Sample expansion method, computer equipment and computer readable storage medium

Abstract

The invention relates to the technical field of model construction, is suitable for the financial field, and provides a sample expansion method, computer equipment and a computer readable storage medium, and the method comprises the steps: obtaining an expanded sample based on the information type of each sample content in a real sample; calculating the semantic similarity between the expansion sample embedding vector and a vector in a real data embedding index corresponding to the real sample; determining a structural consistency score based on the structured data in the expanded sample; determining semantic stability based on the output embedded representation of the extended sample in the plurality of models; determining the credibility of the expanded sample based on the semantic similarity, the structural consistency score, the semantic stability and the output result confidence of the current initial model; and if the credibility is greater than or equal to a preset credibility threshold, determining that the expanded sample is an effective sample. According to the scheme, training data is constructed for rare class services in a targeted manner, and more diversified reliable training samples are provided for a downstream service model.

Description

Technical Field

[0001] This application relates to the field of model building technology, applicable to the financial field, and particularly to a sample expansion method, computer device, and computer-readable storage medium. Background Technology

[0002] In financial business scenarios, including various insurance businesses, the accuracy of machine learning models depends on high-quality training data. However, when faced with rare business scenarios such as rare claims events, abnormal transaction behaviors, special types of customer complaints, and rare fraud cases, real business data is often extremely scarce, directly affecting the performance and robustness of downstream models.

[0003] Current data augmentation methods typically employ relatively simple text synonym replacement or image geometric transformations to generate augmented samples. These methods offer limited data variation and struggle to cover complex and ever-changing real-world business scenarios. Furthermore, the generated augmented samples often deviate from real-world business scenarios in terms of semantics, style, or structure. Introducing these samples into the database for training may introduce noise, leading to decreased model performance. Additionally, samples from few-sample events often require manual screening and annotation by experts, which is costly, time-consuming, and heavily reliant on the annotator's skill level, making it difficult to meet the demands of rapidly iterating financial business needs.

[0004] Therefore, how to automatically generate high-quality augmented samples for rare business scenarios and improve the model's ability to handle rare business scenarios has become an urgent technical problem to be solved. Summary of the Invention

[0005] This application provides a sample expansion method, a computer device, and a computer-readable storage medium, aiming to solve the technical problem in related technologies where the number of real samples in rare category business scenarios is small and the expansion is difficult, resulting in inaccurate model output results.

[0006] In a first aspect, embodiments of this application provide a sample expansion method, including: For each real sample in the real sample set of the current initial model, based on the information type of each sample content in the real sample, the content transformation process is performed on each sample content to obtain the expanded sample; The augmented sample is converted into an augmented sample embedding vector, and the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample is calculated. Based on the structured data in the expanded sample, the structural consistency score of the expanded sample is determined; Based on the output embedding representation of the expanded sample in multiple models, the semantic stability of the expanded sample is determined, wherein the multiple models include the current initial model, the historical models corresponding to the previous multiple iterations of the current initial model, and the teacher model; The credibility of the augmented samples is determined based on the semantic similarity, the structural consistency score, the semantic stability, and the confidence of the output of the current initial model. If the credibility is greater than or equal to a predetermined credibility threshold, the expanded sample is determined to be a valid sample, and the expanded sample is used to train the downstream specific business model of the current initial model.

[0007] Optionally, in one embodiment of this application, if the information type of the sample content includes text, images, and structured data, then the step of performing content transformation processing on each item of sample content based on the information type of each item in the real sample to obtain an expanded sample includes: For any of the sample contents of the real sample, If the information type of the sample content is text, the sample content is processed by one or more of the following methods through a large language model: synonym replacement processing, template variation processing, and contextual expansion processing, to obtain the transformed text, which is used as the sample content of the expanded sample. If the information type of the sample content is an image, a new image is generated based on the sample content and a predetermined conditional diffusion model, and used as the sample content of the expanded sample; or, the sample content is subjected to style transformation processing and / or viewpoint transformation processing to obtain a transformed image, which is used as the sample content of the expanded sample. If the information type of the sample content is structured data, a new structured data with the same format and the same field value relationships is generated based on the sample content and a predetermined condition generation model, and this new structured data is used as the sample content of the expanded sample.

[0008] Optionally, in one embodiment of this application, converting the augmented sample into an augmented sample embedding vector includes: The text, images, and structured data in the augmented sample are encoded separately, and the encoding results are projected into a unified semantic space to obtain the text vector, image vector, and structured data vector of the augmented sample. The text vector, image vector, and structured data vector of the augmented sample are fused using an attention fusion network to obtain the augmented sample embedding vector corresponding to the augmented sample; or, The text vector, image vector, and structured data vector of the augmented sample are concatenated, and the concatenation result is subjected to nonlinear transformation and feature fusion processing through a multilayer perceptron to obtain the augmented sample embedding vector corresponding to the augmented sample.

[0009] In one embodiment of this application, optionally, before performing content transformation processing on each sample content based on the information type of each sample content in the real sample to obtain the expanded sample, the method further includes: The text, images, and structured data in the real sample are encoded respectively, and the encoding results are projected into the unified semantic space to obtain the text vector, image vector, and structured data vector of the real sample; The text vector, image vector, and structured data vector of the real sample are fused using an attention fusion network to determine the real data embedding index corresponding to the real sample; or, The text vector, image vector, and structured data vector of the real sample are concatenated, and the concatenation result is subjected to nonlinear transformation and feature fusion processing through a multilayer perceptron to determine the real data embedding index corresponding to the real sample.

[0010] In one embodiment of this application, optionally, calculating the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample includes: Cluster the vector sets corresponding to the real sample set to obtain multiple vector clusters; Calculate the centroid of each vector cluster, and based on the centroid and the augmented sample embedding vector, determine the first semantic similarity between the augmented sample and each vector cluster; For the vector cluster corresponding to the highest first semantic similarity, the second semantic similarity between the expanded sample and each real sample in the vector cluster is determined based on the vector of each real data embedding index in the vector cluster and the embedding vector of the expanded sample. The highest second semantic similarity is determined as the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample.

[0011] In one embodiment of this application, optionally, determining the structural consistency score of the expanded sample based on the structured data in the expanded sample includes: The system checks whether each field in the structured data meets predetermined field validity conditions, wherein the field validity conditions include: the field is complete, the field type of the field is consistent with the required field type of the field at the location of the field, and the value of the field is within the predetermined field range corresponding to the location of the field. Based on the total number of fields in the structured data and the number of valid fields in the structured data that meet the valid conditions of the fields, the ratio of the number of valid fields to the total number of fields is determined as the structural consistency score of the augmented sample.

[0012] In one embodiment of this application, optionally, determining the semantic stability of the augmented sample based on its output embedding representation in multiple models includes: Based on the output embedding representation of the augmented sample in the multiple models, calculate the cosine similarity of the output embedding representation of the augmented sample for every two models in the multiple models; Based on the cosine similarity and the total number of the multiple models, the semantic stability of the augmented samples is determined, wherein, , This refers to the expanded sample. This indicates the semantic stability of the augmented samples. For the i-th model among the multiple models, For the j-th model among the multiple models, This indicates cosine similarity, where N is the total number of the multiple models.

[0013] In one embodiment of this application, optionally, determining the credibility of the augmented samples based on the semantic similarity, the structural consistency score, the semantic stability, and the confidence level of the output of the current initial model includes: The confidence level of the expanded sample is obtained by weighted summation of the semantic similarity, the structural consistency score, the semantic stability, and the confidence level of the output of the current initial model.

[0014] In a second aspect, embodiments of this application provide a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being configured to perform the method described in the first aspect above.

[0015] Thirdly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions for performing the method described in the first aspect above.

[0016] The above technical solution addresses the technical problem of inaccurate model output results caused by the limited number of real samples and the difficulty in expanding rare categories of business scenarios. For sample content of different information types in real samples, appropriate content transformation methods are used to obtain expanded samples. Subsequently, the expanded samples are mapped to a unified semantic space to generate embedding vectors, and their semantic similarity with real data is comprehensively evaluated. In addition, the compliance of the expanded sample at the structural level, the semantic stability shown in different models, and the confidence of the current initial model in its output results are also evaluated. Finally, a credibility reflecting the quality of the expanded sample is calculated as the basis for judging whether the expanded sample can be used for downstream model training. Therefore, when facing rare business scenarios such as rare claims events, abnormal transaction behaviors, special customer complaint types, and rare fraud cases, it can efficiently and automatically expand the training data of rare categories without manual annotation. This overcomes the shortcomings of related technologies, such as limited variation of data augmentation methods, uncontrollable quality of generated samples, and reliance on manual annotation. It improves the efficiency of sample data expansion, can construct training data for rare business categories in a targeted manner, and ultimately provides more diverse and reliable training samples for downstream business models, which helps to improve the accuracy of the output results of downstream business models. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart of a sample augmentation method according to an embodiment of this application is shown; Figure 2 A flowchart of a sample augmentation method according to another embodiment of this application is shown; Figure 3 A block diagram of a computer device according to one embodiment of this application is shown; Figure 4 A block diagram of a computer device according to another embodiment of this application is shown. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Figure 1 A flowchart of a sample expansion method according to an embodiment of this application is shown.

[0021] like Figure 1 As shown, a sample augmentation method according to an embodiment of this application includes: Step 102: For each real sample in the real sample set of the current initial model, based on the information type of each sample content in the real sample, perform content transformation processing on each sample content to obtain an expanded sample.

[0022] The current initial model serves as the foundation for training downstream models that handle the actual business scenarios, using sample data from real-world business operations. Before training the downstream models, it's essential to obtain a sample set for the current initial model. When facing rare business scenarios such as unusual claims events, abnormal transaction behaviors, special customer complaint types, or rare fraud cases, the number of real samples is insufficient to train a downstream model with sufficiently high accuracy. Therefore, it's necessary to first expand the downstream model with sufficient supplementary samples, based on real data.

[0023] Specifically, for each part of the sample content in the real sample, the corresponding content transformation processing method can be applied to transform the sample content, resulting in expanded samples. Thus, for the specific content and content category in the real sample, expanded samples can be customized to match the real sample in both specific content and content category. The information type of the sample content includes, but is not limited to, text, images, and structured data, or any type that meets the training requirements of the actual sample.

[0024] In one possible design, for any of the sample contents of the real sample, if the information type of the sample content is text, the sample content is processed by one or more of the following through a large language model: synonym substitution, template variation, and contextual expansion, to obtain the transformed text, which is then used as the sample content of the expanded sample.

[0025] Synonym substitution replaces words in the original text with synonyms or near-synonyms without altering its core semantics. Template variation fills the template with the original text according to its fixed structure, generating new text, or fills or transforms the variable parts of the template based on a fixed structure, generating new text that is structurally similar but different in content. Contextual expansion extends the original text by adding background information, detailed descriptions, and other relevant descriptions, while retaining its core information. These three transformation methods preserve the original semantics or, while preserving the original semantics, horizontally expand or deeply explore the original text, thus providing reasonable expanded text that matches the real sample.

[0026] In one possible design, for any of the sample contents of the real sample, if the information type of the sample content is an image, a new image is generated based on the sample content and a predetermined conditional diffusion model, and used as the sample content of the expanded sample; or, the sample content is subjected to style transformation processing and / or perspective transformation processing to obtain a transformed image, which is used as the sample content of the expanded sample.

[0027] Conditional diffusion models aim to guide the generation of new images using target conditions, synthesizing semantically consistent and detailed new images from input images. Style transformation aims to alter the visual effect of images to simulate their appearance under different shooting conditions. Viewpoint transformation is used to adjust the viewing angle of an image, thereby simulating the visual effect of photographing the same object from different positions. These three methods effectively enrich and expand the diversity of image samples, helping to improve the model's robustness to visual changes, thus providing training samples with stronger generalization capabilities for downstream tasks.

[0028] In one possible design, for any of the sample contents of the real sample, if the information type of the sample content is structured data, a new structured data with the same format and the same field value relationships is generated based on the sample content and a predetermined conditional generation model, and this new structured data is used as the sample content of the expanded sample.

[0029] Conditional generation models can generate new structured data based on structured data from real samples and the relational constraints within that structured data. These new structures have the same format and the same relationships between field values. Therefore, while strictly adhering to the business logic embodied in the structured data, high-quality structured training samples can be automatically synthesized to expand the original sample set, providing downstream models with more high-quality samples adapted to their actual business needs.

[0030] Step 104: Convert the augmented sample into an augmented sample embedding vector, and calculate the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample.

[0031] The augmented sample embedding vector reflects the deep semantic feature representation of the augmented sample in a unified semantic space. Similarly, the vector in the real data embedding index corresponding to the real sample reflects the deep semantic feature representation of the real sample in a unified semantic space. The semantic similarity between the two reflects the similarity level between the augmented sample and the real sample in the semantic dimension, and can be used as one of the criteria for judging whether the augmented sample is a qualified high-quality sample.

[0032] In one possible design, the specific method for converting the augmented sample into an augmented sample embedding vector includes: encoding the text, image, and structured data in the augmented sample separately, and projecting the encoding results into a unified semantic space to obtain the text vector, image vector, and structured data vector of the augmented sample; and fusing the text vector, image vector, and structured data vector of the augmented sample through an attention fusion network to obtain the augmented sample embedding vector corresponding to the augmented sample.

[0033] The text, images, and structured data in the augmented samples are encoded separately. The most suitable encoding method can be selected based on the characteristics of different information types in the augmented samples, thereby obtaining the most accurate feature representation for the augmented samples. The attention fusion network can integrate feature representations from different information types to form a unified vector representation for the augmented samples that comprehensively reflects their semantics.

[0034] In another possible design, the specific method for converting the augmented sample into an augmented sample embedding vector includes: encoding the text, image, and structured data in the augmented sample separately, and projecting the encoding results into a unified semantic space to obtain the text vector, image vector, and structured data vector of the augmented sample; concatenating the text vector, image vector, and structured data vector of the augmented sample, and performing nonlinear transformation and feature fusion processing on the concatenation result through a multilayer perceptron to obtain the augmented sample embedding vector corresponding to the augmented sample.

[0035] By concatenating the text vectors, image vectors, and structured data vectors of the augmented samples, it is ensured that the feature representations under different information types are completely preserved, serving as the basis for subsequent feature fusion. Furthermore, a multilayer perceptron (MLP) is used to perform nonlinear transformations and dimension mapping on the concatenated multimodal vectors to achieve deep feature fusion.

[0036] Of course, before step 102, the real samples also need to undergo the same processing to compare the feature fusion results of the two and calculate semantic similarity. Specifically, the text, image, and structured data in the real samples can be encoded separately, and the encoding results can be projected into the unified semantic space to obtain the text vector, image vector, and structured data vector of the real samples; the text vector, image vector, and structured data vector of the real samples can be fused through an attention fusion network to determine the real data embedding index corresponding to the real samples; or, the text vector, image vector, and structured data vector of the real samples can be concatenated, and the concatenation result can be subjected to nonlinear transformation and feature fusion processing through a multilayer perceptron to determine the real data embedding index corresponding to the real samples.

[0037] Step 106: Determine the structural consistency score of the expanded sample based on the structured data in the expanded sample.

[0038] Structured data is a structured representation of the attributes involved in the business to which the sample belongs and the logical relationships between the attributes. The resulting structure consistency score reflects the performance of the expanded sample in terms of field completeness, field type compliance, and numerical validity, and can be used as one of the criteria for judging whether the expanded sample is a qualified high-quality sample.

[0039] Step 108: Determine the semantic stability of the expanded sample based on the output embedding representation of the expanded sample in multiple models, wherein the multiple models include the current initial model, the historical models corresponding to the previous multiple iterations of the current initial model, and the teacher model.

[0040] The output embeddings of augmented samples across multiple models reflect the differences in the understanding of the same augmented sample by the current initial model, historical models corresponding to previous iterations of the current initial model, and teacher models. The resulting semantic stability reflects the semantic stability of the same augmented sample when facing semantic analysis from different models. In other words, the smaller the difference in semantic evaluation of the same augmented sample across different models, the higher its semantic stability. That is, semantic stability is an evaluation of the augmented sample using different models as different dimensions, and therefore can be used as one of the criteria for judging whether an augmented sample is a qualified high-quality sample.

[0041] Step 110: Determine the credibility of the expanded samples based on the semantic similarity, the structural consistency score, the semantic stability, and the confidence of the output of the current initial model.

[0042] The confidence score of the current initial model's output can be obtained after the initial model has been trained. For example, it can be derived from the internal evaluation score of the current initial model's own prediction results, such as the maximum class probability output by a classification model or the confidence estimate of the prediction interval by a regression model. Based on this, the semantic similarity, structural consistency score, semantic stability, and output confidence score all reflect the reliability of the expanded samples as training samples for downstream specific business models in their respective dimensions. Therefore, the credibility of the expanded samples can be calculated using these four parameters, and the credibility of the expanded samples serves as the final basis for determining whether the expanded samples should be used as training samples for downstream specific business models.

[0043] Step 112: If the credibility is greater than or equal to a predetermined credibility threshold, the expanded sample is determined to be a valid sample, and the expanded sample is used to train the downstream specific business model of the current initial model.

[0044] If the credibility is greater than or equal to the predetermined credibility threshold, it indicates that the quality of the expanded sample is sufficient to support its use as a training sample for a downstream specific business model. In this case, the expanded sample can be selected as a training sample for the downstream specific business model.

[0045] The above technical solution employs adapted content transformation methods to obtain augmented samples for different information types in real samples. These augmented samples are then mapped to a unified semantic space to generate embedding vectors. The semantic similarity between these vectors and the real data is comprehensively evaluated. Furthermore, the structural compliance, semantic stability across different models, and confidence level of the current initial model's output are also assessed. Finally, a credibility index reflecting the quality of the augmented sample is calculated, serving as the basis for determining whether it can be used for downstream model training. Therefore, when facing rare business scenarios such as unusual claims events, abnormal transaction behaviors, special customer complaint types, and rare fraud cases, this solution can efficiently and automatically augment training data for rare categories without manual annotation. It overcomes the shortcomings of related technologies, such as limited variation in data augmentation methods, uncontrollable sample quality, and reliance on manual annotation. This improves the efficiency of sample data augmentation, enabling targeted construction of training data for rare business categories. Ultimately, it provides more diverse and reliable training samples for downstream business models, contributing to improved accuracy of their output.

[0046] Figure 2 A flowchart of a sample expansion method according to another embodiment of this application is shown.

[0047] like Figure 2 As shown, a sample expansion method according to another embodiment of this application includes: Step 202: For each real sample in the real sample set of the current initial model, based on the information type of each sample content in the real sample, perform content transformation processing on each sample content to obtain an expanded sample.

[0048] Step 204: Convert the augmented sample into an augmented sample embedding vector.

[0049] Step 206: Cluster the vector set corresponding to the real sample set to obtain multiple vector clusters.

[0050] Step 208: Calculate the centroid of each vector cluster, and determine the first semantic similarity between the expanded sample and each vector cluster based on the centroid and the expanded sample embedding vector.

[0051] Each vector cluster is a group of semantically highly similar real samples, and the centroid of each vector cluster is the semantic core feature of that group of real samples in a unified semantic space. By clustering the set of real samples into multiple clusters, and then calculating the first semantic similarity between the centroids of multiple clusters and the expanded samples, the vector cluster corresponding to the highest first semantic similarity can be determined. This can narrow the similar sample range of the expanded samples from the entire set of real samples to this vector cluster.

[0052] Step 210: For the vector cluster corresponding to the highest first semantic similarity, determine the second semantic similarity between the expanded sample and each real sample in the vector cluster based on the vector of each real data embedding index in the vector cluster and the embedded vector of the expanded sample.

[0053] Step 212: The highest second semantic similarity is determined as the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample.

[0054] After narrowing the similarity interval of the expanded sample from the entire set of real samples to this vector cluster, the second semantic similarity between each real sample in this vector cluster and the expanded sample can be calculated. The real sample corresponding to the highest second semantic similarity is then selected. In this way, the similarity interval of the expanded sample can be narrowed down from this vector cluster to the most similar real sample. The resulting semantic similarity is the degree of similarity between the most similar real sample and the expanded sample.

[0055] Step 214: Detect whether each field in the structured data meets the predetermined field validity conditions, wherein the field validity conditions include: the field is complete, the field type of the field is consistent with the required field type of the field at the location of the field, and the value of the field is within the predetermined field range corresponding to the location of the field.

[0056] Step 216: Based on the total number of fields in the structured data and the number of valid fields in the structured data that meet the valid conditions of the fields, determine the ratio of the number of valid fields to the total number of fields as the structural consistency score of the expanded sample.

[0057] The presence of complete fields indicates that the structured data in the expanded sample covers all necessary information items for the corresponding business in terms of structural integrity. Consistent field types indicate that the data format of each field value in the expanded sample's structured data conforms to the requirements of the corresponding business. Field values ​​within a predetermined range reflect that the specific values ​​of the fields in the expanded sample's structured data meet the numerical range requirements of the corresponding business.

[0058] Based on this, the ratio of the number of valid fields in the structured data of the expanded sample to the total number of fields is calculated as the structure consistency score of the expanded sample. This can quantify the degree of matching between the structured data of the expanded sample and the actual corresponding business, and serve as one of the criteria for judging whether the expanded sample is a qualified high-quality sample.

[0059] Step 218: Based on the output embedding representation of the expanded sample in the plurality of models, calculate the cosine similarity of the output embedding representation of the expanded sample to every two models in the plurality of models.

[0060] Step 220: Determine the semantic stability of the augmented samples based on the cosine similarity and the total number of the multiple models.

[0061] , This refers to the expanded sample. This indicates the semantic stability of the augmented samples. For the i-th model among the multiple models, For the j-th model among the multiple models, This indicates cosine similarity, where N is the total number of the multiple models.

[0062] The cosine similarity of the output embeddings between any two models reflects the consistency of their semantic understanding of the same augmented sample, while the sum of the cosine similarities of the output embeddings between any two models across all models... This reflects the consistency of semantic representation understanding of the augmented sample across all participating models. In other words, the semantic stability of the final augmented sample reflects the average level of semantic understanding consistency among all participating models. The higher this average level, the more similar the understanding results the augmented sample can achieve across different models, and the higher the quality of the augmented sample. Therefore, the semantic stability of the augmented sample can be used as one of the criteria for judging whether an augmented sample is a qualified high-quality sample.

[0063] Step 222: Perform a weighted summation of the semantic similarity, the structural consistency score, the semantic stability, and the confidence score of the output of the current initial model to obtain the confidence score of the expanded sample.

[0064] Finally, the semantic similarity, structural consistency score, semantic stability, and confidence score of the current initial model output are weighted and summed. The weights of these four components reflect their respective contributions to the quality evaluation of the expanded samples.

[0065] The above technical solution, based on real samples, performs targeted content transformations according to different information types to generate augmented samples. These augmented samples are then embedded into a unified semantic space to form vector representations. Further, clustering and multi-level similarity calculations determine the optimal semantic closeness between the augmented samples and the real data. Simultaneously, compliance checks are performed on the fields of the generated structured data, and the semantic stability of the augmented samples in multiple models is calculated. Finally, combined with the confidence level of the initial model, the usability of the augmented samples is determined. Therefore, when facing rare business scenarios such as rare claims events, abnormal transaction behaviors, special customer complaint types, and rare fraud cases, the solution can efficiently and automatically augment training data for rare categories without manual annotation. This overcomes the shortcomings of related technologies, such as limited variation in data augmentation methods, uncontrollable sample quality, and reliance on manual annotation. It improves the efficiency of sample data augmentation, enabling targeted construction of training data for rare business categories, ultimately providing more diverse and reliable training samples for downstream business models, thus helping to improve the accuracy of downstream business model output results.

[0066] In one possible real-world scenario, an insurance company is training an intelligent model to automatically review "glass breakage alone" claims in auto insurance. Cases involving only glass breakage without other vehicle damage represent a very small percentage of overall claims, resulting in a limited number of real-world samples available for training. Optionally, real-world samples may include the following three components.

[0067] 1. Text content, i.e., the claim description: "While driving, the windshield was struck by a stone kicked up by a truck in front, causing radial cracks in the glass. No other parts were damaged." 2. Image content, i.e., an accident photograph: a clear photograph showing the radial cracks in the vehicle's windshield. 3. Structured data, i.e., the claim form: an electronic form containing fields such as the time of the accident, license plate number, vehicle model, damaged parts (windshield), and estimated repair cost.

[0068] At this point, firstly, based on the real samples, expanded samples are generated through content transformation.

[0069] During text expansion, synonym replacement, template variation, and contextual expansion can be performed.

[0070] Specifically, in the synonym replacement processing, a large language model can be used to replace synonyms in the original text, generating descriptions such as "During driving, gravel kicked up by the tires of the vehicle in front struck the windshield, causing spiderweb-like cracks in the glass; the rest of the vehicle was intact." In the template variation processing, descriptions with different structures can be generated based on the fixed sentence template of the claims report, such as "The complainant stated: The windshield of the vehicle was damaged due to an impact from a foreign object. Upon investigation, the damage was limited to the glass." In the context expansion processing, details can be added to generate detailed descriptions such as "While driving on the highway, following a truck fully loaded with sand and gravel, the truck's rear wheel kicked up a stone that struck the lower left corner of the windshield of my vehicle at high speed, immediately creating radial cracks about the size of a palm. Upon inspection, it was confirmed that the paint, bumper, and headlights were free of any scratches or damage."

[0071] In image augmentation, new images can be generated based on a conditional diffusion model. That is, using the original photo of the glass crack as a condition, a new photo with different car models and different crack patterns can be generated, but the semantic meaning is still "only glass is broken". Style / viewpoint transformation can also be performed to process the original photo into visual effects under different weather conditions, or to simulate the perspective of taking a photo of the glass crack from inside the car or from the side to generate new images.

[0072] In the expansion of structured data, new data can be generated based on the conditional generation model. That is, a new set of structured data is generated based on the structure of the original claim form and the relationship between its fields—the correspondence between vehicle model and estimated repair amount. In this new set of data, the values ​​of fields such as license plate number, accident time, and specific amount change, but the damaged part is still the windshield. All field types, value ranges, and logical relationships comply with the rules of auto insurance business.

[0073] Next, semantic similarity is calculated. The generated augmented samples are multimodal fusion encoded to obtain their augmented sample embedding vectors. Similarly, all real historical claims samples are encoded and a real data embedding index is constructed. The cosine similarity between this augmented sample vector and all real sample vectors in the index is calculated, and the highest value is taken as its semantic similarity. This semantic similarity reflects how closely the generated glass breakage case is semantically similar to the real case.

[0074] Furthermore, determine the structural consistency score. Specifically, this can be done by checking the newly generated claim form, including the following aspects.

[0075] 1. Are all fields complete? Are any required fields such as the complainant and the loss adjuster missing?

[0076] 2. Are the field types reasonable? Is the accident time in date format? Is the estimated repair cost a number?

[0077] 3. Whether the value is within the business's allowable range: Whether the repair cost is within the reasonable range of the market price of glass for this model.

[0078] Based on this, the proportion of compliant fields to the total number of fields is calculated to obtain the structural consistency score. For example, if all 10 fields are compliant, the score is 1.0.

[0079] Next, semantic stability is determined by inputting the same generated text description of "broken glass" into the current initial model, several historical models, and the teacher model. For the embedding results output by each model, the cosine similarity between each pair of embedding results is calculated, and the average of all cosine similarities is taken to obtain the semantic stability.

[0080] Finally, the semantic similarity, structural consistency score, semantic stability, and the fraud risk probability identified by the current initial model for this generated case are weighted and fused to obtain the credibility score.

[0081] A predetermined confidence threshold is set to 0.85. If the confidence of an augmented sample is only 0.82, the sample is invalid and will not be adopted by the downstream business model. If the confidence of another augmented sample reaches 0.90, the sample will be judged as a valid sample and will be adopted by the downstream business model.

[0082] Ultimately, all the high-confidence augmented samples that pass the screening will be merged with the original small number of real "isolated glass breakage" samples to form a training set. This set will be used to train or fine-tune the downstream automatic claims review model, thereby significantly improving the accuracy and generalization ability of the automatic claims review model in identifying such rare but important claims cases.

[0083] In one embodiment of this application, a sample augmentation system is provided. This system uses the prediction error, uncertainty, and cross-version consistency of the business model as core signals to dynamically guide the generation of high information gain and high reliability data samples for model orientation synthesis. It also uses a multi-dimensional quality evaluation mechanism to screen high-quality samples and prevent noisy samples from entering the training process, thereby achieving continuous improvement in model performance with minimal manual annotation costs.

[0084] The system comprises several functional modules, which form a strict sequential dependency and feedback loop.

[0085] An embedded index refers to a vectorized retrieval index structure constructed by mapping historical real business data (text, images, structured forms, etc.) into vector representations, used to support efficient similar sample search and matching. In this invention, the embedded index serves to provide a real business reference space for generated samples; determine whether generated samples deviate from the real data distribution; and support subsequent credibility assessment and screening mechanisms.

[0086] In the embedded index module, the process for building the embedded index is as follows: Step S1: Real data collection, obtaining a real sample set from historical business systems: Each of them It can contain: text (claim description, complaint text, transaction notes, etc.), images (accident photos, scanned invoices), and structured fields (amount, time, type of insurance, field key-value pairs).

[0087] Step S2: Multimodal unified embedding encoding, encoding different modalities separately and projecting them into a unified semantic space. The text encoding result is as follows: , The text in the sample can be encoded using the FinBERT / LLaMA-Fin model. The image encoding result is as follows: , For the images in the sample, the CLIP / ViT model can be used. The structured data encoding result is: , This refers to the structured data in the sample.

[0088] Step S3: Multimodal fusion, fusing multimodal embeddings into a unified vector. : ,in, It can be used for splicing and MLP processing, or for attention fusion networks.

[0089] Step S4: Construct a vector index by embedding all real samples into a vector set and then storing it in a vector database to form a real data embedding index. .

[0090] This index supports Top-K similar sample search, and supports searches based on cosine similarity, inner product, Euclidean distance, and other metrics.

[0091] In the model error and uncertainty analysis module, the current business model is analyzed. The prediction error is calculated on the validation set or online samples. .in, For business samples; For true labels or high-confidence reference labels; For the model prediction results, The loss function is such as cross-entropy or mean squared error.

[0092] When identifying the target generation region, the learning blind spots of the model can be identified by using thresholding or clustering methods. }, and This is the threshold baseline. This area indicates that the model has a large prediction error and the model is uncertain about the outcome, but it is representative of real-world business scenarios.

[0093] In the reverse drive generation control module, the conditional control signal is: ,in, Consistent with the dimensions of the latent space within the generative model, the vector sources are as follows.

[0094]

[0095] That is, extract the sample embeddings within the R_target region from the embedding index and calculate their semantic centers. This vector is not designed manually, but is the semantic center of the capability blind zone defined in reverse by model errors and uncertainties. The modulation generation distribution is achieved by adding modulation to the intermediate layer of the generation model, which serves as a latent space bias term. .in, This is the hidden state of layer l. For learnable modulation coefficients, Instead of being used as an explicit textual condition, it serves as a semantic bias vector for the generation path, influencing the entire generation trajectory. Generation is no longer random or balanced sampling, but is driven inversely by model error.

[0096] This significantly reduces invalid synthetic data, generates a sample set to repair blind spots in the model's capabilities, and improves model performance faster at the same generation cost.

[0097] In summary, this invention constructs a sustainable, self-evolving, and low-cost few-sample data generation system by introducing a targeted generation mechanism driven by model error and uncertainty, and a sample quality discrimination mechanism based on cross-model version semantic stability. This system significantly differs from traditional random generation or rule-driven data augmentation schemes. Specifically, it achieves a significant increase in the amount of rare category data by automatically expanding the few-sample categories through model generation, retrieval matching, and refinement filtering, thereby improving model training coverage. It also ensures high-quality data, with generated samples closely resembling the real domain, and guarantees data quality through hierarchical credibility control and manual sampling. Furthermore, it demonstrates strong multimodal applicability, supporting various data types such as text, images, and structured forms, adapting to complex scenarios in financial and insurance businesses. Finally, it closes the entire process of generation, filtering, training, and feedback without extensive manual intervention, continuously improving the performance of few-sample category models. This makes it highly applicable to financial and insurance scenarios such as claims anomalies, fraud events, rare complaint categories, and special customer behaviors, reducing operating costs and improving the efficiency of intelligent risk control, claims review, and customer service.

[0098] In another embodiment, this application provides a computer device, which may be a server, and its internal structure diagram may be as follows. Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it can implement the methods described in any of the above embodiments.

[0099] In one embodiment, this application also provides a computer device, which can be a client, and its internal structure diagram can be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it can implement the methods described in any of the above embodiments.

[0100] Any of the computer devices described in the embodiments of this application exist in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0101] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, etc.

[0102] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes: audio and video players, handheld game consoles, e-books, as well as smart toys, wearable devices, and portable car navigation devices.

[0103] (4) Server: A device that provides computing services. The components of a server include a processor, hard disk, memory, system bus, etc. Servers are similar to general computer architectures, but because they need to provide highly reliable services, they have higher requirements in terms of processing power, stability, reliability, security, scalability, and manageability.

[0104] (5) Other electronic devices with data interaction functions.

[0105] Additionally, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which are used to perform the following steps: For each real sample in the real sample set of the current initial model, based on the information type of each sample content in the real sample, the content transformation process is performed on each sample content to obtain the expanded sample; The augmented sample is converted into an augmented sample embedding vector, and the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample is calculated. Based on the structured data in the expanded sample, the structural consistency score of the expanded sample is determined; Based on the output embedding representation of the expanded sample in multiple models, the semantic stability of the expanded sample is determined, wherein the multiple models include the current initial model, the historical models corresponding to the previous multiple iterations of the current initial model, and the teacher model; The credibility of the augmented samples is determined based on the semantic similarity, the structural consistency score, the semantic stability, and the confidence of the output of the current initial model. If the credibility is greater than or equal to a predetermined credibility threshold, the expanded sample is determined to be a valid sample, and the expanded sample is used to train the downstream specific business model of the current initial model.

[0106] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0107] The technical solution of this application has been described in detail above with reference to the accompanying drawings. Through the technical solution of this application, when facing rare categories of business scenarios such as rare claims events, abnormal transaction behaviors, special customer complaint types, and rare fraud cases, the purpose of efficiently and automatically expanding the training data of rare categories can be achieved without manual annotation. This overcomes the shortcomings of related technologies, such as limited variation of data augmentation methods, uncontrollable quality of generated samples, and reliance on manual annotation. It improves the efficiency of sample data expansion, can construct training data for rare categories of business in a targeted manner, and ultimately provides more diverse and reliable training samples for downstream business models, which helps to improve the accuracy of the output results of downstream business models.

[0108] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0109] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. The singular forms “a,” “the,” and “the” used in the embodiments of this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0111] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units.

[0112] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0113] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A sample augmentation method, characterized in that, include: For each real sample in the real sample set of the current initial model, based on the information type of each sample content in the real sample, the content transformation process is performed on each sample content to obtain the expanded sample; The augmented sample is converted into an augmented sample embedding vector, and the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample is calculated. Based on the structured data in the expanded sample, the structural consistency score of the expanded sample is determined; Based on the output embedding representation of the expanded sample in multiple models, the semantic stability of the expanded sample is determined, wherein the multiple models include the current initial model, the historical models corresponding to the previous multiple iterations of the current initial model, and the teacher model; The credibility of the augmented samples is determined based on the semantic similarity, the structural consistency score, the semantic stability, and the confidence of the output of the current initial model. If the credibility is greater than or equal to a predetermined credibility threshold, the expanded sample is determined to be a valid sample, and the expanded sample is used to train the downstream specific business model of the current initial model.

2. The method according to claim 1, characterized in that, The information types of the sample content include text, images, and structured data. Therefore, based on the information type of each item in the real sample, the process of performing content transformation on each item of sample content to obtain expanded samples includes: For any of the sample contents of the real sample, If the information type of the sample content is text, the sample content is processed by one or more of the following methods through a large language model: synonym replacement processing, template variation processing, and contextual expansion processing, to obtain the transformed text, which is used as the sample content of the expanded sample. If the information type of the sample content is an image, a new image is generated based on the sample content and a predetermined conditional diffusion model, and used as the sample content of the expanded sample; or, the sample content is subjected to style transformation processing and / or viewpoint transformation processing to obtain a transformed image, which is used as the sample content of the expanded sample. If the information type of the sample content is structured data, a new structured data with the same format and the same field value relationships is generated based on the sample content and a predetermined condition generation model, and this new structured data is used as the sample content of the expanded sample.

3. The method according to claim 1, characterized in that, The step of converting the augmented sample into an augmented sample embedding vector includes: The text, images, and structured data in the augmented sample are encoded separately, and the encoding results are projected into a unified semantic space to obtain the text vector, image vector, and structured data vector of the augmented sample. The text vector, image vector, and structured data vector of the augmented sample are fused using an attention fusion network to obtain the augmented sample embedding vector corresponding to the augmented sample; or, The text vector, image vector, and structured data vector of the augmented sample are concatenated, and the concatenation result is subjected to nonlinear transformation and feature fusion processing through a multilayer perceptron to obtain the augmented sample embedding vector corresponding to the augmented sample.

4. The method according to claim 3, characterized in that, Before performing content transformation processing on each item of sample content based on the information type of each item of sample content in the real sample to obtain the expanded sample, the method further includes: The text, images, and structured data in the real sample are encoded respectively, and the encoding results are projected into the unified semantic space to obtain the text vector, image vector, and structured data vector of the real sample; The text vector, image vector, and structured data vector of the real sample are fused using an attention fusion network to determine the real data embedding index corresponding to the real sample; or, The text vector, image vector, and structured data vector of the real sample are concatenated, and the concatenation result is subjected to nonlinear transformation and feature fusion processing through a multilayer perceptron to determine the real data embedding index corresponding to the real sample.

5. The method according to any one of claims 1 to 4, characterized in that, The step of calculating the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample includes: Cluster the vector sets corresponding to the real sample set to obtain multiple vector clusters; Calculate the centroid of each vector cluster, and based on the centroid and the augmented sample embedding vector, determine the first semantic similarity between the augmented sample and each vector cluster; For the vector cluster corresponding to the highest first semantic similarity, the second semantic similarity between the expanded sample and each real sample in the vector cluster is determined based on the vector of each real data embedding index in the vector cluster and the embedded vector of the expanded sample. The highest second semantic similarity is determined as the semantic similarity between the augmented sample embedding vector and the vector in the real data embedding index corresponding to the real sample.

6. The method according to claim 5, characterized in that, The determination of the structural consistency score of the expanded sample based on the structured data in the expanded sample includes: The system checks whether each field in the structured data meets predetermined field validity conditions, wherein the field validity conditions include: the field is complete, the field type of the field is consistent with the required field type of the field at the location of the field, and the value of the field is within the predetermined field range corresponding to the location of the field. Based on the total number of fields in the structured data and the number of valid fields in the structured data that meet the valid conditions of the fields, the ratio of the number of valid fields to the total number of fields is determined as the structural consistency score of the augmented sample.

7. The method according to claim 6, characterized in that, The determination of the semantic stability of the expanded samples based on their output embedding representations in multiple models includes: Based on the output embedding representation of the augmented sample in the multiple models, calculate the cosine similarity of the output embedding representation of the augmented sample for every two models in the multiple models; Based on the cosine similarity and the total number of the multiple models, the semantic stability of the augmented samples is determined, wherein, ， This refers to the expanded sample. This indicates the semantic stability of the augmented samples. For the i-th model among the multiple models, For the j-th model among the multiple models, This indicates cosine similarity, where N is the total number of the multiple models.

8. The method according to claim 6, characterized in that, The process of determining the credibility of the augmented samples based on the semantic similarity, structural consistency score, semantic stability, and the confidence score of the output of the current initial model includes: The confidence level of the expanded sample is obtained by weighted summation of the semantic similarity, the structural consistency score, the semantic stability, and the confidence level of the output of the current initial model.

9. A computer device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, the instructions being configured to cause the processor to perform the method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The device stores computer-executable instructions configured to perform the method as described in any one of claims 1 to 8.

Images