Model training method and device, computer equipment and readable storage medium

Abstract

The invention relates to the technical field of machine learning, and provides a model training method and device, computer equipment and a readable storage medium, and the method comprises the steps: for a data set corresponding to a preset field, calculating a comprehensive score of each piece of sample data in the data set, so as to determine high-value data; preliminarily training a preset semantic understanding model according to the high-value data; performing weighted fusion on global semantic features output by the semantic understanding model and local features corresponding to the high-value data by using an attention mechanism to form fusion features; obtaining error example data corresponding to the preset field, and updating the high-value data according to the error example data and the fusion feature; service data of a preset field is collected in real time, and a training set corresponding to the semantic comprehension model is updated according to the service data and the updated high-value data; and completing the training of the semantic understanding model according to the updated training set. According to the method, the core problems of information asymmetry, poor model adaptability and the like in the fields of financial insurance and medical health old-age care are systematically solved.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a model training method, apparatus, computer device, and readable storage medium. Background Technology

[0002] In intelligent data processing within the financial insurance and healthcare / elderly care sectors, accurate semantic understanding, efficient feature extraction, and dynamic model adaptation capabilities are core technical challenges. Taking insurance as an example, insurance policy texts are characterized by dense technical terminology and complex semantic structures. Traditional models rely on random sampling for training, making it difficult to selectively filter high-value samples containing complex semantics. This leads to biases in the model's semantic understanding of insurance policy terms, impacting the accuracy of tasks such as customer needs identification and policy classification. In insurance claims review within the healthcare / elderly care sector, erroneous data (such as semantically ambiguous or logically broken claims texts) is prevalent. Traditional data repair methods lack deep integration with business context and domain rules, making it difficult to generate high-quality samples that match actual business scenarios. This results in low claims review efficiency and a high false positive rate. Furthermore, dynamic scenarios such as financial policy adjustments, insurance product iterations, and updates to elderly care rules require models to have real-time adaptability. However, existing data feedback mechanisms lack domain-adaptive data augmentation methods (such as synonym replacement for insurance terminology and professional optimization of syntactic structures), and incremental learning strategies do not incorporate the dynamic characteristics of business data, causing model updates to lag behind actual business needs.

[0003] The core shortcomings of existing technologies are: (1) Sample selection relies on experience or random strategies, and a quantitative evaluation system is not designed for the professional terminology characteristics of insurance texts; (2) The feature extraction model is singular and it is difficult to take into account both the global semantics and local fine-grained features of the text; (3) Error correction lacks business rule guidance, resulting in low quality of generated samples; (4) The data feedback mechanism is out of sync with the dynamic changes in the domain, resulting in poor model adaptability. These problems are particularly prominent in financial insurance and medical health and elderly care scenarios, and there is an urgent need for an intelligent data processing solution that integrates active learning, multimodal feature fusion, generative adversarial networks and dynamic feedback technologies to solve key problems such as inaccurate semantic understanding and weak model generalization ability in complex business scenarios. Summary of the Invention

[0004] This application provides a model training method, apparatus, computer equipment and readable storage medium, aiming to solve the core defects of the prior art, which are: (1) sample selection relies on experience or random strategies and does not design a quantitative evaluation system for the professional terminology characteristics of insurance texts; (2) the feature extraction model is singular and it is difficult to take into account both the global semantics and local fine-grained features of the text; (3) the error correction lacks business rule guidance and the generated sample quality is low; (4) the data feedback mechanism is out of touch with the dynamic changes in the domain and the model has poor adaptability.

[0005] Firstly, this application provides a model training method, including: For a dataset corresponding to a preset domain, a comprehensive score is calculated for each sample data in the dataset to determine high-value data based on the comprehensive score; the comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a difference score based on the predicted label and the manually labeled label. A pre-set semantic understanding model is initially trained based on the high-value data; the global semantic features corresponding to the high-value data output by the semantic understanding model are weighted and fused with the local features corresponding to the high-value data using an attention mechanism to form fused features; Obtain the faulty instance data corresponding to the preset domain, and update the high-value data based on the faulty instance data and the fusion features; Real-time collection of business data in the preset domain is used to update the training set corresponding to the semantic understanding model based on the business data and the updated high-value data; the semantic understanding model is trained based on the updated training set to improve the semantic understanding model's semantic understanding capability in the preset domain.

[0006] In some embodiments, calculating the comprehensive score of each sample data in the dataset to determine high-value data based on the comprehensive score includes: for each sample data, calculating the distribution of professional terms in the text using information entropy to obtain the uncertainty score; calculating the text similarity between the sample data and a preset scene library using cosine similarity to obtain the representativeness score; obtaining the difference score by comparing the corresponding initial prediction label with the professional annotation result; weighting and summing the uncertainty score, representativeness score, and difference score to obtain the comprehensive score; and sorting the sample data according to the comprehensive score to filter out a preset proportion of high-value data.

[0007] In some embodiments, the semantic understanding model is a BERT model; the preliminary training of the preset semantic understanding model based on the high-value data includes: inputting the high-value data into the BERT model for training, optimizing the parameters of the BERT model through the semantic features of insurance clause text in the high-value data, and improving the accuracy of the BERT model in semantic understanding of insurance domain text.

[0008] In some embodiments, the step of weightedly fusing the global semantic features corresponding to the high-value data output by the semantic understanding model with the local features corresponding to the high-value data using an attention mechanism to form a fused feature includes: performing local sliding processing of multi-scale convolutional kernels on the high-value data through a preset convolutional neural network model to extract local features of different granularities; inputting the global semantic feature vector output by the semantic understanding model and the local features into the attention mechanism, dynamically allocating weights according to the semantic correlation between the global semantic features and the local features, and generating a fused feature vector for insurance clause classification or customer demand identification tasks.

[0009] In some embodiments, updating the high-value data based on the erroneous sample data and the fused features includes: the erroneous sample data being misjudged or missed samples generated in a business process corresponding to a preset domain; optimizing the erroneous sample data using a generative adversarial network, generating new samples by combining the textual features, contextual information, and associated insurance clauses of the erroneous sample data with a generator, and judging the semantic features, syntactic structure, and contextual consistency of the new samples by a discriminator; and through adversarial training between the generator and the discriminator, making the generator output an optimized result that approximates real high-quality samples, and using the optimized result as the updated high-value data.

[0010] In some embodiments, updating the training set corresponding to the semantic understanding model based on the business data and the updated high-value data includes: collecting the business data in real time through a stream processing framework, merging the business data with the updated high-value data according to a dynamically adjusted update frequency; performing data augmentation processing on the merged text data, including synonym replacement based on word vector models, sentence rearrangement based on dependency parsing, and randomly inserting insurance-related noise words to form an updated training set.

[0011] In some embodiments, training the semantic understanding model based on the updated training set includes: updating the semantic understanding model online using an incremental learning strategy, gradually inputting the updated training set into the semantic understanding model, and adjusting the model parameters to adapt the semantic understanding model to dynamic scenarios in a preset domain, thereby improving the semantic understanding model's ability to understand complex insurance scenarios.

[0012] Secondly, this application provides a model training apparatus, comprising: The data calculation unit is used to calculate the comprehensive score of each sample data in the dataset corresponding to the preset domain, so as to determine high-value data based on the comprehensive score; the comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a difference score based on the predicted label and the manual annotation. The initial training unit is used to initially train a preset semantic understanding model based on the high-value data; the global semantic features corresponding to the high-value data output by the semantic understanding model are weighted and fused with the local features corresponding to the high-value data using an attention mechanism to form fused features; The data update unit is used to acquire the error data corresponding to the preset domain, and update the high-value data according to the error data and the fusion features; The training completion unit is used to collect business data in the preset domain in real time, update the training set corresponding to the semantic understanding model based on the business data and the updated high-value data, and complete the training of the semantic understanding model based on the updated training set to improve the semantic understanding model's semantic understanding ability in the preset domain.

[0013] Thirdly, this application also provides a computer device, comprising: Memory and processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, in executing the computer program, implement the steps of the model training method as described in the first aspect above.

[0014] Fourthly, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to implement the steps of the model training method described in the first aspect above.

[0015] This invention provides a model training method that quantifies sample value through a three-dimensional evaluation system of information entropy (uncertainty), cosine similarity (representativeness), and prediction difference (label consistency). It prioritizes training the BERT model with data that most significantly improves the model, overcoming the blindness of traditional random training. It integrates BERT's global semantic encoding capabilities with CNN's multi-scale local feature extraction capabilities, dynamically weighting and generating fused features through an attention mechanism to provide multi-level semantic expression for tasks such as clause classification and demand identification. For business errors, it generates high-quality samples by combining text features, context, and domain clauses to correct data defects such as semantic ambiguity and improve review accuracy. Through real-time stream processing, domain-adaptive data augmentation (synonym replacement, syntactic rearrangement, etc.), and incremental learning strategies, the model responds in real-time to dynamic scenarios such as policy changes and product iterations, forming a data-driven closed-loop optimization.

[0016] The three-dimensional sample screening strategy specifically addresses the complexity of insurance text terminology, improving the training efficiency of the BERT model and enhancing the accuracy of semantic understanding of terms compared to traditional methods. The multimodal fusion architecture overcomes the limitations of single-model feature extraction, improving the F1 score in customer needs identification tasks and providing higher-quality feature inputs for downstream tasks such as intelligent customer service and risk assessment. The GAN-based error correction technology improves the semantic completeness of claims text, reduces the false negative rate, and significantly reduces the cost of manual review. The real-time feedback mechanism combined with domain augmentation technology reduces the model's response time to new product terms and policy adjustments from 72 hours in traditional solutions to minutes, continuously maintaining the synchronization between model performance and business scenarios.

[0017] This solution systematically addresses core issues such as information asymmetry and poor model adaptability in the fields of finance, insurance, healthcare, and elderly care through technological integration and process innovation. It constructs an intelligent data processing framework with independent intellectual property rights, demonstrating significant technological advancement and industrial application value.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart illustrating the steps of a model training method provided in an embodiment of this application; Figure 2 This is a schematic diagram of the architecture of a model training method provided in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a model training device provided in one embodiment of this application; Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application.

[0021] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Detailed Implementation

[0022] 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.

[0023] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content and operations / steps, nor does it necessarily have to be performed in the order described. For example, some operations / steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0024] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0025] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the scope of the application. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0026] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0028] In intelligent data processing within the financial insurance and healthcare / elderly care sectors, accurate semantic understanding, efficient feature extraction, and dynamic model adaptation capabilities are core technical challenges. Taking insurance as an example, insurance policy texts are characterized by dense technical terminology and complex semantic structures. Traditional models rely on random sampling for training, making it difficult to selectively filter high-value samples containing complex semantics. This leads to biases in the model's semantic understanding of insurance policy terms, impacting the accuracy of tasks such as customer needs identification and policy classification. In insurance claims review within the healthcare / elderly care sector, erroneous data (such as semantically ambiguous or logically broken claims texts) is prevalent. Traditional data repair methods lack deep integration with business context and domain rules, making it difficult to generate high-quality samples that match actual business scenarios. This results in low claims review efficiency and a high false positive rate. Furthermore, dynamic scenarios such as financial policy adjustments, insurance product iterations, and updates to elderly care rules require models to have real-time adaptability. However, existing data feedback mechanisms lack domain-adaptive data augmentation methods (such as synonym replacement for insurance terminology and professional optimization of syntactic structures), and incremental learning strategies do not incorporate the dynamic characteristics of business data, causing model updates to lag behind actual business needs.

[0029] The core shortcomings of existing technologies are: (1) Sample selection relies on experience or random strategies, and a quantitative evaluation system is not designed for the professional terminology characteristics of insurance texts; (2) The feature extraction model is singular and it is difficult to take into account both the global semantics and local fine-grained features of the text; (3) Error correction lacks business rule guidance, resulting in low quality of generated samples; (4) The data feedback mechanism is out of sync with the dynamic changes in the domain, resulting in poor model adaptability. These problems are particularly prominent in financial insurance and medical health and elderly care scenarios, and there is an urgent need for an intelligent data processing solution that integrates active learning, multimodal feature fusion, generative adversarial networks and dynamic feedback technologies to solve key problems such as inaccurate semantic understanding and weak model generalization ability in complex business scenarios.

[0030] To resolve the above issues, please refer to [link / reference]. Figure 1 , Figure 1 This is a schematic flowchart of a model training method provided in one embodiment of this application. The model training method can be implemented by a computer device, which can be deployed on a single server or a server cluster. It can also be deployed on a handheld terminal, laptop, wearable device, or robot, etc.

[0031] It should be noted that the acquisition of any information mentioned in the provided methods is in compliance with relevant regulations and is carried out with the user's consent, and will not infringe on the user's privacy or violate relevant laws and regulations.

[0032] To solve the above problem, please refer to Figure 1 Specifically, such as Figure 1As shown, the provided model training method includes steps S101 to S104. Details are as follows: Step S101. For the dataset corresponding to the preset domain, calculate the comprehensive score of each sample data in the dataset, so as to determine high-value data based on the comprehensive score; the comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a difference score based on the predicted label and the manual annotation.

[0033] Specifically, a quantitative evaluation system is constructed using three-dimensional indicators: information entropy (uncertainty), cosine similarity (representativeness), and prediction difference (label consistency). The comprehensive score of the samples is calculated, and the samples with the highest value for improving the model are selected as high-value data.

[0034] In the financial insurance scenario, information entropy calculation targets insurance policy texts, statistically analyzing the distribution density of professional terms such as "exclusion clauses," "deductible," and "waiting period," quantifying term complexity using the information entropy formula (higher entropy values ​​indicate more dispersed term distribution and higher semantic uncertainty). Cosine similarity calculation uses pre-set business scenario text sets such as "critical illness insurance claims scenario library" and "auto insurance accident description library," vectorizing samples with scenario library texts (e.g., TF-IDF or BERT word vectors), calculating cosine similarity, and identifying samples with high matching degree to high-frequency claims scenarios. Predictive difference calculation uses an initial model (e.g., basic BERT) to generate predicted labels for samples, compares them with the results of manual annotation by legal personnel, and quantifies the difference using the Jaccard coefficient or edit distance. The comprehensive score is calculated by weighting and summing three indicators (e.g., uncertainty 40% + representativeness 30% + difference 30%), selecting the top 20% of samples as high-value data (e.g., interpretation of complex clauses, disputed claims texts).

[0035] In the medical, health, and elderly care scenarios, information entropy calculation quantifies the degree of semantic ambiguity by statistically analyzing the distribution of medical-insurance cross-terms such as "chronic disease medication," "rehabilitation treatment projects," and "medical insurance reimbursement catalog" in medical claim texts. Cosine similarity calculation identifies typical elderly care or health insurance claim samples by constructing scenario databases in fields such as "hypertension complication claim scenario database" and "elderly care service demand database" through semantic vector matching. Prediction difference calculation quantifies the deviation in understanding review rules by comparing the model's prediction results for tags such as "compliance of diagnosis and treatment projects" and "reasonableness of medication dosage" with the annotations of medical insurance review experts. The comprehensive score selects high-value samples involving rare pathology claims and cross-departmental diagnosis and treatment descriptions by emphasizing the weight of medical terminology complexity and the matching degree of review rules (such as uncertainty 35% + representativeness 35% + difference 30%).

[0036] It avoids the problem of "redundancy in simple samples and lack of complex samples" caused by traditional random sampling, improves the training coverage of complex clause samples in financial insurance scenarios, and improves the efficiency of cross-domain terminology sample screening in medical and elderly care scenarios; it automatically identifies weak samples in the model by predicting difference indicators, reduces most of the invalid annotation work in financial scenarios, and improves the targeting of expert annotation in medical scenarios.

[0037] Step S102. Initially train a preset semantic understanding model based on the high-value data; use an attention mechanism to weightedly fuse the global semantic features corresponding to the high-value data output by the semantic understanding model with the local features corresponding to the high-value data to form fused features.

[0038] Specifically, the BERT semantic understanding model is trained using high-value data to extract global semantic features, and local fine-grained features are extracted using CNN multi-scale convolutional kernels. These features are then dynamically weighted and fused based on an attention mechanism to form fused features for downstream tasks.

[0039] In financial insurance scenarios, BERT model training involves inputting high-value insurance policy texts (such as participating insurance benefit clauses and liability insurance exclusion clauses), optimizing parameters through Masked Language Modeling (MLM) and Next Sentence Prediction (NSP) tasks, and outputting a global feature vector (e.g., 768-dimensional) containing the overall semantics of the policy. CNN local feature extraction uses 3 / 5 / 7 scale convolution kernels to slide across the text sequence, extracting local syntactic features such as "subject-verb-object structure" and "conditional adverbial clauses" (e.g., the conditional logic feature of "if the insured is diagnosed with a critical illness during the waiting period, no compensation will be paid"). Attention fusion calculates the semantic correlation between global and local features through a self-attention mechanism (e.g., the association weight between the "waiting period" term and the "diagnosed with a critical illness" scenario), and generates weighted fused features for insurance policy classification (e.g., liability / exclusion clauses) or customer consultation intent recognition (e.g., claims application / policy consultation).

[0040] In the medical and health care elderly care scenario, the BERT model training captures cross-domain semantic associations such as "medical drugs outside the medical insurance catalog" and "designated medical institutions" by inputting high-value medical claim texts (such as chronic disease outpatient reimbursement applications and rehabilitation aid purchase reimbursement forms); CNN local feature extraction focuses on the triple structure of "visit time-treatment items-cost amount" and extracts local features of "drug name + dosage + indication" through small-scale convolution kernel (3) and long-distance dependency features of "hospitalization days + nursing level + reimbursement ratio" through large-scale kernel (7); attention fusion strengthens the semantic alignment of "service agreement terms" and "cost details" through attention mechanism for the complex logic of "nursing service type-service duration-payment method" in the elderly care scenario, and generates fusion features for the compliance judgment of claim review.

[0041] In financial scenarios, the accuracy of clause classification has improved from 82% of the traditional single BERT, and the misjudgment rate of customer demand identification has decreased; in medical and elderly care scenarios, the accuracy of understanding texts such as "changes in reimbursement rules caused by updates to medical insurance policies" has improved, and the semantic mapping problem of "generic name of drug - brand name" and "treatment item code - name" has been effectively handled.

[0042] Step S103. Obtain the error data corresponding to the preset domain, and update the high-value data based on the error data and the fusion features.

[0043] Specifically, for misjudgments / missed judgments in business processes, high-quality repair samples are generated by combining generative adversarial networks (GANs) with domain rules to update high-value datasets.

[0044] In the financial insurance scenario, the error input is collected from car insurance claims for "vague accident liability descriptions" (such as "multi-vehicle collisions without a clear primary responsible party") and health insurance claims for "incomplete disclosure of pre-existing conditions," extracting text features, related clauses, and contextual information (such as accident scene photo descriptions). The generator mechanism is built based on the Transformer architecture, inputting the error feature vector and combining it with insurance clause rules (such as "liability division must include the accident report number") to generate logically complete claim description text. The discriminator mechanism is trained to distinguish between generated samples and real high-quality samples, focusing on verifying "semantic consistency" (such as the matching of liability division with clause definitions) and "syntactic compliance" (such as the completeness of time / location / responsible party elements). Through adversarial training, the pass rate of generated samples reaches over 90%.

[0045] In the healthcare and elderly care scenarios, error input is achieved by collecting errors such as "mismatch between treatment items and disease diagnosis" in outpatient reimbursement (e.g., "hypertensive patients reimbursing ophthalmology examination fees") and "missing key time points in service records" in elderly care claims, and linking them with knowledge in areas such as medical insurance reimbursement catalogs and elderly care service standards; the generator mechanism introduces a medical terminology knowledge base (e.g., the ICD-10 disease coding library), and when fixing the error of "missed reporting of diabetic complications," the generator automatically supplements necessary information such as "frequency of blood glucose testing" and "diagnostic basis for complications"; the discriminator mechanism adds "compliance of treatment logic" judgment for medical scenarios (e.g., "medical correlation between medication dosage and weight and age") to ensure that the generated samples comply with clinical guidelines and medical insurance reimbursement rules.

[0046] Step S104. Collect business data from the preset domain in real time, and update the training set corresponding to the semantic understanding model based on the business data and the updated high-value data; complete the training of the semantic understanding model based on the updated training set, and improve the semantic understanding model's semantic understanding ability in the preset domain.

[0047] Specifically, business data is collected through a real-time stream processing framework, and the training set is updated by combining domain-adaptive data augmentation (synonym replacement, syntactic rearrangement, and noise injection). An incremental learning strategy is used to achieve dynamic optimization of the model.

[0048] In the financial insurance scenario, real-time stream processing is used: data streams generated by the release of new insurance product terms and updates to regulatory policies are captured in real time through the Flink / Kafka framework and fed back to the training set at a frequency of minutes; data augmentation is performed by performing domain-specific augmentations on insurance terminology (such as replacing "premium waiver" with "premium waiver" and rearranging "exclusion of liability" to "exclusion of liability"), and randomly inserting role-confusion noise such as "policyholder" and "insured" to simulate the diversity of expressions in customer consultations; incremental learning uses the Elastic Weight Consolidation (EWC) algorithm to retain the ability to understand existing terms when updating the model, while quickly adapting to new product knowledge such as "new coverage of long-term care insurance", avoiding "catastrophic forgetting".

[0049] In the healthcare and elderly care scenarios, real-time stream processing collects business data generated by dynamic adjustments to the medical insurance catalog (such as the addition of reimbursement for anti-cancer drugs) and updates to elderly care service standards (such as the revision of detailed rules for disability level assessment) in real time, and updates the training set at an hourly frequency; data augmentation, based on medical dependency parsing, rearranges "coronary artery bypass grafting" to "coronary artery bypass surgery" and inserts noisy words in fields such as "Class A and B drugs" and "day surgery" to enhance the model's ability to identify variations of medical terms; incremental learning, combined with the security requirements of medical scenarios, uses knowledge distillation technology to encode expert review rules into soft labels during incremental training, ensuring the model's accuracy in responding to low-frequency scenarios such as "rare disease reimbursement policies".

[0050] In the financial sector, the model's understanding delay of the terms of newly released insurance products has been reduced from 48 hours in the traditional approach to 15 minutes, while the accuracy of policy interpretation remains above 95%. In the medical and elderly care sector, data augmentation has improved the model's accuracy in handling issues such as "different names for the same drug" and "cross-system terminology mapping" by 25%, and the incremental learning strategy has ensured that the model's stability has been improved by 30% under frequent adjustments to medical insurance policies.

[0051] By adjusting the weights of the three-dimensional evaluation system, embedding domain rules into the GAN generator, and processing professional terminology for data augmentation, the system achieves differentiated coverage of financial insurance, medical and health care, and elderly care scenarios. It forms a self-optimizing cycle of "high-value data screening → multimodal feature enhancement → error correction → dynamic feedback," solving the chain problem of "poor data quality - low training efficiency - slow adaptation speed" in traditional models. In financial scenarios, it improves intelligent underwriting efficiency by more than 30%, and reduces the claims dispute rate by 20% in medical and elderly care scenarios, significantly optimizing the efficiency of human-machine collaboration and decision-making accuracy in complex business scenarios.

[0052] In some embodiments, calculating the comprehensive score of each sample data in the dataset to determine high-value data based on the comprehensive score includes: for each sample data, calculating the distribution of professional terms in the text using information entropy to obtain the uncertainty score; calculating the text similarity between the sample data and a preset scene library using cosine similarity to obtain the representativeness score; obtaining the difference score by comparing the corresponding initial prediction label with the professional annotation result; weighting and summing the uncertainty score, representativeness score, and difference score to obtain the comprehensive score; and sorting the sample data according to the comprehensive score to filter out a preset proportion of high-value data.

[0053] The comprehensive score of the sample is calculated using a three-dimensional quantitative index system: Uncertainty score: based on information entropy to quantify the complexity of the distribution of professional terms in the text; Representativeness score: calculated by cosine similarity to determine the matching degree between the sample and the preset scene library; Difference score: compared with the difference between the model's initial predicted label and the manually labeled result; After weighted summation, high-value data (such as the top 20%) are selected according to the preset proportion based on the score.

[0054] In the financial insurance scenario, uncertainty scoring is achieved by extracting professional terms such as "exclusion clauses," "waiting period," and "cash value" from insurance policy texts and calculating the information entropy of the frequency of these terms (higher entropy indicates more dispersed term distribution and more ambiguous semantics). Representativeness scoring is achieved by vectorizing samples (e.g., using BERT word vectors) and calculating cosine similarity with scenario texts from pre-defined scenario libraries such as "auto insurance accident liability description library" and "critical illness insurance health disclosure library." Higher similarity indicates stronger representativeness (e.g., matching claims text in the "multi-vehicle accident liability division" scenario). Dissimilarity scoring is achieved by using initial model-predicted sample labels (e.g., "liability attribution" and "exclusion") and calculating the Jaccard coefficient with legal personnel annotations; greater dissimilarity results in a higher score (e.g., samples where the model misclassifies "drunk driving exemption clause" as "claimable clause"). A weighted strategy, using 40% uncertainty + 30% representativeness + 30% dissimilarity, selects the top 20% of complex clauses or controversial claims texts as high-value data.

[0055] In the healthcare and elderly care scenarios, uncertainty scores are calculated by statistically analyzing the information entropy of cross-domain terms such as "ICD-10 code," "medications not covered by medical insurance," and "rehabilitation and nursing projects" in medical claim texts to quantify semantic ambiguity (e.g., texts involving both "diabetic complications" and "traditional Chinese medicine physiotherapy" have higher entropy values). Representativeness scores are generated by constructing "hypertension outpatient reimbursement scenario libraries" and "disability level assessment scenario libraries," and by matching the similarity between samples and these scenario libraries using semantic vectors to identify typical elderly care or health insurance claim scenarios (e.g., matching application texts for "long-term care service duration determination" scenarios). Discrepancy scores are calculated by comparing the model's predictions for the "medical treatment project compliance" label with the annotations from medical insurance experts, calculating the edit distance (e.g., samples where the model missed classifying "treatment at non-designated institutions" as non-compliant have higher discrepancy scores). A weighted strategy, using 35% uncertainty + 35% representativeness + 30% discrepancy, selects high-value samples involving rare disease reimbursement and cross-departmental medical treatment descriptions.

[0056] By accurately screening complex clause samples, the coverage of complex semantic samples during BERT model training is increased by 60%, and the accuracy of clause understanding is improved; the claims text with ambiguous cross-domain terminology is automatically located, reducing most invalid sample annotations and improving expert annotation efficiency.

[0057] In some embodiments, the semantic understanding model is a BERT model; the preliminary training of the preset semantic understanding model based on the high-value data includes: inputting the high-value data into the BERT model for training, optimizing the parameters of the BERT model through the semantic features of insurance clause text in the high-value data, and improving the accuracy of the BERT model in semantic understanding of insurance domain text.

[0058] The BERT model is used as the semantic understanding model. High-value data is input for domain-adaptive training. The model parameters are optimized by using the semantic features of insurance / medical texts to improve the ability to understand professional terms and complex sentence structures.

[0059] In financial insurance scenarios, input data includes high-value insurance policy texts (such as "annuity insurance benefit calculation clauses" and "liability insurance exclusion clauses") and disputed claim application texts (such as "texts on disputes over claims for undisclosed pre-existing conditions"). Training strategies include using a masked language model (MLM) task to randomly mask technical terms in the policy terms (such as "deductible" and "cooling-off period") and train the model to predict missing words; strengthening the understanding of the contextual logic of the policy terms through the next sentence prediction (NSP) task (such as the connection between "underwriting conditions" and "insurance liability"); and optimizing the Transformer encoder parameters of BERT, especially the attention head weights related to the semantics of insurance terminology.

[0060] In the healthcare and elderly care scenarios, input data includes high-value medical claim texts (such as "targeted drug reimbursement application" and "rehabilitation aids expense list") and elderly care service agreement terms (such as "disability level assessment details" and "nursing service pricing standards"). The training strategy increases the proportion of domain-specific masks in the MLM task for medical-insurance cross-terms (such as "chronic disease outpatient care" and "day surgery"); introduces medical insurance reimbursement rules as implicit supervision signals, and optimizes parameters by comparing the semantic differences between "compliant" and "non-compliant" claim texts; and focuses on optimizing the model's encoding ability for cross-domain semantic associations such as "treatment items-reimbursement ratio" and "drug dosage-indication".

[0061] The BERT model improved its accuracy in parsing ambiguous sentences in insurance clauses from 75% to 92%, and reduced the misjudgment rate of customer inquiry intent identification by 30%. In the medical field, the model improves the accuracy of judging the "rationality of drugs used outside the medical insurance catalog," reducing claims disputes caused by misunderstandings of terminology.

[0062] In some embodiments, the step of weightedly fusing the global semantic features corresponding to the high-value data output by the semantic understanding model with the local features corresponding to the high-value data using an attention mechanism to form a fused feature includes: performing local sliding processing of multi-scale convolutional kernels on the high-value data through a preset convolutional neural network model to extract local features of different granularities; inputting the global semantic feature vector output by the semantic understanding model and the local features into the attention mechanism, dynamically allocating weights according to the semantic correlation between the global semantic features and the local features, and generating a fused feature vector for insurance clause classification or customer demand identification tasks.

[0063] Local text features (character / word-level fine-grained semantics) are extracted using CNN multi-scale convolutional kernels and dynamically weighted and fused with the global semantic features output by BERT through an attention mechanism to generate fused features that take into account both global semantics and local details for use in downstream tasks.

[0064] In the financial insurance scenario, local feature extraction uses three scales of convolution kernels (3, 5, and 7) to perform sliding processing on the insurance clause text: the 3-scale kernel extracts "subject-predicate structure" (such as "the insurer assumes liability for compensation"); the 5-scale kernel extracts "conditional adverbial clauses" (such as "if the insured is diagnosed with a critical illness, then..."); and the 7-scale kernel extracts "long-distance dependencies" (such as "the relationship between the insured's age, payment period, and sum insured").

[0065] Attention fusion calculates the semantic correlation between global features (BERT output) and local features (such as the attention weight of the global feature of "disclaimer" to local keywords such as "drunk driving" and "driving without a license"), and generates fused features by weighting them. These features are used for clause classification (liability-based / disclaimer-based) or customer need identification (claims application / clause consultation).

[0066] In the medical, health and elderly care scenarios, local feature extraction is performed on medical claim texts using 2 / 4 / 6 scale convolution kernels: 2-scale kernels extract "drug name + dosage" (e.g., "Gliclazide 80mg"); 4-scale kernels extract "visit time + treatment items" (e.g., "2024-01-05 Echocardiography"); and 6-scale kernels extract "disease diagnosis + reimbursement basis" (e.g., "Hypertension III based on the 2023 Medical Insurance Catalog").

[0067] Attention fusion strengthens the attention weight of the global feature of "nursing service type" to local features such as "service duration" and "payment method" in the elderly care scenario, and generates fusion features for disability level assessment, thus solving the semantic completion problem of "missing service record time nodes".

[0068] In some embodiments, updating the high-value data based on the erroneous sample data and the fused features includes: the erroneous sample data being misjudged or missed samples generated in a business process corresponding to a preset domain; optimizing the erroneous sample data using a generative adversarial network, generating new samples by combining the textual features, contextual information, and associated insurance clauses of the erroneous sample data with a generator, and judging the semantic features, syntactic structure, and contextual consistency of the new samples by a discriminator; and through adversarial training between the generator and the discriminator, making the generator output an optimized result that approximates real high-quality samples, and using the optimized result as the updated high-value data.

[0069] For erroneous data resulting from misjudgments or omissions in business operations, a Generative Adversarial Network (GAN) is used in conjunction with domain rules to generate high-quality repair samples: the generator integrates erroneous text features, context, and related terms / rules; the discriminator verifies the semantic compliance, syntactic integrity, and contextual consistency of the generated samples.

[0070] In financial insurance scenarios, error types include "vague descriptions of accident liability" in auto insurance claims (e.g., "multiple vehicle collisions without specifying the order of collisions") and "logical breaks in health disclosures" in health insurance claims (e.g., "treatment time for pre-existing conditions not filled in"). The generator mechanism: Inputting the BERT feature vector of the error text, related clauses, and context (description of accident scene images), a Transformer generator completes the missing liability elements (e.g., "Vehicle A collided with Vehicle B from the left front, and the traffic police determined Vehicle A was fully responsible").

[0071] The discriminator mechanism includes checking whether the generated samples contain the three elements of "responsible party", "cause of accident" and "basis of terms", and whether the semantics comply with insurance industry standards (such as the accurate use of terms such as "full responsibility" and "primary responsibility"). Training stops when the discrimination accuracy reaches 90% or higher.

[0072] In the context of healthcare and elderly care, common errors include "treatment items are not related to the diagnosis" in outpatient reimbursement (e.g., "patients with lumbar disc herniation are reimbursed for ophthalmological examinations") and "service records are missing key dates" in elderly care (e.g., "the start date of nursing services is not filled in").

[0073] The generator mechanism includes the introduction of the medical insurance reimbursement catalog and the ICD-10 disease coding library as prior knowledge. When the generator corrects the error of "diabetic patients missing blood glucose testing", it automatically supplements diagnostic evidence such as "fasting blood glucose 7.8 mmol / L, which meets the medical insurance reimbursement scope".

[0074] The discriminator mechanism ensures that the generated samples pass compliance checks by verifying the compliance of the diagnostic and treatment logic of the generated samples (such as the medical association between "drug indications" and "disease diagnosis") and the matching degree of medical insurance rules (such as the correspondence between "drugs in the catalog" and reimbursement ratios).

[0075] In some embodiments, updating the training set corresponding to the semantic understanding model based on the business data and the updated high-value data includes: collecting the business data in real time through a stream processing framework, merging the business data with the updated high-value data according to a dynamically adjusted update frequency; performing data augmentation processing on the merged text data, including synonym replacement based on word vector models, sentence rearrangement based on dependency parsing, and randomly inserting insurance-related noise words to form an updated training set.

[0076] Business data is collected through a real-time stream processing framework, merged with updated high-value data, and then augmented with domain-adapted data to form a dynamic training set. Data augmentation methods include: synonym replacement based on word vectors, sentence rearrangement based on dependency parsing, and insertion of domain-specific noisy words.

[0077] In financial and insurance scenarios, real-time data acquisition involves capturing text data streams of new insurance product terms (such as new regulations for "personal tax-deferred commercial pension insurance") and regulatory policy updates (such as revisions to the "Measures for the Management of Insurance Funds") using the Flink framework. Data augmentation includes: synonym replacement (e.g., "premium waiver" → "premium waiver," "liability exemption" → "liability exemption"); syntactic rearrangement (e.g., rearranging "If the insured is diagnosed with a critical illness during the waiting period, the company will not pay compensation" to "If the insured is diagnosed with a critical illness during the waiting period, the company will not be liable for compensation"); and noise insertion (randomly inserting colloquial words such as "excuse me" and "trouble" into customer inquiry texts to simulate real-world interaction scenarios).

[0078] In the healthcare and elderly care scenarios, real-time data is collected via Kafka, reflecting dynamic adjustments to the medical insurance catalog (such as the addition of "PD-1 inhibitor" reimbursement) and updates to elderly care service standards (such as the revision of the "Standards for Assessment of Elderly Capacity"). Data augmentation is achieved through synonym replacement: "coronary artery bypass grafting" → "coronary artery bypass surgery"; "medical insurance catalog" → "medical insurance drug catalog"; syntactic rearrangement, such as rearranging "patients receive rehabilitation treatment at designated medical institutions and meet the reimbursement conditions" to "rehabilitation treatment that meets the reimbursement conditions must be carried out at designated medical institutions"; and noise insertion, such as inserting ambiguous words like "suspected" and "possible" into medical claim text, to enhance the model's robustness to semantic uncertainty.

[0079] In some embodiments, training the semantic understanding model based on the updated training set includes: updating the semantic understanding model online using an incremental learning strategy, gradually inputting the updated training set into the semantic understanding model, and adjusting the model parameters to adapt the semantic understanding model to dynamic scenarios in a preset domain, thereby improving the semantic understanding model's ability to understand complex insurance scenarios.

[0080] By employing an incremental learning strategy to update the semantic understanding model online and gradually inputting a dynamic training set, the model can be adapted to dynamic changes in the domain (such as product iterations and policy updates) by adjusting parameters, thus avoiding "catastrophic forgetting".

[0081] In financial insurance scenarios, by adopting Elastic Weight Consolidation (EWC), when updating new data such as "new coverage of long-term care insurance", penalty constraints are imposed on key parameters of old knowledge (such as traditional health insurance terms) to retain existing semantic understanding capabilities.

[0082] The update strategy uses a mixed input approach, with new product terms accounting for 30% and historical high-value data accounting for 70%. It focuses on optimizing the semantic encoding parameters of attention heads related to the new coverage, such as terms like "nursing service type" and "payment cycle".

[0083] In the healthcare and elderly care scenarios, incremental learning algorithms combine knowledge distillation to encode the medical insurance expert review rules into soft labels (such as the implicit logic of "patients need to pay for drugs outside the catalog"), which guide the updating of model parameters during incremental training and ensure the retention of knowledge in low-frequency scenarios (such as "special reimbursement for rare diseases").

[0084] The update strategy merges newly added medical insurance policy data with historical claims data on a weekly basis to form a progressive training set; it also performs targeted optimization on cross-modal correlation parameters such as "age-drug dosage" and "disease stage-reimbursement ratio" to avoid misjudgment of old rules caused by policy updates.

[0085] The model's adaptation cycle to "insurance product terms iteration" was shortened from 72 hours to 4 hours, and its ability to understand complex terms remained stable (accuracy fluctuation ≤5%). Incremental learning improved the model's response accuracy to "annual adjustment of medical insurance catalog" by 30%, reduced the false negative rate in rare disease reimbursement scenarios by 40%, and significantly enhanced long-term stability.

[0086] In some embodiments, taking the insurance sector within the financial field as an example, insurance terms are complex and highly specialized, filled with numerous technical terms, intricate sentence structures, and nested conditional settings. Taking critical illness insurance as an example, the definitions of diseases in the terms often cite medical terminology, and the compensation conditions involve a variety of complex combinations of medical indicators and treatment methods. Ordinary customers, lacking professional medical and insurance knowledge, find it difficult to understand the meaning of the terms. Let the customer comprehension difficulty coefficient be , the technical terminology density be , the proportion of complex sentence structures be , and the number of nested conditional layers be , which can be roughly expressed as , where is the comprehensive influence function. The high difficulty coefficient makes it difficult for customers to accurately understand their rights and obligations when purchasing insurance, easily leading to disputes during claims, resulting in decreased customer satisfaction and increased policy cancellation rates.

[0087] Traditional risk assessment relies on static historical data. In the auto insurance sector, for example, it estimates the probability of an accident based on limited information such as the driver's past accident records and vehicle model, using simple statistical models, typically linear regression (as the error term). This ignores dynamic, multi-source information about the insured, such as social activity reflecting life stability, consumption behavior suggesting economic risk preferences, and health data directly related to personal insurance risk. This one-sided assessment leads to unreasonable insurance pricing: overcharging low-risk customers causes them to leave, while undercharging high-risk customers increases claims costs, impacting the insurance company's profitability and market competitiveness.

[0088] Some insurance companies have introduced AI-powered customer service, mostly based on simple keyword matching algorithms. Let the customer's inquiry text be and the keyword database be . If the text contains the keyword , a preset answer is given. However, when faced with complex questions, such as "How is the payout amount determined if a customer suffers from multiple complications during critical illness insurance coverage?", keyword matching is prone to failure. It cannot deeply understand semantic relationships, cannot accurately interpret the insurance terms, resulting in low accuracy and a poor customer experience.

[0089] Some companies use traditional machine learning algorithms to process risk assessment data, such as logistic regression and decision trees. Logistic regression has poor fitting ability for non-linear relationship data and cannot effectively construct a high-dimensional feature space when facing the challenge of fusing multi-source heterogeneous data. Taking health insurance risk assessment as an example, if the policyholder's social, consumption, and medical data are simply concatenated as feature vectors, the logistic regression model (with the label indicating whether an claim has occurred, the weight vector, and the bias) struggles to learn complex interaction features, resulting in large risk assessment biases and failing to meet the needs of refined insurance business.

[0090] This embodiment focuses on the insurance field, with the core idea of ​​improving the efficiency and accuracy of each link in the insurance business. On the one hand, it constructs an active learning sample selection strategy, using methods such as information entropy and cosine similarity to measure the uncertainty, representativeness, and difference from the model's prediction results of insurance samples. High-value samples are selected based on comprehensive scores to train the BERT model, strengthening its semantic understanding of insurance terms and accurately answering customer inquiries. On the other hand, it integrates BERT and CNN, using attention mechanisms to weightedly fuse features and employing multi-scale convolutional kernels to extract multi-level information. Simultaneously, for errors in insurance business, GAN technology is used to optimize errors by designing generators and discriminators, generating high-quality samples. Furthermore, it constructs a real-time data feedback and enhancement system, employing a stream processing framework to process data streams in real time, incremental learning strategies to update the model, and data augmentation techniques such as synonym replacement, enabling the model to adapt to changes in the insurance market, accurately understand customer needs, and promote the continuous development of the insurance business. See the specific structure below. Figure 2 .

[0091] Given the vast and complex nature of the insurance knowledge system, an active learning sample selection strategy is constructed. Let the insurance domain dataset be *x*, and *x* be a sample. Sample uncertainty is calculated, such as using information entropy to measure the distribution of technical terms in the text (where *x* is a technical term and *x* is its probability of occurrence), where *pi* represents the probability of each technical term occurring. ; Representativeness is achieved by calculating the text similarity between the sample and a database of common claims scenarios, using cosine similarity, where x and y represent the vector representations in the sample and claims scenario, respectively. ; To compare the model's current predictions with the labels initially predicted by the model, we use the Kronecker function (δ) to represent the difference between the model's initial predicted label y′ and the label y of the legal professionals. ; The overall score (as a weighted coefficient) is used to select high-value samples for priority training of the BERT model, improving the model's accuracy in understanding the semantics of insurance terms, quickly and accurately answering customer inquiries, and reducing the risk of information asymmetry. Here, w1, w2, and w3 are the weighted coefficients for three aspects: ; The samples are sorted according to their overall scores, and high-value samples are selected first for training the BERT model.

[0092] The BERT model training optimization based on active learning aims to improve the accuracy of the model's semantic understanding of insurance terms by prioritizing the training of BERT models with reasonably selected samples. This enables the model to quickly and accurately answer customer inquiries and reduce the risk of information asymmetry. Cross-model fusion and error example precision optimization leverage the advantages of BERT and CNN to enhance the perception of insurance data features, and utilize GANs to optimize and correct errors, thereby improving the accuracy and efficiency of insurance claims review, reducing misjudgments and omissions, and protecting the interests of both parties. All of these efforts are aimed at improving the quality and efficiency of insurance business-related processes.

[0093] 1. Model Fusion: BERT and CNN are combined for insurance business applications. BERT globally encodes the semantics of insurance text to obtain feature vectors, while CNN extracts features locally through convolutional kernels. These features are then fused (semicolons indicate concatenation) for subsequent tasks, such as insurance clause classification and customer needs identification, fully leveraging the advantages of both to enhance the model's ability to perceive insurance data features. Details of BERT and CNN fusion: (1) Feature fusion strategy: The global semantic feature vector generated by BERT can be weighted and fused with the local features extracted by CNN, rather than simply concatenating them. The weights of BERT and CNN features can be dynamically adjusted through the attention mechanism to ensure that the model selects the most relevant features in different tasks.

[0094] (2) Hierarchical feature extraction: In the CNN part, multi-scale convolution kernels (such as 3x3, 5x5, 7x7) can be used to extract local features of different granularities to ensure that the model can capture multi-level information in the insurance text.

[0095] 2. Error Optimization: After identifying errors in the insurance business process, GAN optimization is used. The generator generates new samples based on the characteristics of the errors, and the discriminator judges the sample quality. The goal is to maximize (the number of truly high-quality samples). The adversarial game between the two forces the generator to generate samples that approach high-quality insurance data, correcting problems such as semantic ambiguity and logical inconsistencies in claims case texts, improving the accuracy and efficiency of insurance claims review, reducing misjudgments and omissions, and protecting the interests of both parties.

[0096] (1) Generator design: The generator can adopt a variational autoencoder (VAE) or Transformer structure to ensure that the generated samples are more semantically coherent. The input of the generator can include the text features of the error cases, contextual information, and relevant insurance terms.

[0097] (2) Discriminator Design: The discriminator can employ a multilayer perceptron (MLP) or a convolutional neural network (CNN), combined with an adversarial training strategy to ensure that the generated samples can pass the discriminator's test. The input to the discriminator can include the semantic features, syntactic structure, and contextual consistency of the generated samples. For a multilayer perceptron (MLP), its loss function can be minimizing the cross-entropy loss: ; where y is the true label and y^ is the predicted probability distribution.

[0098] A real-time feedback mechanism is established, where optimized insurance data is immediately updated to update the training set. The update frequency is set to be dynamically adjusted based on business changes, ensuring the training set keeps pace with insurance market trends, such as changes in data characteristics caused by insurance product updates and regulatory policy changes. Details are as follows: (1) Real-time data stream processing: Stream processing frameworks (such as Apache Kafka or Apache Flink) can be used to process the data streams generated in insurance business in real time. Through real-time data stream processing, it is ensured that the model can obtain the latest insurance data in a timely manner and perform online learning and updates.

[0099] (2) Incremental learning: During the model update process, an incremental learning strategy can be adopted to ensure that the model can gradually learn new data distributions while retaining existing knowledge. Incremental learning can be achieved through techniques such as Elastic Weight Consolidation (EWC) or Memory Replay.

[0100] Data augmentation involves manipulating the text data of insurance policy terms using data augmentation techniques. Let the original text be , then be obtained through synonym replacement, sentence rearrangement, and random insertion of noise words. The augmented data is then merged with the original data to update the training set, improving model adaptability. For example, in insurance product recommendation scenarios, based on real-time updates to diverse training sets, it can accurately understand changes in customer needs, recommend suitable products, improve customer satisfaction and purchase conversion rates, and drive the continuous development of the insurance business. The following techniques can be used: (1) Synonym replacement: Pre-trained word vector models (such as Word2Vec, GloVe or FastText) can be used to find synonyms in the insurance field to ensure that the replaced text remains semantically consistent.

[0101] (2) Sentence rearrangement: Dependency parsing can be used to analyze the grammatical structure of sentences to ensure that the rearranged sentences are still grammatically correct.

[0102] (3) Noise insertion: Noise words related to the insurance field (such as "claims", "premium", "policy") can be randomly inserted into the text to ensure that the model can better handle noisy data during training.

[0103] In terms of customer service, the BERT model, trained and optimized based on active learning, can accurately understand insurance terms, quickly answer complex customer inquiries, reduce information asymmetry, improve customer satisfaction, and reduce policy cancellation disputes. In the model fusion and error example optimization phase, the combination of BERT and CNN enhances the perception of insurance data features, while GAN optimizes errors, making insurance claims review more accurate and efficient, avoiding misjudgments and omissions, and protecting the interests of both customers and insurance companies. A dynamic, real-time data feedback and enhancement system ensures that the training set keeps pace with market changes. Combined with data augmentation technology, the model becomes more adaptable, accurately identifying needs when recommending insurance products, improving customer purchase conversion rates, enhancing the profitability and market competitiveness of insurance companies, and comprehensively promoting the vigorous development of the insurance business.

[0104] Please see Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of the model training device 200 provided in the embodiments of this application. The model training device 200 is used to execute the steps of the model training methods shown in the above embodiments. The model training device 200 can be a single server or a server cluster, or the model training device 200 can be a terminal, such as a handheld terminal, a laptop computer, a wearable device, or a robot.

[0105] like Figure 3 As shown, the model training device 200 includes: The data calculation unit 201 is used to calculate the comprehensive score of each sample data in the dataset corresponding to the preset domain, so as to determine high-value data based on the comprehensive score; the comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a difference score based on the predicted label and the manual annotation. The preliminary training unit 202 is used to initially train a preset semantic understanding model based on the high-value data; the global semantic features corresponding to the high-value data output by the semantic understanding model are weighted and fused with the local features corresponding to the high-value data using an attention mechanism to form fused features; Data update unit 203 is used to obtain the error data corresponding to the preset domain, and update the high-value data according to the error data and the fusion features; The training completion unit 204 is used to collect business data in the preset domain in real time, update the training set corresponding to the semantic understanding model according to the business data and the updated high-value data, and complete the training of the semantic understanding model according to the updated training set to improve the semantic understanding model's semantic understanding ability in the preset domain.

[0106] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the model training device and each module described above can be referred to the corresponding processes in the model training method embodiments described above, and will not be repeated here.

[0107] The model training method described above can be implemented as a computer program, which can be used in, for example... Figure 3 It runs on the device shown.

[0108] Please see Figure 4 , Figure 4 This is a schematic block diagram of the structure of a computer device provided in an embodiment of this application. The computer device includes a processor, a memory, and a network interface connected via a device bus, wherein the memory may include a storage medium and internal memory.

[0109] The storage medium can store operating devices and computer programs. The computer program includes program instructions that, when executed, cause the processor to perform any model training method.

[0110] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0111] Internal memory provides an environment for the execution of computer programs stored in non-volatile storage media. When these computer programs are executed by the processor, the processor can perform any model training method.

[0112] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the terminal to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0113] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0114] In one embodiment, the processor is configured to run a computer program stored in memory to perform the following steps: For a dataset corresponding to a preset domain, a comprehensive score is calculated for each sample data in the dataset to determine high-value data based on the comprehensive score; the comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a difference score based on the predicted label and the manually labeled label. A pre-set semantic understanding model is initially trained based on the high-value data; the global semantic features corresponding to the high-value data output by the semantic understanding model are weighted and fused with the local features corresponding to the high-value data using an attention mechanism to form fused features; Obtain the faulty instance data corresponding to the preset domain, and update the high-value data based on the faulty instance data and the fusion features; Real-time collection of business data in the preset domain is used to update the training set corresponding to the semantic understanding model based on the business data and the updated high-value data; the semantic understanding model is trained based on the updated training set to improve the semantic understanding model's semantic understanding capability in the preset domain.

[0115] In some embodiments, calculating the comprehensive score of each sample data in the dataset to determine high-value data based on the comprehensive score includes: for each sample data, calculating the distribution of professional terms in the text using information entropy to obtain the uncertainty score; calculating the text similarity between the sample data and a preset scene library using cosine similarity to obtain the representativeness score; obtaining the difference score by comparing the corresponding initial prediction label with the professional annotation result; weighting and summing the uncertainty score, representativeness score, and difference score to obtain the comprehensive score; and sorting the sample data according to the comprehensive score to filter out a preset proportion of high-value data.

[0116] In some embodiments, the semantic understanding model is a BERT model; the preliminary training of the preset semantic understanding model based on the high-value data includes: inputting the high-value data into the BERT model for training, optimizing the parameters of the BERT model through the semantic features of insurance clause text in the high-value data, and improving the accuracy of the BERT model in semantic understanding of insurance domain text.

[0117] In some embodiments, the step of weightedly fusing the global semantic features corresponding to the high-value data output by the semantic understanding model with the local features corresponding to the high-value data using an attention mechanism to form a fused feature includes: performing local sliding processing of multi-scale convolutional kernels on the high-value data through a preset convolutional neural network model to extract local features of different granularities; inputting the global semantic feature vector output by the semantic understanding model and the local features into the attention mechanism, dynamically allocating weights according to the semantic correlation between the global semantic features and the local features, and generating a fused feature vector for insurance clause classification or customer demand identification tasks.

[0118] In some embodiments, updating the high-value data based on the erroneous sample data and the fused features includes: the erroneous sample data being misjudged or missed samples generated in a business process corresponding to a preset domain; optimizing the erroneous sample data using a generative adversarial network, generating new samples by combining the textual features, contextual information, and associated insurance clauses of the erroneous sample data with a generator, and judging the semantic features, syntactic structure, and contextual consistency of the new samples by a discriminator; and through adversarial training between the generator and the discriminator, making the generator output an optimized result that approximates real high-quality samples, and using the optimized result as the updated high-value data.

[0119] In some embodiments, updating the training set corresponding to the semantic understanding model based on the business data and the updated high-value data includes: collecting the business data in real time through a stream processing framework, merging the business data with the updated high-value data according to a dynamically adjusted update frequency; performing data augmentation processing on the merged text data, including synonym replacement based on word vector models, sentence rearrangement based on dependency parsing, and randomly inserting insurance-related noise words to form an updated training set.

[0120] In some embodiments, training the semantic understanding model based on the updated training set includes: updating the semantic understanding model online using an incremental learning strategy, gradually inputting the updated training set into the semantic understanding model, and adjusting the model parameters to adapt the semantic understanding model to dynamic scenarios in a preset domain, thereby improving the semantic understanding model's ability to understand complex insurance scenarios.

[0121] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method provided in any embodiment of this application.

[0122] The computer-readable storage medium may be an internal storage unit of the computer device described in the foregoing embodiments, such as the hard disk or memory of the computer device. The computer-readable storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, SmartMedia Card (SMC), Secure Digital (SD) card, or Flash Card equipped on the computer device.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A model training method, characterized in that, include: For a dataset corresponding to a preset domain, calculate the comprehensive score of each sample data in the dataset, and determine high-value data based on the comprehensive score; The comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a score based on the difference between predicted labels and manually labeled labels. A pre-set semantic understanding model is initially trained based on the high-value data; the global semantic features corresponding to the high-value data output by the semantic understanding model are weighted and fused with the local features corresponding to the high-value data using an attention mechanism to form fused features; Obtain the faulty instance data corresponding to the preset domain, and update the high-value data based on the faulty instance data and the fusion features; Real-time collection of business data in the preset domain is used to update the training set corresponding to the semantic understanding model based on the business data and the updated high-value data; the semantic understanding model is trained based on the updated training set to improve the semantic understanding model's semantic understanding capability in the preset domain.

2. The method according to claim 1, characterized in that, The calculation of a comprehensive score for each sample in the dataset, and the determination of high-value data based on the comprehensive score, includes: For each sample data, the distribution of professional terms in the text is calculated by information entropy to obtain the uncertainty score, the similarity between the sample data and the text of the preset scene library is calculated by cosine similarity to obtain the representativeness score, and the difference between the corresponding initial prediction label and the professional annotation result is obtained to obtain the difference score. The uncertainty score, representativeness score, and difference score are weighted and summed to obtain the comprehensive score; the sample data are sorted according to the comprehensive score, and a preset proportion of high-value data is selected.

3. The method according to claim 1, characterized in that, The semantic understanding model is the BERT model; the preliminary training of the preset semantic understanding model based on the high-value data includes: The high-value data is input into the BERT model for training. The parameters of the BERT model are optimized by the semantic features of the insurance clause text in the high-value data, thereby improving the accuracy of the BERT model in semantic understanding of insurance text.

4. The method according to claim 1, characterized in that, The step of weightedly fusing the global semantic features corresponding to the high-value data output by the semantic understanding model with the local features corresponding to the high-value data using an attention mechanism to form fused features includes: The high-value data is processed by local sliding of multi-scale convolutional kernels using a pre-defined convolutional neural network model to extract local features of different granularities. The global semantic feature vector output by the semantic understanding model is input into the attention mechanism along with the local features. Weights are dynamically allocated based on the semantic correlation between the global and local features to generate a fused feature vector, which is then used for insurance clause classification or customer need identification tasks.

5. The method according to claim 1, characterized in that, The step of updating the high-value data based on the faulty data and the fused features includes: The error sample data are samples of misjudgment or omission generated in the business process corresponding to the preset domain; Generative adversarial networks are used to optimize the faulty sample data. A generator combines the text features, contextual information, and related insurance clauses of the faulty sample data to generate new samples. A discriminator judges the semantic features, syntactic structure, and contextual consistency of the new samples. Through adversarial training between the generator and the discriminator, the generator outputs optimized results that approximate real high-quality samples, and these optimized results are used as updated high-value data.

6. The method according to claim 1, characterized in that, The step of updating the training set corresponding to the semantic understanding model based on the business data and the updated high-value data includes: The business data is collected in real time through a stream processing framework, and the business data is merged with the updated high-value data according to a dynamically adjusted update frequency. The merged text data is augmented by a combination of methods, including synonym replacement based on word vector models, sentence rearrangement based on dependency parsing, and random insertion of insurance-related noisy words to form an updated training set.

7. The method according to claim 1, characterized in that, The step of training the semantic understanding model based on the updated training set includes: An incremental learning strategy is adopted to update the semantic understanding model online. The updated training set is gradually input into the semantic understanding model. By adjusting the model parameters, the semantic understanding model is adapted to the dynamic scene of the preset domain, thereby improving the semantic understanding model's ability to understand complex insurance scenarios.

8. A model training device, characterized in that, include: The data calculation unit is used to calculate the comprehensive score of each sample data in the dataset corresponding to the preset domain, so as to determine high-value data based on the comprehensive score; The comprehensive score includes an uncertainty score based on information entropy, a representativeness score based on cosine similarity, and a score based on the difference between predicted labels and manually labeled labels. The initial training unit is used to initially train a preset semantic understanding model based on the high-value data; the global semantic features corresponding to the high-value data output by the semantic understanding model are weighted and fused with the local features corresponding to the high-value data using an attention mechanism to form fused features; The data update unit is used to acquire the error data corresponding to the preset domain, and update the high-value data according to the error data and the fusion features; The training completion unit is used to collect business data in the preset domain in real time, update the training set corresponding to the semantic understanding model based on the business data and the updated high-value data, and complete the training of the semantic understanding model based on the updated training set to improve the semantic understanding model's semantic understanding ability in the preset domain.

9. A computer device, characterized in that, The computer device includes a memory and a processor; The memory is used to store computer programs; The processor is configured to execute the computer program and, in executing the computer program, implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to implement the method as described in any one of claims 1 to 7.

Images