that those skilled in the art can clearly understand that all or part of the steps in the above-mentioned embodiment method can be implemented by means of software plus a general hardware platform. Based on such an understanding, the technical solution of the present application can be embodied in the form of a software product, which can be stored in a storage medium, such as a read-only memory (ROM) / RAM, a magnetic 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 communication device such as a router) to execute the methods described in each embodiment of the present application or some parts of the embodiments.

[0139] Each embodiment in this specification is described in a progressive manner, and the same and similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. Some or all of the modules can 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 work.

[0140] The above description is merely an exemplary embodiment of the present application and is not intended to limit the protection scope of the present application.

Claims

1. A method for mining insurance liability rules, characterized in that: The method comprises: The large accountability model, based on the accountability rule set, conducts accountability on the expense items of the cases included in the accountability data set of the target batch, obtains the accountability results and compares them with the accountability labels in the accountability data set; If there is a discrepancy in the alignment, an error feedback is generated; The rule combing big model executes the rule combing instruction to modify the judgment rule set based on the error feedback, and randomly samples the output of the rule combing big model, calculates the reward value of each output in combination with the reward function, determines the optimal judgment rule set and updates it; The next batch of judgment data sets are used as the new target batch of judgment data sets, and the judgment model is executed based on the judgment rule set and subsequent steps until the iterative training of the target number of batches of judgment data sets is completed; A preference data set for this round of iterative training is constructed based on the randomly sampled outputs and reward values ​​corresponding to each rule combing instruction in this round of iterative training. Based on the preference data set and the direct preference optimization algorithm, the rule combing large model is trained and updated, and the accountability large model is executed based on the accountability rule set and subsequent steps until the target number of cycles are completed to obtain the final accountability rule set.

2. The method according to claim 1, characterized in that: The method further comprises: Initializing and constructing the liability judgment rule set according to the exemption clause of the insurance policy, wherein the liability judgment rule set includes a plurality of liability judgment rules and a number of each liability judgment rule; Obtaining a judgment data set, wherein the judgment data set includes case information and judgment labels, wherein the judgment labels are judgment information marked for each expense item of each case, and the judgment information includes the expense item, a judgment on whether to judge, and a judgment reason for violating the number of the judgment rule in the judgment rule set; The judgment data set is divided into a plurality of batches.

3. The method according to claim 1, characterized in that If there is a discrepancy between the alignments, an error response is generated, including: When the accountability model judges the target cost item as violating the accountability rule No. X, and the accountability label of the target cost item is that the target cost item is not to be judged, generating feedback that the target cost item does not violate the accountability rule No. X and should not be judged; When the accountability model does not judge the target cost item, but the accountability label of the target cost item is that the target cost violates the Xth judgment rule, feedback is generated that the target cost item violates the Xth judgment rule and should be judged, and the Xth judgment rule is one of the judgment rules included in the judgment rule set.

4. The method according to claim 1, characterized in that The method further comprises: If there is no difference in the comparison, the step of using the next batch of judgment data sets as the new target batch of judgment data sets and subsequent steps are performed.

5. The method according to claim 1, characterized in that The rule combing model modifies the judgment rule set based on the error feedback, randomly samples the output of the rule combing model, calculates the reward value of each output in combination with the reward function, and determines the optimal judgment rule set, including: Setting the hyperparameters of the rule combing model, and constructing rule combing instructions in combination with the target batch of judgment data sets, the judgment rule set, the judgment results and the error feedback; Input the rule combing instruction M times into the rule combing big model to obtain M outputs of the rule combing big model; The reward function is used to calculate the reward value of each output in the M outputs, and the output with the highest reward value is taken as the optimal judgment rule set.

6. The method according to claim 5, characterized in that The reward value of the Nth output among the M outputs is calculated through the reward function, including: Based on the Nth output, the accuracy of the judgment of the judgment data validation set is used as the task reward R(task) for the Nth output. N , the judgment data verification set includes case information and judgment labels; The inverse of the ratio of the length of the Nth output to the longest output length among the M outputs is used as the generated length reward R (token) for the Nth output. N ; The inverse of the language model loss of the N-th output of the judgment model is used as the language fluency reward R(loss) of the N-th output. N ; According to the task reward R(task) N , Generate length reward R (token) N and language fluency reward R(loss) N , calculate the reward value of the Nth output through the reward function; The reward function of the Nth output is: Reward function (Nth output) = R(task) N +λ(token)R(token) N +λ(loss)R(loss) N , where λ(token) and λ(loss) are both hyperparameters.

7. The method according to claim 5, characterized in that The preferred data set of this round of iterative training is constructed according to the randomly sampled output and reward value corresponding to each rule combing instruction in this round of iterative training, including: Calculate the difference between the output with the highest reward value and the output with the lowest reward value in the randomly sampled output corresponding to each rule combing instruction in this round of iterative training; If the difference is greater than a threshold, the output with the highest reward value, the output with the lowest reward value, and the rule combing instruction are taken as a piece of preference data in the preference data set.

8. The method according to claim 1, characterized in that The method further comprises: The case information to be judged and the final judgment rule set are used to construct a judgment task instruction, which is input into the judgment big model to obtain the judgment result of the case information to be judged.

9. A device for mining insurance liability rules, characterized in that: The device comprises: A judgment unit, which is used for judging the cost items of the cases included in the judgment data set of the target batch based on the judgment rule set by the judgment big model, obtaining the judgment result and comparing it with the judgment label in the judgment data set; a generating unit, configured to generate an error feedback if a difference is found in the comparison; An updating unit, used for the rule combing model to modify the judgment rule set based on the error feedback, by randomly sampling the output of the rule combing model, calculating the reward value of each output in combination with the reward function, and determining and updating the optimal judgment rule set; A processing unit, used to use the next batch of judgment data sets as the new target batch of judgment data sets, and execute the judgment model based on the judgment rule set and subsequent steps until the iterative training of the target number of batches of judgment data sets is completed; The optimization unit is used to construct a preference data set for this round of iterative training according to the output and reward value of random sampling corresponding to each batch of the judgment data set in this round of iterative training, train and update the rule combing model based on the preference data set and the direct preference optimization algorithm, and execute the judgment model based on the judgment rule set and subsequent steps until the target number of cycles is completed to obtain the final judgment rule set.

10. The device according to claim 9, characterized in that The generating unit is further configured to, if no difference exists in the comparison, execute the step of using the next batch of judgment data sets as the new target batch of judgment data sets and subsequent steps.

Images