Model training methods and claims element extraction methods

Abstract

This application discloses a model training method and a claims element extraction method. The method includes: labeling entities in each insurance clause within an insurance clause dataset to obtain a first dataset, wherein the entities include: claims elements, element values ​​corresponding to the claims elements, and semantic fragments containing the claims elements; for each insurance clause in the first dataset, using a target template, and generating training instructions based on the insurance clause and the entities corresponding to the insurance clause, wherein the training instructions include a command part, an input part, and an output part, the input part corresponding to the insurance clause, and the output part corresponding to the labeled entities in the insurance clause; training a pre-set claims element extraction model using a second dataset to obtain a trained claims element extraction model, wherein the second dataset includes training instructions corresponding to each insurance clause.

Description

Technical Field

[0001] This application belongs to the field of computer technology, specifically relating to a model training method and a method for extracting claims elements. Background Technology

[0002] Traditional information extraction methods typically treat information as a whole, resulting in an inability to deeply analyze and extract specific details, leading to overly granular data. Due to this lack of refined processing, the extracted information is inaccurate, prone to redundancy or errors, and negatively impacts subsequent analysis and application. Therefore, traditional information extraction methods are insufficient to effectively improve the accuracy of extracting elements in the insurance field. Summary of the Invention

[0003] This application provides a model training method and a claims element extraction method, which can solve the problem of low element extraction accuracy in traditional information extraction methods.

[0004] In a first aspect, embodiments of this application provide a model training method, which includes: labeling entities in each insurance clause within an insurance clause dataset to obtain a first dataset, wherein the entities include: claim elements, element values ​​corresponding to the claim elements, and semantic fragments containing the claim elements; for each insurance clause in the first dataset, using a target template, and generating training instructions based on the insurance clause and the entities corresponding to the insurance clause, wherein the training instructions include a command part, an input part, and an output part, the input part corresponding to the insurance clause, and the output part corresponding to the entities labeled in the insurance clause; and training a preset claim element extraction model using a second dataset to obtain a trained claim element extraction model, wherein the second dataset includes the training instructions corresponding to each insurance clause. Secondly, embodiments of this application provide a method for extracting claims elements, the method comprising: obtaining an insurance clause to be processed; using a large-scale claims element extraction model to extract claims elements and corresponding element values ​​from the insurance clause to be processed, wherein the large-scale claims element extraction model is a model trained using the model training method described in the first aspect; and outputting the claims elements and corresponding element values.

[0005] Thirdly, embodiments of this application provide a model training apparatus, which includes: The annotation module is used to annotate entities in each insurance clause within the insurance clause dataset to obtain a first dataset, wherein the entities include: claim elements, element values ​​corresponding to the claim elements, and semantic fragments containing the claim elements; the generation module is used to generate training instructions for each insurance clause in the first dataset using a target template and based on the insurance clause and the entities corresponding to the insurance clause, wherein the training instructions include a command part, an input part, and an output part, the input part corresponding to the insurance clause, and the output part corresponding to the entities annotated in the insurance clause; the training module is used to train a preset claim element extraction model using a second dataset to obtain a trained claim element extraction model, wherein the second dataset includes the training instructions corresponding to each insurance clause.

[0006] Fourthly, embodiments of this application provide a claims element extraction device, which includes: an acquisition module for acquiring insurance terms to be processed; an extraction module for extracting claims elements and corresponding element values ​​from the insurance terms to be processed using a claims element extraction model, wherein the claims element extraction model is a model trained using the model training method described in the second aspect; and an output module for outputting the claims elements and corresponding element values.

[0007] Fifthly, embodiments of this application provide an electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the steps of the method described in the first aspect, or implement the steps of the method described in the second aspect.

[0008] In a sixth aspect, embodiments of this application provide a computer-readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect, or implement the steps of the method described in the second aspect.

[0009] In a seventh aspect, embodiments of this application provide a chip, the chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the method as described in the first aspect, or to implement the steps of the method as described in the second aspect.

[0010] Eighthly, embodiments of this application provide a computer program product comprising at least one computer program that, when loaded and executed by a processor, implements the method described in the first aspect or the steps of the method described in the second aspect.

[0011] In this embodiment, a first dataset is obtained by labeling entities in each insurance clause within the insurance clause dataset. The entities include: claim elements, element values ​​corresponding to the claim elements, and semantic fragments containing the claim elements. Then, for each insurance clause in the first dataset, a target template is used to generate training instructions based on the insurance clause and the corresponding entities. Each training instruction includes a command part, an input part, and an output part. The input part corresponds to the insurance clause, and the output part corresponds to the labeled entities in the insurance clause. Finally, a pre-defined claim element extraction model is trained using a second dataset to obtain a trained claim element extraction model. The second dataset includes the training instructions corresponding to each insurance clause, making the trained claim element extraction model more robust and improving the accuracy and effectiveness of claim element extraction, thereby meeting subsequent data processing needs. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of a model training method provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a method for extracting claims elements provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a model training device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a claims element extraction device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] The following description, in conjunction with the accompanying drawings, details a model training method and a claims element extraction method provided in this application through specific embodiments and application scenarios.

[0015] Figure 1This application illustrates a model training method according to an embodiment of the present application. This method can be executed by an electronic device, which may include a server and / or a terminal device. In other words, the method can be executed by software or hardware installed on the electronic device, and the method includes the following steps: S110: Label the entities in each insurance clause within the insurance clause data set to obtain the first dataset.

[0016] The entity includes: a claim element, an element value corresponding to the claim element, and a semantic fragment containing the claim element.

[0017] S110 may include the following steps: S111: Obtain the insurance terms data set, wherein the insurance terms data set includes multiple insurance terms.

[0018] S112: For each of the aforementioned insurance clauses, mark the semantic fragment containing the claims element.

[0019] S113: For each semantic segment containing the claim element, label the claim element and the element value corresponding to the claim element.

[0020] In other words, first, the complete semantic segments containing key claim information are marked, and then each claim element is marked from each segment. Claim elements include, but are not limited to, claim time limits, compensation amount, policyholder, and insured. For example, suppose an insurance clause includes the following: After deducting other expenses, the remaining balance will be paid as follows: For expenses not reimbursed by medical insurance, the deductible is 100 yuan, and the reimbursement rate is 90%; for expenses reimbursed by medical insurance, the deductible is 0 yuan, and the reimbursement rate is 90%. The semantic fragments containing claim elements in the insurance terms include: for expenses not reimbursed by medical insurance, the deductible is 100 yuan and the reimbursement rate is 90%; for expenses reimbursed by medical insurance, the deductible for the remaining reimbursable expenses is 0 yuan and the reimbursement rate is 90%. The claim elements in these semantic fragments include: medical insurance deductible, medical reimbursement rate, non-medical insurance deductible, and non-medical reimbursement rate. Specifically, the element value corresponding to the medical insurance deductible is 0 yuan, the element value corresponding to the medical reimbursement rate is 90%, the element value corresponding to the non-medical insurance deductible is 100 yuan, and the element value corresponding to the non-medical reimbursement rate is 90%.

[0021] Optionally, after obtaining the first dataset, the method further includes cleaning the first dataset, i.e., deleting insurance clauses for entities that are not labeled.

[0022] S120: For each insurance clause in the first dataset, a training instruction is generated using a target template based on the insurance clause and the entity corresponding to the insurance clause. The training instruction includes a command part, an input part, and an output part. The input part corresponds to the insurance clause, and the output part corresponds to the entity marked in the insurance clause.

[0023] It is understood that after obtaining the labeled first dataset, for each insurance clause in the first dataset, a training instruction is generated based on the insurance clause and the insurance clause itself. This training instruction can be a prompt instruction. Furthermore, the training instruction can be generated based on a pre-set target template. For example, the target template can be: { “instruct”: “Please extract key information about claims elements such as deductible, reimbursement ratio, and reimbursement limit from the following insurance terms. Note that the insurance terms are long and need to be extracted step by step to improve the accuracy of claims element extraction.” "input": "Insurance Terms" "output": "In the above insurance clause, the semantic fragment containing the claim elements is: [span_1, span_2, ..., span_n], where [span_1] contains the claim elements [y_1, y_2], the element value corresponding to [y_1] is [x_1], the element value corresponding to [y_2] is [x_2], [span_2] contains the claim elements [y_3, y_4], the element value corresponding to [y_3] is [x_3], the element value corresponding to [y_4] is [x_4], ... " }

[0024] In this context, "instruct" indicates the command part, "input" indicates the input part, and "output" indicates the output part. S130: Use the second dataset to train the pre-set large model for extracting claims elements, and obtain the trained large model for extracting claims elements.

[0025] The second dataset includes the training instructions corresponding to each of the insurance terms.

[0026] In other words, the second dataset consists of the training instructions corresponding to each of the aforementioned insurance clauses, and a large-scale claims element extraction model is then trained based on this second dataset. This large-scale claims element extraction model can be a Large Language Model (LLM). After training, a large language model possesses broad language understanding capabilities, capable of handling various complex language structures and contexts. It can not only identify explicit elements but also extract key elements from implicit semantics. For example, assuming an insurance clause states that the maximum cost of vaccines for animal bites or scratches is 52 yuan per dose, then the trained large-scale claims element extraction model can extract the claims element: the animal-caused single-dose limit, which corresponds to a value of 52 yuan. Furthermore, the large language model can be fine-tuned by adjusting parameters or incorporating additional data to achieve element extraction at different granularities.

[0027] In this step, training instructions using natural language allow for a clearer definition of extraction targets and refined requirements. This enables the claims element extraction model to focus more explicitly on specific information when performing tasks, significantly reducing errors caused by task ambiguity and thus improving extraction accuracy. Furthermore, when dealing with texts with complex grammatical structures or long dependencies, these training instructions help the claims element extraction model better understand the semantic structure of the text, thereby improving extraction accuracy. For example, in the first dataset, nested entities in the insurance clauses were annotated, allowing the claims element extraction model to more accurately resolve relationships between different entities and reduce the error rate. For instance, the entity "semantic fragment containing the claims element" belongs to the first level, while the entities "claims element" and "the element value corresponding to the claims element" belong to the second level.

[0028] In this embodiment, a first dataset is obtained by labeling entities in each insurance clause within the insurance clause dataset. The entities include: claim elements, element values ​​corresponding to the claim elements, and semantic fragments containing the claim elements. Then, for each insurance clause in the first dataset, a target template is used to generate training instructions based on the insurance clause and the corresponding entities. Each training instruction includes a command part, an input part, and an output part. The input part corresponds to the insurance clause, and the output part corresponds to the labeled entities in the insurance clause. Finally, a pre-defined claim element extraction model is trained using a second dataset to obtain a trained claim element extraction model. The second dataset includes the training instructions corresponding to each insurance clause, making the trained claim element extraction model more robust and improving the accuracy and effectiveness of claim element extraction, thereby meeting subsequent data processing needs.

[0029] In one implementation, before generating training instructions for each insurance clause in the first dataset using a target template and based on the insurance clause and the entity corresponding to the insurance clause, the method further includes: performing augmentation processing on the first dataset based on the labeled entities.

[0030] Understandably, augmenting the first dataset increases the diversity of training data, helping the large claims element extraction model learn linguistic features and semantic relationships more comprehensively, thus exhibiting higher extraction accuracy and robustness when facing complex real-world scenarios. Furthermore, augmenting the first dataset can generate more training samples, preventing insufficient training data from hindering the learning ability of the large claims element extraction model.

[0031] Furthermore, since the labeled entities include semantic fragments containing claim elements and feature values ​​corresponding to claim elements, enhancements can be applied separately to the semantic fragments containing claim elements and the feature values ​​corresponding to claim elements: (1) In one implementation, the enhancement of the first dataset based on the labeled entity includes: performing a first enhancement process on the first dataset based on the labeled semantic fragment containing the claim element to obtain at least one first restructured insurance clause; and adding the at least one first restructured insurance clause to the first dataset.

[0032] The first enhancement process includes at least one of the following: Exchange the semantic segments containing the claim elements in any two of the aforementioned insurance clauses; For any of the aforementioned insurance clauses, the semantic segment containing the claim element is deleted from the first position, and the semantic segment containing the claim element is inserted into the second position of the insurance clause, wherein the second position is any other position in the insurance clause besides the first position; For any of the aforementioned insurance clauses, insert a stop word or an invalid word into the insurance clause; For any of the aforementioned insurance clauses, a stop word or invalid word is inserted into at least one semantic segment containing the claim element corresponding to the insurance clause.

[0033] Understandably, the methods for performing data augmentation on semantic fragments to generate the first reconstructed insurance clause can include: randomly swapping a semantic fragment with other insurance clause texts to generate the first reconstructed insurance clause; deleting a semantic fragment from its original position in the original text and randomly inserting it into other positions to generate the first reconstructed insurance clause; randomly inserting stop words or invalid words into the insurance clause to generate the first reconstructed insurance clause; or randomly inserting stop words or invalid words into the semantic fragment containing the claim elements corresponding to the insurance clause to generate the first reconstructed insurance clause, wherein the stop words or invalid words are used to increase noise to improve the robustness of the large model for claim element extraction.

[0034] (2) In one implementation, the enhancement of the first dataset based on the labeled entities includes: performing a second enhancement process on the first dataset based on the element values ​​corresponding to the labeled claims elements to obtain at least one second restructured insurance clause; and adding the at least one second restructured insurance clause to the first dataset.

[0035] The second enhancement process includes at least one of the following: For any of the aforementioned insurance clauses, the element value corresponding to at least one of the claim elements in the insurance clauses is modified to a first value, wherein the first value is within a preset range, and the first value corresponding to different claim elements is different; For any of the aforementioned insurance clauses, the element value corresponding to at least one of the claim elements in the insurance clauses is modified to a second value, and the second value is marked with a result label, wherein the second value is outside a preset range, and the result label is used to indicate that the second value is incorrect, and the second value corresponding to different claim elements is different.

[0036] Understandably, enhancements to the claims elements may include: within a preset range, randomly modifying the element value corresponding to at least one of the claims elements, for example, replacing "payout ratio is 85%" with "payout ratio is 25%"; outside the preset range, randomly modifying the element value corresponding to at least one of the claims elements, for example, replacing "payout ratio is 85%" with "payout ratio is 135%", wherein "135%" is not within the preset range and can therefore be marked as incorrect to help the model understand and thus improve the model's learning ability.

[0037] In one implementation, training a pre-defined claims element extraction model using a second dataset to obtain a trained claims element extraction model includes: dividing the training instructions in the second dataset into N batches of training data to obtain a third dataset consisting of N batches of training data, where N is an integer greater than 0, and each batch includes a pre-defined number of training instructions; and using the third dataset to perform multiple rounds of iterative training on the pre-defined claims element extraction model, with each round of training using one batch of training data from the third dataset.

[0038] Understandably, the entire training process involves multiple iterations, i.e., traversing multiple batches, until all batches of data have been trained or the preset number of iterations has been reached. Each batch trained is called an epoch, i.e., one round of training. Through multiple epochs, the model's weights can be continuously optimized, gradually improving its extraction accuracy across the entire dataset.

[0039] In this implementation, since there are new sample data in each iteration, the large model for extracting claims elements can adapt to changes more quickly. Therefore, small-batch training can update the model weights more frequently, thereby accelerating the convergence process.

[0040] Furthermore, in another implementation, the step of using the third dataset to perform multiple rounds of iterative training on the preset claims element extraction model includes: in each round of training, for each training instruction, calculating the target cross-entropy loss between the standard probability corresponding to the labeled output part and the predicted probability corresponding to the output part predicted by the claims element extraction model; summing the target cross-entropy losses corresponding to each training instruction to obtain the total cross-entropy loss; and updating the weights of the claims element extraction model using the stochastic gradient descent algorithm based on the total cross-entropy loss.

[0041] Understandably, in each iteration, for each training instruction in the current batch, the claims element extraction model calculates the target cross-entropy loss based on the predicted output and the ground truth annotations. This target cross-entropy loss quantifies the difference between the predicted entity and the pre-labeled entity. Then, the target cross-entropy losses for all training instructions in the current batch are summed to obtain the total loss value for that batch, i.e., the total cross-entropy loss. This total cross-entropy loss reflects the overall performance of the claims element extraction model in the current batch. In this way, by using an appropriate batch size, the claims element extraction model can gradually adjust its weights and balance gradient fluctuations and stability, further improving the model's final performance.

[0042] Specifically, for each training instruction, calculating the target cross-entropy loss between the standard probability corresponding to the labeled entity and the predicted probability of the predicted entity output by the large model for extracting claims elements refers to serializing the training instruction, for example: <istct>{instruct} <eistct> <input> {input} <einput> <response>{output} <eresponse>,in, <istct> 、 <eistct> 、 <input> 、 <einput> 、 <response> 、 <eresponse>This is used to indicate the beginning or end of the corresponding content, and then the predicted probability of the sequence is maximized, for example: L(X) = ; Where L(X) represents the predicted probability of the sequence. This refers to each character in the sequence. It should be noted that, in the training instructions, since the `instruction` and `input` parts are known information, in order to reduce computational load, increase computational speed, and ensure fine-tuning effects, in this embodiment, only the cross-entropy of the output text is summed. For example, the target cross-entropy loss can be expressed by the following formula: Target-Cross-Entropy(p, = sum(plog( )) ,i>index of <einput>.

[0043] in, <einput>Indicates the end of the input section, i>index of <einput>This represents the character in the output part, where P is the standard probability. To predict probabilities.

[0044] like Figure 2 As shown in the embodiments of this application, a method for extracting claims elements is also provided, applying the above-mentioned... Figure 1 The model training method in this embodiment is used to train a model for recognition. This method includes the following steps: S210: Obtain pending insurance terms.

[0045] S220: Using a large model for extracting claims elements, extract the claims elements and their corresponding element values ​​from the insurance clauses to be processed.

[0046] The large-scale model for extracting claims elements adopts... Figure 1 The model is obtained by training the model using the model training method in the illustrated embodiment.

[0047] S230: Output the claim element and the element value corresponding to the claim element.

[0048] In this embodiment, the insurance terms to be processed are first obtained, then a large-scale claims element extraction model is used to extract claims elements and their corresponding element values ​​from the insurance terms to be processed, and finally the claims elements and their corresponding element values ​​are output. Since the large-scale claims element extraction model is implemented through the above... Figure 1 The model trained by the method provided in the illustrated embodiment has strong robustness and high extraction accuracy, resulting in more accurate final output claims elements.

[0049] Figure 3 This specification shows a schematic diagram of the structure of a model training device provided in an embodiment, as shown below. Figure 3 As shown, the model training device 300 may include: a labeling module 310, a generation module 320, and a training module 330.

[0050] In this embodiment, the annotation module 310 is used to annotate entities in each insurance clause within the insurance clause data set to obtain a first dataset, wherein the entities include: claim elements, element values ​​corresponding to the claim elements, and semantic fragments containing the claim elements; the generation module 320 is used to generate training instructions for each insurance clause in the first dataset using a target template and based on the insurance clause and the entities corresponding to the insurance clause, wherein the training instructions include a command part, an input part, and an output part, the input part corresponding to the insurance clause, and the output part corresponding to the entities annotated in the insurance clause; the training module 330 is used to train a preset claim element extraction model using a second dataset to obtain a trained claim element extraction model, wherein the second dataset includes the training instructions corresponding to each insurance clause.

[0051] In one implementation, the model training device 300 may further include an enhancement module for enhancing the first dataset based on the labeled entities.

[0052] In one implementation, the enhancement module is further specifically configured to perform a first enhancement process on the first dataset based on the labeled semantic fragments containing the claim element, to obtain at least one first restructured insurance clause; and to add the at least one first restructured insurance clause to the first dataset; wherein the first enhancement process includes at least one of the following: exchanging the semantic fragments containing the claim element in any two insurance clauses; for any insurance clause, deleting the semantic fragment containing the claim element from a first position and inserting the semantic fragment containing the claim element into a second position of the insurance clause, wherein the second position is any other position in the insurance clause besides the first position; for any insurance clause, inserting a stop word or invalid word into the insurance clause; for any insurance clause, inserting a stop word or invalid word into at least one semantic fragment containing the claim element corresponding to the insurance clause.

[0053] In one implementation, the enhancement module is further specifically used to perform a second enhancement process on the first dataset based on the element values ​​corresponding to the labeled claim elements, to obtain at least one second restructured insurance clause; and to add the at least one second restructured insurance clause to the first dataset; wherein the second enhancement process includes at least one of the following: for any one of the insurance clauses, modifying the element value corresponding to at least one claim element in the insurance clause to a first value, wherein the first value is within a preset range, and the first value corresponding to different claim elements is different; for any one of the insurance clauses, modifying the element value corresponding to at least one claim element in the insurance clause to a second value, and labeling the second value, wherein the second value is outside the preset range, and the labeling is used to indicate that the second value is incorrect, and the second value corresponding to different claim elements is different.

[0054] In one implementation, the training module 330 is further configured to divide the training instructions in the second dataset into N batches of training data to obtain a third dataset consisting of N batches of training data, wherein N is an integer greater than 0, and each batch includes a preset number of the training instructions; using the third dataset, the preset claims element extraction model is subjected to multiple rounds of iterative training, and each round of training uses one batch of training data from the third dataset.

[0055] In one implementation, the training module 330 is further configured to, during each training round, calculate, for each training instruction, the target cross-entropy loss between the standard probability corresponding to the labeled output portion and the predicted probability corresponding to the output portion predicted by the claims element extraction large model; sum the target cross-entropy losses corresponding to each training instruction to obtain the total cross-entropy loss; and based on the total cross-entropy loss, use the stochastic gradient descent algorithm to update the weights of the claims element extraction large model.

[0056] The model training device provided in this application embodiment can achieve... Figure 1 The various processes implemented in the method embodiments shown will not be described again here to avoid repetition.

[0057] Figure 4 This application illustrates a claims element extraction device 400, which may include an acquisition module 410, an extraction module 420, and an output module 430.

[0058] In this embodiment, the acquisition module 410 is used to acquire the insurance terms to be processed; the extraction module 420 is used to extract claim elements and corresponding element values ​​from the insurance terms to be processed using a claim element extraction model, wherein the claim element extraction model is the one described above. Figure 1 The model is trained using the model training method shown; the output module 430 is used to output the claim elements and the element values ​​corresponding to the claim elements.

[0059] The claims element extraction device provided in this application embodiment can achieve... Figure 2 The various processes implemented in the method embodiments shown will not be described again here to avoid repetition.

[0060] The model training device and the claims element extraction device in the embodiments of this application can be devices, or components, integrated circuits, or chips in electronic devices. The embodiments of this application are not specifically limited.

[0061] The model training device and the claims element extraction device in this application embodiment can be devices with an operating system. The operating system can be Android, iOS, or other possible operating systems; this application embodiment does not specifically limit the specific operating system.

[0062] Optional, such as Figure 5 As shown, this application embodiment also provides an electronic device 500, including a processor 510, a memory 520, and a program or instructions stored in the memory 520 and executable on the processor 510. When the program or instructions are executed by the processor 510, they implement the various processes of the above-described model training method embodiment or the various processes of the above-described claims element extraction method embodiment, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0063] This application also provides a computer-readable storage medium storing a program or instructions. When the program or instructions are executed by a processor, they implement the various processes of the above-described model training method embodiment or the various processes of the above-described claims element extraction method embodiment, and can achieve the same technical effect. To avoid repetition, they will not be described again here.

[0064] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes computer-readable storage media, such as computer read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0065] This application embodiment also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described model training method embodiment or the various processes of the above-described claims element extraction method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0066] It should be understood that the chip mentioned in the embodiments of this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0067] This application also provides a computer program product, which includes at least one computer program. When the computer program is loaded and executed by a processor, it implements the various processes of the above-described model training method embodiment or the various processes of the above-described claims element extraction method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0068] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0069] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0070] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.< / einput> < / einput> < / einput> < / eresponse> < / response> < / einput> < / eistct> < / istct> < / eresponse> < / response> < / einput> < / eistct> < / istct>

Claims

1. A model training method, characterized in that, include: The entities in each insurance clause within the insurance clause data set are labeled to obtain the first dataset, wherein the entities include: claims elements, the element values ​​corresponding to the claims elements, and semantic fragments containing the claims elements; For each insurance clause in the first dataset, a target template is used to generate a training instruction based on the insurance clause and the entity corresponding to the insurance clause. The training instruction includes a command part, an input part, and an output part. The input part corresponds to the insurance clause, and the output part corresponds to the entity marked in the insurance clause. The pre-set claim element extraction model is trained using the second dataset to obtain the trained claim element extraction model, wherein the second dataset includes the training instructions corresponding to each of the insurance clauses. The method further includes, before generating training instructions for each insurance clause in the first dataset using a target template and based on the insurance clause and the entity corresponding to the insurance clause, performing augmentation processing on the first dataset based on the labeled entities. The annotation-based entities enhance the first dataset, including: Based on the annotated semantic fragments containing the claim elements, the first dataset is subjected to a first augmentation process to obtain at least one first restructured insurance clause; Add the at least one first restructuring insurance clause to the first dataset; The first enhancement process includes at least one of the following: Exchange the semantic segments containing the claim elements in any two of the aforementioned insurance clauses; For any of the aforementioned insurance clauses, the semantic segment containing the claim element is deleted from the first position, and the semantic segment containing the claim element is inserted into the second position of the insurance clause, wherein the second position is any other position in the insurance clause besides the first position; For any of the aforementioned insurance clauses, insert a stop word or an invalid word into the insurance clause; For any of the aforementioned insurance clauses, a stop word or invalid word is inserted into at least one semantic segment containing the claim element corresponding to the insurance clause.

2. The method according to claim 1, characterized in that, The annotation-based entities enhance the first dataset, including: Based on the element values ​​corresponding to the labeled claim elements, the first dataset is subjected to a second augmentation process to obtain at least one second restructured insurance clause. Add the at least one second restructuring insurance clause to the first dataset; The second enhancement process includes at least one of the following: For any of the aforementioned insurance clauses, the element value corresponding to at least one of the claim elements in the insurance clauses is modified to a first value, wherein the first value is within a preset range, and the first value corresponding to different claim elements is different; For any of the aforementioned insurance clauses, the element value corresponding to at least one of the claim elements in the insurance clauses is modified to a second value, and the second value is marked with a result label, wherein the second value is outside a preset range, and the result label is used to indicate that the second value is incorrect, and the second value corresponding to different claim elements is different.

3. The method according to claim 1, characterized in that, The step of training a pre-defined large-scale model for extracting claims elements using a second dataset to obtain a trained large-scale model for extracting claims elements includes: The training instructions in the second dataset are divided into N batches of training data to obtain a third dataset consisting of N batches of training data, where N is an integer greater than 0, and each batch includes a preset number of training instructions. Using the third dataset, the preset large model for extracting claims elements is trained in multiple rounds of iterations, with each round of training using a batch of training data from the third dataset.

4. The method according to claim 3, characterized in that, The step of using the third dataset to perform multiple rounds of iterative training on the preset claims element extraction model includes: During each training round, for each training instruction, the target cross-entropy loss is calculated between the standard probability corresponding to the labeled output part and the predicted probability corresponding to the output part predicted by the large model for claim element extraction. The total cross-entropy loss is obtained by summing the target cross-entropy loss corresponding to each training instruction. Based on the total cross-entropy loss, the weights of the claims element extraction model are updated using the stochastic gradient descent algorithm.

5. A method for extracting claim elements, characterized in that, The method includes: Obtain pending insurance terms; Using a large-scale claims element extraction model, claims elements and corresponding element values ​​are extracted from the insurance clauses to be processed, wherein the large-scale claims element extraction model is a model trained using the model training method described in any one of claims 1-4. Output the claim elements and the corresponding element values.

6. A model training device, characterized in that, include: The annotation module is used to annotate the entities in each insurance clause within the insurance clause data set to obtain the first dataset, wherein the entities include: claims elements, the element values ​​corresponding to the claims elements, and semantic fragments containing the claims elements; The generation module is used to generate training instructions for each insurance clause in the first dataset, using a target template and based on the insurance clause and the entity corresponding to the insurance clause. The training instructions include a command part, an input part, and an output part. The input part corresponds to the insurance clause, and the output part corresponds to the entity marked in the insurance clause. The training module is used to train a pre-set large model for extracting claims elements using a second dataset to obtain a trained large model for extracting claims elements. The second dataset includes the training instructions corresponding to each of the insurance clauses. The model training device further includes an enhancement module for enhancing the first dataset based on the labeled entities. The enhancement module is further specifically configured to perform a first enhancement process on the first dataset based on the labeled semantic fragments containing the claim element, to obtain at least one first restructured insurance clause; and to add the at least one first restructured insurance clause to the first dataset; wherein the first enhancement process includes at least one of the following: exchanging the semantic fragments containing the claim element in any two insurance clauses; for any insurance clause, deleting the semantic fragment containing the claim element from a first position and inserting the semantic fragment containing the claim element into a second position of the insurance clause, wherein the second position is any other position in the insurance clause besides the first position; for any insurance clause, inserting a stop word or invalid word into the insurance clause; for any insurance clause, inserting a stop word or invalid word into at least one semantic fragment containing the claim element corresponding to the insurance clause.

7. A claims element extraction device, characterized in that, The device includes: The acquisition module is used to acquire insurance terms to be processed. The extraction module is used to extract claim elements and corresponding element values ​​from the insurance clauses to be processed using a large model for claim element extraction. The large model for claim element extraction is a model trained using the model training method described in any one of claims 1-4. The output module is used to output the claim elements and the corresponding element values.

8. An electronic device, characterized in that, It includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein when the program or instructions are executed by the processor, they implement the steps of the model training method as described in any one of claims 1-4, or the steps of the claims element extraction method as described in claim 5.

9. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the model training method as described in any one of claims 1-4, or the steps of the claims element extraction method as described in claim 5.

10. A computer program product, characterized in that, The computer program product includes program instructions that, when executed by a computer, cause the computer to implement the steps of the model training method as described in any one of claims 1-4, or the steps of the claims element extraction method as described in claim 5.

Images