Material intelligent analysis method, device, equipment, medium and program product

Abstract

The application provides a material intelligent analysis method, device, equipment, medium and program product, and relates to the fields of financial technology or intelligent document processing technology. Feature information of a material to be processed is extracted first, the complexity of analysis of the material is evaluated in combination with the features, a corresponding analysis strategy is adaptively matched according to the complexity, and the material analysis output result is completed accordingly. The method can improve the analysis efficiency and stability, reduce the waste of computing power and processing delay, and enhance the automatic processing capability and result reliability in the high-concurrency and complex material scene by implementing hierarchical processing on materials of different types, formats and qualities.

Description

Technical Field

[0001] This application relates to the fields of financial technology or intelligent document processing technology, and in particular to a method, apparatus, device, medium and program product for intelligent material analysis. Background Technology

[0002] In bank credit approval, existing materials are usually processed through manual review, automatic entry by rule engines, or information extraction and analysis of images, scanned documents and PDF materials using models.

[0003] However, when faced with scenarios involving diverse material types, significant format variations, and unstructured content, the aforementioned methods often struggle to balance processing efficiency and parsing accuracy. Manual processing is time-consuming and inconsistent, rule-based solutions lack adaptability, and single-model solutions are prone to issues such as high computational consumption, large response delays, or insufficient stability in complex semantic parsing. Summary of the Invention

[0004] This application provides a method, apparatus, equipment, medium, and program product for intelligent material analysis, which addresses the technical problems of long processing cycles and limited consistency of manual methods, insufficient adaptability of rule-based schemes to changes, and high computational consumption, large response delays, or insufficient stability of complex semantic analysis by single-model schemes.

[0005] Firstly, this application provides a method for intelligent material analysis, the method comprising:

[0006] Obtain the material characteristic data corresponding to the target material;

[0007] The analytical complexity corresponding to the target material is determined based on the material characteristic data.

[0008] Determine the target processing strategy corresponding to the target material based on the analytical complexity;

[0009] The target material is analyzed based on the target processing strategy to obtain the analysis results.

[0010] Secondly, this application provides a smart material analysis device, comprising:

[0011] The acquisition module is used to acquire the material characteristic data corresponding to the target material;

[0012] The determination module is used to determine the analytical complexity corresponding to the target material based on the material feature data;

[0013] The determination module is used to determine the target processing strategy corresponding to the target material based on the analytical complexity.

[0014] The processing module is used to perform analytical processing on the target material based on the target processing strategy to obtain the analytical results.

[0015] Thirdly, embodiments of this application provide a smart material analysis device, including: a memory and a processor;

[0016] The memory stores computer-executed instructions;

[0017] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0018] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0019] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0020] The intelligent material analysis method provided in this application can effectively distinguish the processing difficulty of different materials by acquiring the material feature data corresponding to the target material and determining the analysis complexity of the target material based on the material feature data. By determining the target processing strategy corresponding to the target material based on the analysis complexity and performing analysis processing on the target material based on the target processing strategy, the analysis process can match the actual complexity of the material, thereby improving the processing efficiency and accuracy of material analysis and improving the rationality of resource utilization. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0022] Figure 1 Flowchart of the intelligent material analysis method provided for this application Figure 1 ;

[0023] Figure 2 Flowchart of the intelligent material analysis method provided for this application Figure 2 ;

[0024] Figure 3 A schematic diagram of the intelligent material analysis device provided in this application;

[0025] Figure 4 A schematic diagram of the intelligent material analysis device provided in this application.

[0026] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0027] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0028] It should be noted that the intelligent material analysis method, apparatus, equipment, media, and program products provided in this application can be used in the fields of financial technology or intelligent document processing technology, or in any field other than financial technology or intelligent document processing technology. This application does not limit the application fields of the intelligent material analysis method, apparatus, equipment, media, and program products.

[0029] Intelligent document processing in bank loan approval is mainly applied to business processes such as pre-loan investigation, loan review, and post-loan management. In this scenario, the business system typically receives documents such as identity certificates, business qualifications, financial statements, and contracts submitted by customers, and completes the approval process by combining document access, information extraction, review and analysis, and result feedback.

[0030] Current credit document processing typically involves manual review, automatic data entry via rule engines, or the use of a single model to identify and analyze images, scanned documents, and PDFs. The basic approach involves receiving the target materials, then manually reviewing and entering information item by item, or extracting fixed fields according to preset rules, or directly calling a unified model to parse text, tables, and semantic content.

[0031] However, credit materials vary significantly in type, format, clarity, and business attributes, making it difficult to adapt a single processing method to all scenarios. While manual methods possess some judgment capability, they suffer from long processing cycles and insufficient standardization. Rule-based solutions rely on predefined templates and fixed logic, making them less adaptable to material variations, field changes, or unstructured content. Single-model solutions tend to waste computational resources on simple tasks and face issues such as response delays or insufficient parsing stability on complex tasks.

[0032] To address the aforementioned issues, the intelligent material analysis method provided in this application first acquires the material feature data corresponding to the target material, then determines the analysis complexity of the target material based on the material feature data, and determines the target processing strategy corresponding to the target material based on the analysis complexity. Finally, the target material is analyzed and processed based on the target processing strategy to obtain the analysis result. This method takes into account processing efficiency, analysis accuracy, and rational resource utilization during the material analysis process.

[0033] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0034] Figure 1 Flowchart of the intelligent material analysis method provided for this application Figure 1 In this embodiment, the executing entity is, for example, a material intelligent analysis system. Figure 1 As shown, the method includes:

[0035] S101: Obtain the material characteristic data corresponding to the target material.

[0036] Among them, material characteristic data is a set of input information for the target material.

[0037] The target materials can be identity documents, business licenses, financial statements, credit authorization letters, purchase and sale contracts, guarantee documents, bank statements, and other paper or electronic materials submitted by the client during the pre-loan investigation, loan review, or post-loan management stages.

[0038] Specifically, when acquiring material feature data corresponding to the target material, the system can receive the target material file and its associated information, and form a feature data object corresponding to the target material. The material feature data can include, but is not limited to, one or more pieces of information that can characterize the properties of the target material; the specific content can be configured according to the actual application scenario. The acquired information is then uniformly written into the feature data object, forming a structured feature record that corresponds one-to-one with the target material, and this structured feature record is output to the complexity evaluation module.

[0039] S102: Determine the analytical complexity of the target material based on the material characteristic data.

[0040] Among them, analytical complexity is used to represent the evaluation result of the difficulty of analyzing the target material.

[0041] Specifically, the analytical difficulty of the target material is assessed based on material characteristic data to obtain the corresponding analytical complexity. This analytical complexity can be determined using an appropriate evaluation mechanism, such as rule-based, model-based, or other complexity assessment methods. Furthermore, in some implementations, the analytical complexity result can be adjusted based on historical processing data or current application requirements to make the obtained result more consistent with actual analytical needs.

[0042] S103: Determine the target processing strategy corresponding to the target material based on the analytical complexity.

[0043] The target processing strategy is used to indicate the analytical method used for the target material.

[0044] Specifically, the target processing strategy reads the parsing complexity corresponding to the target material and determines the corresponding processing strategy identifier based on the parsing complexity. The target processing strategy can be different parsing methods, different processing paths, or different resource call schemes configured for different parsing complexities.

[0045] S104: Analyze the target material based on the target processing strategy to obtain the analysis results.

[0046] The parsing results can be structured data, visualization reports, or a combination of both.

[0047] Structured data can include field names, field values, page number positions, coordinate regions, confidence levels, source fragments, and validation status.

[0048] Visualized reports can include a comparison interface between the original text and the extracted results, highlighted areas of doubt, missing field prompts, and review suggestions.

[0049] Specifically, based on the target processing strategy, the corresponding parsing capabilities are invoked to process the target material, completing content recognition, information extraction, and result generation. After parsing, post-processing can be performed on the parsing results, and the parsing results can be output or sent back after meeting the requirements. By executing a differentiated parsing process according to the target processing strategy, the generation path of the parsing results is consistent with the complexity of the target material, which not only achieves hierarchical processing of different target materials, but also matches the parsing resource investment with the actual difficulty.

[0050] The intelligent material analysis method provided in this embodiment obtains material feature data corresponding to the target material, determines the analysis complexity of the target material based on the material feature data, determines the target processing strategy corresponding to the target material based on the analysis complexity, and performs analysis processing on the target material based on the target processing strategy to obtain the analysis result. In this application, by obtaining material feature data corresponding to the target material and determining the analysis complexity accordingly, and then selecting an appropriate processing strategy and executing the analysis processing based on the analysis complexity, a continuous data transmission relationship is established between material access, complexity assessment, strategy decision-making, and analysis execution, thereby balancing processing efficiency, analysis accuracy, and rational resource utilization in the material analysis process.

[0051] Figure 2 Flowchart of the intelligent material analysis method provided for this application Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the examples, the intelligent material analysis method is described in detail, which includes:

[0052] S201: Obtain the material characteristic data corresponding to the target material.

[0053] Optionally, the material characteristic data may include at least one of the following: material type information, material format information, material quality information, historical analysis information, and business attribute information.

[0054] Specifically, upon receiving the target materials, the aforementioned feature information is extracted from the target material content, file metadata, and business system tags, and then uniformly packaged into material feature data. For material type information, identification can be completed based on a pre-set category dictionary; for material format information, it can be determined based on file extension, container structure, or image encoding method; for material quality information, corresponding values ​​can be obtained by combining image clarity detection, missing area detection, and character readability detection; for historical parsing information, parsing results of similar target materials can be retrieved from the historical task database and statistical values ​​can be generated; for business attribute information, business scenario identifiers can be read from approval documents, customer tags, or product rules.

[0055] The method of constructing material feature data enables the simultaneous acquisition of input information from three levels: the characteristics of the target material itself, the characteristics of its historical analysis, and the characteristics of the business scenario. Based on this feature data, complexity scoring and level classification can be further completed, and matching analysis resources and processing methods can be selected accordingly, thereby making the target material analysis results more in line with the requirements of actual approval scenarios.

[0056] S202: Extract multiple feature parameters from material feature data.

[0057] Specifically, after receiving the material feature data of the target material, the data is first parsed to extract multiple feature parameters, which are then mapped to calculable numerical features. For material type information, different types such as identity documents, business qualifications, financial statements, and contracts can be assigned corresponding type parameter values. For material format information, format parameter values ​​can be set according to image, scanned document, PDF (Portable Document Format), or mixed format. For material quality information, quality parameter values ​​can be generated by combining clarity, tilt, occlusion level, resolution, and noise level. For historical parsing information, historical parameter values ​​can be generated by combining historical recognition accuracy, field missing rate, and number of manual reviews. For business attribute information, business parameter values ​​can be generated by combining the business sensitivity of the target material in pre-loan investigation, loan review, or post-loan management.

[0058] S203: Calculate the complexity score based on the weights corresponding to each feature parameter.

[0059] Specifically, after obtaining each feature parameter, the system performs a weighted calculation on each parameter according to pre-configured weights to generate a complexity score. The weights can be stored in a parameter configuration table and updated based on business scenarios, material categories, or historical statistical results. The complexity score can be generated using a weighted summation method, where each feature parameter value is multiplied by its corresponding weight and then summed to obtain the score result. Alternatively, it can be normalized before combining the results to ensure that feature parameters of different dimensions are within the same scoring scale.

[0060] S204: Determine the analytical complexity level of the target material based on the complexity score.

[0061] Specifically, the calculated complexity score is then matched with a preset threshold range to determine the analytical complexity level of the target material.

[0062] Specifically, by quantifying and representing multiple feature parameters in the material characteristic data and combining them with weights to comprehensively evaluate the analytical difficulty, the complexity score can reflect the overall analytical burden of the target material. Based on this, a hierarchical analytical complexity level is output. This level result allows the target material to be assigned to the corresponding processing channel under different analytical difficulties. The analytical difficulty of the target material no longer depends solely on a single feature judgment, but rather on a unified score formed by weighting multiple features and mapping it to a level result, making the expression of analytical complexity more stable.

[0063] Optionally, the parsing complexity levels include: low complexity level, medium complexity level, and high complexity level.

[0064] Specifically, after obtaining the material characteristic data of the target material, a comprehensive evaluation can be conducted on the target material's layout regularity, character clarity, field stability, degree of text-image integration, and historical hit rate. The comprehensive evaluation results can then be mapped to low, medium, or high complexity levels. For low complexity levels, the target material can be directly fed into the lightweight parsing model channel, and the structured recognition module will output the field results. For medium complexity levels, the lightweight parsing model can first complete the layout positioning and initial field extraction, and then the initial extraction results can be input into the large language model for semantic correction. For high complexity levels, the large language model can be directly called to jointly parse the full text, tables, and contextual information, and output the complete parsing results.

[0065] The analysis complexity level serves as the grading basis for target material analysis, mapping different target material characteristics to different processing intensities, and enabling the matching of corresponding analysis capabilities based on the difficulty of the target material. Since the low, medium, and high levels correspond to different analysis resource allocation methods, simple target materials can be processed with lighter processing, medium-sized target materials can be processed collaboratively, and complex target materials can be processed deeply.

[0066] Using this analytical complexity level can establish a correspondence with the actual analytical difficulty of the target material, thereby obtaining more stable analytical results under different target material scenarios and keeping the allocation of computing resources consistent with the task difficulty.

[0067] S205: Obtain the business level corresponding to the target material.

[0068] S206: Adjust the selection criteria for processing strategies based on business level.

[0069] Specifically, during runtime, the business level of the target material is first obtained, and then this business level is written into the strategy decision unit. The strategy decision unit recalculates the processing strategy selection conditions based on this level, and inputs the revised conditions into the strategy matching logic after parsing complexity assessment, thereby outputting a target processing strategy adapted to the current business level. Since the same target material triggers different processing strategy selection conditions under different business levels, parsing resources can be allocated to parsing paths that better meet business requirements, and higher-level business materials can receive stricter strategy control.

[0070] The business attributes of the target materials can participate in the dynamic determination process of the target processing strategy. The selection conditions for the processing strategy are no longer fixed, but are adjusted according to the business level. Therefore, the allocation of parsing resources can be kept consistent with the importance of the business, and the determination result of the target processing strategy can better meet the risk control requirements in the credit approval scenario.

[0071] S207: Determine the target processing strategy corresponding to the target material based on the analytical complexity.

[0072] Step S207 is similar to step S103, and will not be described again here.

[0073] S208: Divide the analysis task corresponding to the target material into multiple sub-tasks.

[0074] S209: Distribute multiple subtasks to multiple parsing models for parallel execution.

[0075] S210: Merge the processing results of multiple subtasks to obtain the parsed result.

[0076] Specifically, full-text recognition, field extraction, table parsing, key information verification, or semantic relationship analysis of the target material can be constructed into different sub-tasks, and assigned to the corresponding parsing models for parallel processing according to the sub-task type. Each parsing model can output the processing results corresponding to each sub-task based on its own inputs and outputs, and then perform alignment, deduplication, conflict resolution, and structured integration on the processing results to form a unified parsing result. If the parsing task involves different content areas such as text pages, table pages, and stamp pages, it can also be divided into sub-tasks according to the region, and then executed in parallel by the text parsing model, table parsing model, and layout analysis model respectively, to improve the adaptability of the processing.

[0077] Based on the aforementioned parallel splitting and fusion mechanism, the parsing load of the target material can be distributed among multiple models. No single model needs to independently handle all the parsing content, and different types of processing results can form a unified output during the fusion process, ensuring both structural consistency and content integrity in the parsing results. Thus, in processing bank loan materials, when faced with target materials containing multiple pages of text, tables, and mixed formats, the computational burden can be distributed across multiple parsing models, shortening the overall parsing latency and making the final parsing results more suitable for direct use by subsequent approval systems.

[0078] The intelligent material analysis method provided in this application, in this embodiment, Figure 2 Based on the embodiments, the target processing strategy is described in detail, and the method includes:

[0079] After determining the parsing complexity based on the material's characteristic data, a target processing strategy matching that complexity is invoked. When the target material is in a low-complexity scenario, text and table regions from images, scanned documents, or PDFs are directly input into a lightweight parsing model, which performs field extraction, layout recognition, and basic structuring processing, outputting corresponding field information, structural information, or preliminary analysis information. When the target material is in a medium-complexity scenario, the lightweight parsing model first performs pre-parsing to obtain layout frames, text blocks, table cells, and candidate fields. This result, along with the target material's context, is then input into a large language model, which performs supplementary analysis on field meanings, contextual relationships, and cross-paragraph information, thus forming a collaborative parsing result. When the target material is in a high-complexity scenario, the original target material and its context information are directly input into the large language model to perform comprehensive parsing of unstructured content, implicit relationships, and complex semantics, outputting the final parsing result.

[0080] The target processing strategy assigns parsing tasks of different parsing complexities to corresponding models, or has two types of models work together to complete the parsing, so that the parsing results of the target material can simultaneously contain structured fields and semantic analysis content, and maintain the consistency between the output results and the content of the target material.

[0081] Low-complexity target materials can be processed quickly by lightweight parsing models, medium-complexity target materials can have their structural and semantic information supplemented through collaborative parsing, and high-complexity target materials can be parsed deeply by large language models, thus making the final parsing results more complete and the resource allocation method that matches the complexity of the target materials more reasonable.

[0082] Target processing strategy 1: Use a lightweight parsing model to perform parsing processing.

[0083] The lightweight parsing model can be constructed using a convolutional neural network OCR (Optical Character Recognition) model, a table detection network, or a layout analysis network. It has a small number of model parameters and is suitable for running quickly under high concurrency conditions.

[0084] Target processing strategy two: Use a lightweight parsing model and a large language model to perform parsing processing in a coordinated manner.

[0085] Among them, the large language model can adopt a generative model with text understanding, information induction and relational reasoning capabilities. It can pass intermediate results with the lightweight parsing model through an interface. When necessary, business rules can also constrain and verify the output of the large language model.

[0086] Optionally, a lightweight analytical model can be used to perform preliminary analysis of the target material;

[0087] Obtain the confidence level corresponding to the output of the lightweight analytical model;

[0088] Output the parsing results if the confidence level meets the preset conditions;

[0089] If the confidence level does not meet the preset conditions, the large language model is invoked for enhanced parsing.

[0090] Specifically, the lightweight parsing model can employ a text recognition model, field extraction model, or multimodal parsing model with fewer parameters to perform preliminary parsing of the text, tables, and layout information in the target material, and output candidate fields and their corresponding probability values. Upon receiving this output, the probability values ​​of the candidate fields are aggregated and calculated to form an overall confidence score, or a minimum confidence score is set for each key field, and the minimum value is used as the overall judgment criterion. The large language model can be deployed on the server side, receiving the candidate results output by the lightweight parsing model, the original target material text, or a partial image of the target material through an interface. It then performs secondary inference based on the contextual semantics, thereby enhancing the parsing of missing, ambiguous, or low-confidence fields, and outputting the corrected parsing results.

[0091] For different types of target materials, such as identity documents, contracts, or financial statements, the lightweight parsing model first performs an initial screening. When the target material is clear, well-formatted, and has clearly defined fields, its confidence level is usually higher than the preset conditions, and the system directly outputs the parsing results. When the target material has occlusions, tilting, compression distortion, or non-standard field expressions, the confidence level of the lightweight parsing model decreases, triggering the large language model to complete, correct, and verify the parsed content. Through this collaborative mechanism, while ensuring the processing efficiency of routine target materials, enhanced parsing support can be provided for complex or uncertain target materials, thereby maintaining consistency in the accuracy and stability of the parsing results.

[0092] Target processing strategy three: Use a large language model to perform parsing processing.

[0093] Optionally, identify key information entities in the target material;

[0094] Analyze the relationships between key information entities;

[0095] Structured parsing results and attribute analysis results are generated based on the relationships.

[0096] Specifically, the large language model receives the text content, table content, or recognized character content corresponding to the target material, and extracts entities from the semantic units to generate a set of candidate entities related to credit business. Subsequently, based on contextual semantics, field positional relationships, and cross-paragraph referential relationships, the large language model infers the associations between candidate entities, determining the correspondence and constraints of each entity in the business chain. Based on these associations, the large language model maps entities, attributes, and their relationships to preset structured field templates, outputting structured parsing results that can be directly written to the business database. Simultaneously, it combines entity integrity, field consistency, and business logic consistency to form attribute analysis results. If the target material contains multiple pages of text or multiple related attachments, the large language model can also perform joint parsing of entities across pages and attachments to maintain the consistency of the structured results.

[0097] The above processing method enables the large language model not only to extract information from the target material, but also to further identify the semantic relationships between key information entities and form a structured output, thus making the parsing results both storable and verifiable. By linking key information entities, relationships, structured parsing results, and attribute analysis results, the completeness of parsing complex target materials and the accuracy of business matching can be improved.

[0098] The attribute analysis results include at least one of the following: protocol attribute analysis results, data attribute analysis results, object attribute analysis results, and data consistency analysis results.

[0099] The agreement attribute analysis results represent the analytical conclusions of the agreement-related attributes in the target material. Agreement-related attributes include at least one of the following: clause content, rights and obligations stipulations, constraints on the signing parties, and time limits. When identifying target materials such as contracts, agreements, or authorization letters, the above content can be extracted based on key information entities and their contextual semantics, and the extraction results can be mapped to agreement attribute analysis results to characterize the agreement elements contained in the target material. The data attribute analysis results represent the analytical conclusions of the data characteristics or data semantics in the target material. Data characteristics may include at least one of the following: amount, quantity, date, proportion, number, and field meaning. They can be parsed in conjunction with table areas, text areas, and entity relationships to generate corresponding data attribute analysis results. The object attribute analysis results represent the analytical conclusions of the object characteristics in the target material. Object characteristics may include at least one of the following: entity name, identity category, affiliation, status attribute, and business role. Object attribute analysis results can be formed based on the identified entities and the relationships between entities. Data consistency analysis results are used to express the analytical conclusion of whether the data in the target material are consistent. It can compare the entity values ​​at different locations within the same target material or the corresponding entity values ​​between different target materials, and output the judgment results of consistency, partial consistency or inconsistency.

[0100] When using a large language model for parsing, the model first identifies key information entities in the target material, then generates structured parsing results based on contextual relationships, and simultaneously generates attribute analysis results. For agreement attribute analysis results, the model can extract contract titles, clause numbers, liability for breach of contract, and conditions for effectiveness from the target material, and combine semantic relationships to form a judgment on the agreement's binding relationships. For data attribute analysis results, the model can classify and interpret numerical fields, form fields, and implicit data semantics in the text of the target material, and output their respective data attribute types. For object attribute analysis results, the model can merge the attributes of entities such as borrowers, guarantors, account holders, or transaction partners involved in the target material to form object status descriptions. For data consistency analysis results, the model can cross-validate information such as the name, document number, and amount of the same entity in identity documents, transaction records, contract texts, and application forms, and output consistency conclusions.

[0101] The above attribute analysis results can be returned as part of the parsing results along with the structured fields. By incorporating the protocol, data, object, and consistency analysis results into the output of the large language model, the model can supplement the judgment information at the attribute level of the target material while completing semantic parsing, thereby forming a more complete parsing conclusion for the target material and improving the semantic coverage and usability of the results for complex target materials.

[0102] The intelligent material analysis method in this application can be applied to the pre-loan review stage of bank lending. Customers submit loan application materials including ID cards, income certificates, financial statements, contracts, etc. These materials need to undergo automated analysis and risk assessment to assist in approval decisions. In a possible example, a customer applying for a personal loan submits an ID card, income certificate, and financial statements. The ID card is a high-resolution image, the income certificate is a PDF containing tables and a handwritten signature, and the financial statements are scanned copies, multi-page, and some content is blurry. The method first obtains the material feature data corresponding to each target material. This material feature data includes at least one of the following: material type information, material format information, material quality information, historical analysis information, and business attribute information. Material type information indicates whether the target material belongs to the category of text, form, invoice, contract, report, or other materials. In this example, it corresponds to an ID card, income certificate, and financial statement, respectively. Material format information indicates the file format, layout features, or data organization form of the target material. In this example, it corresponds to an image, PDF, and scanned copy, and the income certificate includes a form with a handwritten signature, while the financial statement is multi-page. Material quality information indicates clarity, completeness, noise level, missing pages, image tilt, or missing fields. In this example, the ID card has high clarity, the income certificate has medium clarity, and the financial statement has low clarity. The clarity is low and some content is blurry; historical parsing information is used to characterize the success rate, abnormal records, error correction records, or model adaptation of the same or similar materials in previous parsing processes. In this example, the historical parsing success rate of ID cards is 98%, the historical parsing success rate of income certificates is 85%, and the historical parsing success rate of financial statements is 70%; business attribute information is used to characterize the business scenario, business rules, field requirements, or timeliness requirements of the target material. In this example, the business attribute of ID cards is low risk, the business attribute of income certificates is medium risk, and the business attribute of financial statements is high risk. Moreover, the pre-loan review scenario has extremely high requirements for parsing accuracy and response speed.

[0103] After acquiring material feature data, multiple feature parameters are extracted from the data, and a complexity score is calculated based on the weight of each feature parameter. The analytical complexity level for each target material is then determined based on the complexity score. During the calculation, parameters such as material format regularity, field density, image quality, historical anomaly frequency, and business rule complexity are mapped to corresponding values ​​and weighted according to preset weights to obtain a complexity score. This complexity score is then compared with a complexity level classification threshold, and the analytical complexity level matching the target material is output. In this example, the ID card, based on its type (ID card), format (image), high clarity, historical parsing success rate of 98%, and business attribute (low risk), is mapped to low complexity and scores 10 points; the income certificate, based on its type (income certificate), format (PDF), clarity of medium, historical parsing success rate of 85%, and business attribute (medium risk), is mapped to medium complexity and scores 50 points; the financial statement, based on its type (financial statement), format (scanned document), clarity of low, historical parsing success rate of 70%, and business attribute (high risk), combined with multiple pages and partially blurred content, is mapped to high complexity and scores 90 points.

[0104] Before determining the target processing strategy based on parsing complexity, the business level corresponding to the target material is first obtained, and the selection criteria for the processing strategy are adjusted according to the business level. The business level reflects the differences in parsing accuracy, timeliness, structuring degree, or risk control requirements of the business scenario. In this pre-loan review scenario, the ID card corresponds to a low-risk business attribute and can be determined as a lower business level; the income certificate corresponds to a medium-risk business attribute and can be determined as a medium business level; and the financial statement corresponds to a high-risk business attribute and can be determined as a higher business level. When the business level is high, the sensitivity of triggering enhanced parsing or large language model parsing is increased; when the business level is low, the threshold for selecting the processing strategy is relaxed. Combining the selection criteria adjusted by the business level, the target processing strategy corresponding to the target material is determined based on the parsing complexity. The target processing strategy includes at least one of the following: using a lightweight parsing model to perform parsing processing, using a lightweight parsing model and a large language model to perform parsing processing collaboratively, and using a large language model to perform parsing processing. In this example, for low-complexity and low-risk ID cards, the OCR recognition small model is directly called for processing, without the need for a large model. For medium-complexity and medium-risk income certificates, the table detection small model first extracts the table structure, and then the large model performs semantic analysis, including identifying income amount and signature authenticity. For high-complexity and high-risk financial statements, the large model is directly called for in-depth analysis, including extracting the debt-to-equity ratio and judging financial health.

[0105] Before parsing, the parsing task corresponding to the target material is broken down into multiple sub-tasks, which are then assigned to multiple parsing models for parallel execution. The processing results from these sub-tasks are then fused. Sub-tasks include layout recognition, text extraction, field location, entity recognition, relation extraction, attribute judgment, and consistency verification. Different sub-tasks are assigned to lightweight parsing models, large language models, or a combination of both, based on the target processing strategy. In this example, the income certificate parsing task is broken down into two sub-tasks: table structure extraction is performed in parallel by a small table detection model, while semantic analysis is performed in parallel by a large model. When parsing the ID card using a lightweight parsing model, the OCR small model directly recognizes, extracts, and performs structured conversion. When parsing the financial statements using a large language model, the large language model performs deep semantic parsing. During fusion, the parallel output results are aligned, conflict resolved, confidence compared, and results completed to form a unified parsing output.

[0106] In this example, during the small model processing stage, the OCR model recognizes the ID card number and outputs a confidence level of 95%; the table detection model extracts table data from the income certificate, including "Monthly Income: 12,000 Yuan," and outputs a confidence level of 88%. When the target processing strategy is to use a lightweight parsing model and a large language model to perform parsing processing collaboratively, the lightweight parsing model is first used to perform preliminary parsing of the target material to obtain the confidence level corresponding to the output result of the lightweight parsing model, and the confidence level is compared with the preset conditions. In this example, the preset threshold is 90%, and the confidence level of "Monthly Income: 12,000 Yuan" in the income certificate is 88%, which is lower than the preset threshold of 90%. Therefore, the corresponding parsing result is not directly output. Instead, the large language model is called for enhanced parsing, and the large language model supplements and corrects the content that the lightweight parsing model did not accurately identify, that is structurally incomplete, or that has high semantic ambiguity. Based on tabular data, the large language model analyzes the semantic content of the income certificate, judges the authenticity of the signature and the reasonableness of the income, and outputs a confidence level of 92%. A second confirmation is performed on "Monthly Income: 12,000 yuan," ultimately determining the field to be "12,500 yuan," increasing the confidence level to 93%. When the target processing strategy uses the large language model for parsing, the model identifies key information entities in the target material, analyzes the relationships between these entities, and generates structured parsing results and attribute analysis results based on these relationships. In the processing of financial statements, key information entities include monetary information, data item names, and status descriptions. Relationships include monetary correspondence and attribution relationships. The large language model performs deep semantic parsing of the financial statements, extracts a debt-to-equity ratio of 60%, assesses financial risk, and outputs a confidence level of 89%. Attribute analysis results include at least one of the following: agreement attribute analysis results, data attribute analysis results, object attribute analysis results, and data consistency analysis results. In this example, it forms judgments on financial health, income reasonableness, signature authenticity, and related data consistency analysis results.

[0107] The parallel output results are integrated to generate structured data, such as JSON format, and the confidence level of each field is labeled. Simultaneously, the parsing results of the target material are output by combining the results of the parallel subtasks. The structured parsing results include key fields such as ID card number, income amount, and debt-to-asset ratio. The attribute analysis results include relevant information such as the authenticity of the income certificate signature and the customer's financial risk level (moderate). The key fields such as ID card number, income amount, and debt-to-asset ratio are stored in the credit system database, and a visual report containing risk assessment conclusions is generated. The report includes "Customer Financial Risk Level: Medium" and "Income Certificate Signature Authentic." In the quality feedback processing, low-confidence records of the "Monthly Income" field in the income certificate are fed back to the quality monitoring module, triggering model optimization. This includes retraining the OCR model to improve the ability to recognize fuzzy text; at the same time, the parsing results of the financial statements are archived for subsequent manual review.

[0108] Figure 3 A schematic diagram of the intelligent material analysis device provided in this application is shown below. Figure 3 As shown, the intelligent material analysis device 300 provided in this embodiment includes:

[0109] The acquisition module 301 is used to acquire the material characteristic data corresponding to the target material;

[0110] Module 302 is used to determine the analytical complexity of the target material based on the material characteristic data.

[0111] The determination module 302 is also used to determine the target processing strategy corresponding to the target material based on the analytical complexity;

[0112] The processing module 303 is used to perform analytical processing on the target material based on the target processing strategy to obtain the analytical results.

[0113] In one possible implementation, the determining module 302 is further configured to include at least one of the following material characteristic data: material type information, material format information, material quality information, historical parsing information, and business attribute information.

[0114] In one possible implementation, the intelligent material analysis device 300 includes: an extraction module 304 and a calculation module 305;

[0115] Extraction module 304 is used to extract multiple feature parameters from material feature data;

[0116] Calculation module 305 is used to calculate the complexity score based on the weights corresponding to each feature parameter;

[0117] The determination module 302 is also used to determine the analytical complexity level of the target material based on the complexity score.

[0118] In one possible implementation, the determining module 302 is further configured to resolve complexity levels including: low complexity level, medium complexity level, and high complexity level.

[0119] In one possible implementation, the processing module 303 is further configured to perform parsing processing using a lightweight parsing model;

[0120] And / or,

[0121] Processing module 303 is also used to perform parsing processing in collaboration with a lightweight parsing model and a large language model;

[0122] And / or,

[0123] Processing module 303 is also used to perform parsing processing using a large language model.

[0124] In one possible implementation, the processing module 303 is further configured to perform preliminary analysis of the target material using a lightweight analytical model;

[0125] The acquisition module 301 is also used to acquire the confidence level corresponding to the output result of the lightweight analytical model;

[0126] The processing module 303 is also used to output the parsing result when the confidence level meets the preset conditions;

[0127] The processing module 303 is also used to call the large language model for enhanced parsing when the confidence level does not meet the preset conditions.

[0128] In one possible implementation, the intelligent material analysis device 300 includes: an identification module 306, an analysis module 307, and a generation module 308;

[0129] The identification module 306 is used to identify key information entities in the target material;

[0130] Analysis module 307 is used to analyze the relationships between key information entities;

[0131] The generation module 308 is used to generate structured parsing results and attribute analysis results based on the association relationship.

[0132] In one possible implementation, the determining module 302 is further configured to include at least one of the following attribute analysis results: protocol attribute analysis results, data attribute analysis results, object attribute analysis results, and data consistency analysis results.

[0133] In one possible implementation, the intelligent material analysis device 300 includes: an adjustment module 309;

[0134] The acquisition module 301 is also used to acquire the business level corresponding to the target material;

[0135] Adjustment module 309 is used to adjust the selection conditions of processing strategy according to business level.

[0136] In one possible implementation, the intelligent material analysis device 300 includes: a splitting module 309 and an allocation module 310;

[0137] The splitting module 309 is used to split the analysis task corresponding to the target material into multiple sub-tasks;

[0138] The allocation module 310 is used to allocate multiple subtasks to multiple parsing models for parallel execution;

[0139] The processing module 303 is also used to fuse the processing results of multiple subtasks to obtain the parsing result.

[0140] The intelligent material analysis device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0141] Figure 4 A schematic diagram of the intelligent material analysis device provided in this application. Figure 4 As shown, the electronic device of this embodiment may include: at least one processor 401; and a memory 402 communicatively connected to the at least one processor; wherein the memory 402 stores instructions that can be executed by the at least one processor 401, and the instructions are executed by the at least one processor 401 to cause the electronic device to perform the method as described in any of the above embodiments.

[0142] Optionally, the memory 402 can be either standalone or integrated with the processor 401. When the memory 402 is set up independently, the device also includes a bus for connecting the memory 402 and the processor 401.

[0143] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0144] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed by a processor, the methods provided in any of the foregoing embodiments can be implemented.

[0145] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the method provided in any of the foregoing embodiments.

[0146] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0147] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0148] It should be understood that the above-described device embodiments are merely illustrative, and the device of this application can also be implemented in other ways. For example, the division of units / modules in the above embodiments is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units, modules, or components may be combined, or integrated into another system, or some features may be ignored or not executed.

[0149] Furthermore, unless otherwise specified, the functional units / modules in the various embodiments of this application can be integrated into one unit / module, or each unit / module can exist physically separately, or two or more units / modules can be integrated together. The integrated units / modules described above can be implemented in hardware or as software program modules.

[0150] Unless otherwise specified, the processor can be any suitable hardware processor, such as a CPU, GPU, FPGA, DSP, and ASIC. Unless otherwise specified, the storage unit can be any suitable magnetic or magneto-optical storage medium, such as resistive random access memory (RRAM), dynamic random access memory (DRAM), static random access memory (SRAM), enhanced dynamic random access memory (EDRAM), high-bandwidth memory (HBM), hybrid memory cube (HMC), etc.

[0151] If the integrated unit / module is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0152] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0153] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0154] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for intelligent material analysis, characterized in that, include: Obtain the material characteristic data corresponding to the target material; The analytical complexity corresponding to the target material is determined based on the material characteristic data. Determine the target processing strategy corresponding to the target material based on the analytical complexity; The target material is analyzed based on the target processing strategy to obtain the analysis results.

2. The method according to claim 1, characterized in that, The material characteristic data includes at least one of the following: material type information, material format information, material quality information, historical analysis information, and business attribute information.

3. The method according to claim 1, characterized in that, The step of determining the analytical complexity of the target material based on the material feature data includes: Extract multiple feature parameters from the material feature data; Calculate the complexity score based on the weights corresponding to each of the aforementioned feature parameters; The analytical complexity level of the target material is determined based on the complexity score.

4. The method according to claim 3, characterized in that, The parsing complexity levels include: low complexity level, medium complexity level, and high complexity level.

5. The method according to claim 1, characterized in that, The target processing strategy includes: A lightweight parsing model is used to perform parsing processing; And / or, A lightweight parsing model and a large language model are used in conjunction to perform parsing processing; And / or, A large language model is used to perform parsing processing.

6. The method according to claim 5, characterized in that, The method of using a lightweight parsing model and a large language model to perform parsing processing includes: The lightweight analytical model is used to perform preliminary analysis of the target material; Obtain the confidence level corresponding to the output result of the lightweight analytical model; The parsing result is output when the confidence level meets the preset conditions. If the confidence level does not meet the preset conditions, the large language model is invoked for enhanced parsing.

7. The method according to claim 5, characterized in that, The parsing process using a large language model includes: Identify key information entities in the target material; Analyze the relationships between the key information entities; Structured parsing results and attribute analysis results are generated based on the aforementioned relationships.

8. The method according to claim 7, characterized in that, The attribute analysis results include at least one of the following: protocol attribute analysis results, data attribute analysis results, object attribute analysis results, and data consistency analysis results.

9. The method according to claim 1, characterized in that, Before determining the target processing strategy corresponding to the target material based on the analytical complexity, the method further includes: Obtain the business level corresponding to the target material; The processing strategy selection criteria are adjusted based on the business level.

10. The method according to claim 1, characterized in that, The step of performing analytical processing on the target material based on the target processing strategy to obtain analytical results includes: The analysis task corresponding to the target material is divided into multiple sub-tasks; The subtasks are assigned to multiple parsing models for parallel execution. The processing results corresponding to multiple subtasks are fused to obtain the parsing result.

11. A smart material analysis device, characterized in that, include: The acquisition module is used to acquire the material characteristic data corresponding to the target material; The determination module is used to determine the analytical complexity corresponding to the target material based on the material feature data; The determination module is used to determine the target processing strategy corresponding to the target material based on the analytical complexity. The processing module is used to perform analytical processing on the target material based on the target processing strategy to obtain the analytical results.

12. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-10.

14. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-10.

Images