A method, apparatus, device, and medium for assessing the accuracy of bank reports.

By introducing artificial intelligence models and semantic modeling technology, automated consistency verification of bank reports is achieved, solving the problems of time-consuming, labor-intensive, and prone to misjudgment in traditional manual review, and improving review efficiency and accuracy.

CN120706431BActive Publication Date: 2026-11-17杭州展鸿科技有限公司
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510802473.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-06-16
Publication Date
2026-11-17
Estimated Expiration
2045-06-16

AI Technical Summary

Technical Problem

Traditional bank report review relies on manual review, which consumes a lot of time and manpower, and is subject to the risk of misjudgment and omission, making it difficult to guarantee accuracy.

Method used

By employing artificial intelligence models and semantic modeling techniques, the BERTScore model is used for contextual semantic embedding and vector matching, the PXL-BLEU algorithm is combined for terminology normativity comparison, and a trained risk level assessment model is used for risk conclusion consistency verification, thereby achieving automated consistency verification of semantics, terminology, numerical values, and entities.

Benefits of technology

It has improved the efficiency and accuracy of bank report review, reduced labor costs, decreased the risk of misjudgment and omission, and enhanced the intelligence and transparency of the review process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120706431B_ABST
    Figure CN120706431B_ABST
Patent Text Reader

Abstract

This invention relates to the technical field of financial information assessment. Specifically, it relates to a method, apparatus, device, and medium for assessing the accuracy of bank reports. The method includes using the BERTScore model to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information; calculating the cosine similarity between the embedded vectors of the original material information and the embedded vectors of the due diligence information and performing greedy matching to obtain a semantic consistency score; using the PXL-BLEU algorithm combined with standard terms from a pre-set terminology library to obtain a terminology consistency verification result; inputting the original information into a pre-set indicator calculation formula to generate corresponding reference values; constructing a risk feature vector; inputting the risk feature vector into a trained risk level assessment model; and matching the customer entity identification information with the subject identification information in the original material information to obtain an entity identification consistency verification result. This invention improves the accuracy and consistency of report review.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the technical field of financial information assessment, and in particular to a method, apparatus, device, and medium for assessing the accuracy of bank reports. Background Technology

[0002] Traditionally, to ensure the accuracy of bank reports, manual review is typically employed. Reviewers meticulously examine the client's original documentation and pre-loan due diligence information, using their professional knowledge and experience to compare and assess aspects such as semantics, terminology, numerical accuracy, and risk assessment. Additionally, simple rule engines are sometimes used to assist the review process, checking specific content within the report according to pre-defined rules. For example, they might check the completeness of certain formatted information or whether specific numerical ranges meet requirements.

[0003] However, these existing review methods have obvious shortcomings. Manual review not only consumes a lot of time and manpower, but also, due to the subjective factors and individual differences of the reviewers, it is easy to overlook or misjudge, making it difficult to guarantee the accuracy of the review results. Summary of the Invention

[0004] To improve the accuracy and consistency of report review, this application provides a method, apparatus, equipment, and medium for assessing the accuracy of bank reports.

[0005] The above-mentioned objective of this application is achieved through the following technical solution: A method for assessing the accuracy of bank reports, the method comprising: Obtain original material information from the target bank's customers; Artificial intelligence models are used to obtain customer pre-loan due diligence information corresponding to the original material information; The BERTScore model is used to perform contextual semantic embedding processing on the original material information and the customer pre-loan due diligence information to generate the original material information embedding vector and the due diligence information embedding vector. The cosine similarity of the original material information embedding vector and the due diligence information embedding vector is calculated and greedy matching is performed to obtain a semantic consistency score. If the semantic consistency score reaches the preset credibility condition, then financial terminology information is obtained from the customer's pre-loan due diligence information, and the PXL-BLEU algorithm is used in combination with the standard terms in the preset terminology standard library to perform a normative comparison of the financial terminology information to obtain the terminology consistency verification result. The financial indicator information identified from the terminology consistency verification results is used to extract the original information for calculating the financial indicator information from the original material information. The original information is then input into the preset indicator calculation formula to generate the corresponding reference value. The indicator description value is obtained from the customer pre-loan due diligence information. The reference value and the description value are compared to obtain the numerical consistency verification result. If abnormal financial indicator information is found in the numerical consistency verification result, the risk judgment statement information associated with the abnormal financial indicator information in the customer pre-loan due diligence information is identified. Based on the risk judgment statement information, the corresponding risk feature information is extracted from the original material information. Based on the risk feature information, a risk feature vector is constructed and input into the trained risk level assessment model to obtain the predicted risk level. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain the risk conclusion consistency verification result. Based on the semantic consistency score and the terminology consistency verification result, customer entity identification information is extracted from the customer pre-loan due diligence information. The customer entity identification information is then matched with the subject identification information in the original material information to obtain the entity identification consistency verification result.

[0006] By adopting the above technical solutions, artificial intelligence models and semantic modeling techniques can be applied to the content consistency verification process between original material information and customer pre-loan due diligence information. By introducing the BERTScore model for contextual semantic embedding and vector matching, automatic semantic comparison is achieved. The PXL-BLEU algorithm is used for normative matching of financial terms. At the same time, the financial data is analyzed for numerical consistency by combining preset indicator calculation formulas. The consistency of risk conclusions is verified by using a trained risk level assessment model. On this basis, entity recognition consistency verification is further used to ensure the accuracy of subject information. Finally, the verification results of each dimension are summarized into an accuracy score and correction suggestions, which completely replaces the traditional method of manual item-by-item review and rule engine-assisted judgment, effectively improving review efficiency, reducing labor costs, reducing the risk of misjudgment and omission, and enhancing the accuracy of the bank report review process.

[0007] In a preferred example, this application can be further configured as follows: the use of the BERTScore model to perform contextual semantic embedding processing on the original material information and the customer pre-loan due diligence information, respectively, to generate embedding vectors for the original material information and due diligence information, including: The BERTScore model performs multi-layer encoding on the original material information to obtain an original material information embedding representation sequence containing context-dependent features. Vector aggregation processing is then performed on the original material information embedding representation sequence to generate the original material information embedding vector. The BERTScore model performs multi-layer encoding on the customer's pre-loan due diligence information to obtain a due diligence information embedding representation sequence. The vector aggregation process is then performed on the due diligence information embedding representation sequence to generate the due diligence information embedding vector.

[0008] By adopting the above technical solution, the BERTScore model can be used to perform multi-layer encoding processing on the original material information and the customer's pre-loan due diligence information, extract context-dependent semantic features, and generate an overall text embedding representation through vector aggregation. This effectively captures the deep semantic relationships in the text, avoids semantic misjudgment caused by differences in language expression, ambiguity of word meaning, or fragmented context in manual review, and provides a unified, dimensionally corresponding, and semantically rich vector foundation for subsequent semantic consistency calculation. This improves the automation and accuracy of semantic comparison and enhances the intelligence and consistency of the evaluation process.

[0009] In a preferred embodiment, this application can be further configured as follows: The step of calculating the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and performing a greedy matching to obtain a semantic consistency score includes: The original material information embedding vector is divided into a first sub-vector set, and the due diligence information embedding vector is divided into a second sub-vector set; The first sub-vector set and the second sub-vector set are combined, and the cosine similarity calculation is performed between each pair of sub-vectors to generate a similarity score matrix; Based on the similarity score matrix, the greedy matching is used to select, in each round of matching, sub-vector pairs with high similarity scores from the first sub-vector set that has not yet participated in matching and the second sub-vector set that has not yet participated in matching; The subvector pairs are marked as matched, and the subvector pairs are added to the semantic matching set; After all the sub-vector pairs have been selected, the similarity scores of the sub-vector pairs in the semantic matching set are averaged to generate the semantic consistency score.

[0010] By adopting the above technical solution, the embedded vectors of the original material information and the customer's pre-loan due diligence information can be substructured. By constructing a similarity score matrix of pairwise combinations of subvectors, and combining a greedy matching strategy to dynamically select semantically highly corresponding subvector pairs, fine-grained semantic hierarchical matching can be effectively achieved. This avoids the problem of difficulty in aligning key content due to the large amount of information and wide semantic span in manual review. At the same time, by averaging the similarity scores of the matched subvector pairs, the semantic consistency between the original materials and due diligence information can be quantified, thereby improving the automation accuracy and objectivity of semantic comparison and providing a reliable scoring basis for subsequent accuracy assessment.

[0011] In a preferred embodiment, this application can be further configured as follows: The step of obtaining financial terminology information from the customer's pre-loan due diligence information, and using the PXL-BLEU algorithm combined with standard terms from a pre-defined terminology standardization library to perform a normative comparison of the financial terminology information to obtain a terminology consistency verification result, includes: Based on the aforementioned financial terminology information, a set of terms to be compared is constructed; Using the PXL-BLEU algorithm, the terms to be compared in the term set to be compared are compared with the standard terms in the preset term specification library to calculate the similarity score, and the corresponding term matching score set is obtained. Based on the term matching score set, determine whether each term to be compared has a matching relationship with any standard term in the term standard library with a matching score higher than a preset similarity threshold. If it exists, mark the term to be compared as a standard matching state; otherwise, mark the term to be compared as an unmatched state. The terms to be compared corresponding to the unmatched states are summarized to obtain a set of term replacement candidates; The PXL-BLEU algorithm is used to calculate the similarity score between the term to be compared in the candidate term replacement set and the standard term, and update the matching status of the corresponding term to be compared to obtain the updated matching status. The updated matching status is merged with the standard matching status to obtain the final matching status information; Based on the final matching status information, the term matching accuracy is calculated, and based on the term matching accuracy, the term consistency verification result is generated.

[0012] By adopting the above technical solution, financial terminology information from pre-loan due diligence can be structured and organized into a set of terms to be compared. By introducing the PXL-BLEU algorithm, multiple rounds of similarity scoring are performed on each term to be compared with standard terms in the preset terminology standard library. Combined with the preset similarity threshold, it is determined whether there is a standard matching relationship between the terms. For unmatched terms, alternative terminology candidates are generated and a secondary scoring correction is performed, which effectively improves the accuracy and coverage of terminology matching. At the same time, by calculating the terminology matching accuracy rate in the final matching state, the compliance of terminology expression is quantified, avoiding the problem of incorrect judgment caused by differences in terminology expression and lack of industry experience in manual review, thereby realizing the automation and objectification of the terminology standard comparison process.

[0013] In a preferred embodiment, this application can be further configured as follows: constructing a risk feature vector based on the risk feature information, and inputting the risk feature vector into a trained risk level assessment model to obtain a predicted risk level, including: The trained risk level assessment model models the correlation between risk factors in the risk feature vector based on the attention mechanism, and performs weighted processing in combination with the time evolution trend of the risk factors. After completing the modeling of the risk feature vector, it outputs an overall risk level score. According to the preset risk level classification rules, the overall risk level score is mapped to the predicted risk level.

[0014] By adopting the above technical solution, it is possible to model the risk feature vector extracted from the original material information based on the trained risk level assessment model, dynamically capture the correlation between different risk factors using the attention mechanism, and perform weight adjustment in combination with the time evolution trend of risk factors, thereby constructing a comprehensive risk feature expression that is more in line with the actual business change logic. On this basis, an overall risk level score is generated, and the score value is mapped to a specific predicted risk level according to the preset risk level classification rules, effectively avoiding misjudgment caused by the lack of systematic modeling of complex risk factors in manual assessment.

[0015] In a preferred embodiment, this application can be further configured as follows: matching the customer entity identification information with the subject identification information in the original material information to obtain an entity identification consistency verification result includes; Obtain target entity identification information from the customer entity identification information, and obtain reference entity identification information from the original material information; Using a field-level joint matching strategy, field alignment and similarity calculation are performed on the target entity identifier information and the reference entity identifier information to generate an entity matching score matrix; By performing a threshold judgment operation on the entity matching score matrix, the matching relationship status between the target entity identification information and the reference entity identification information is determined, and the entity recognition consistency verification result is obtained.

[0016] By adopting the above technical solution, entity fields with identifying functions can be extracted from customer pre-loan due diligence information and original material information. A field-level joint matching strategy is then used to standardize, align, and calculate the similarity of target entity identification information and reference entity identification information to construct an entity matching score matrix. Based on this, by setting a threshold to compare the strength of the scores, it is determined whether there is a consistent matching relationship between the two sets of entity identification information. This effectively solves the problem of misjudgment of the subject caused by differences in field naming, inconsistent expression methods, or information omissions in traditional manual identification. It realizes the consistency verification and automated verification of customer subject identity, improves the accuracy and stability of data consistency verification, and provides a structured entity-level verification basis for report credibility assessment.

[0017] In a preferred embodiment, this application may be further configured such that: the method for assessing the accuracy of bank reports further includes; The semantic consistency score, the terminology consistency verification result, the numerical consistency verification result, the risk conclusion consistency verification result, and the entity recognition consistency verification result are summarized to obtain a multidimensional consistency index set; The multidimensional consistency index set is input into the accuracy evaluation model to obtain the accuracy score result; Based on the accuracy score, the accuracy of the customer's pre-loan due diligence information is determined, and accuracy assessment information is obtained. Based on the accuracy assessment information, corresponding correction suggestions are generated.

[0018] By adopting the above technical solution, semantic consistency scoring, terminology consistency verification results, numerical consistency verification results, risk conclusion consistency verification results, and entity recognition consistency verification results can be summarized from multiple dimensions to construct a unified consistency index set. This index set is then input into an accuracy assessment model for comprehensive scoring, thereby quantitatively determining the overall accuracy of the customer's pre-loan due diligence information. Based on this, structured accuracy assessment information is generated according to the scoring results, and targeted correction suggestions are automatically generated in conjunction with various sources of inconsistency. This achieves a closed-loop processing logic from verification and scoring to feedback, effectively reducing the risk of subjective intervention in manual judgment, improving the systematicness, transparency, and interpretability of the assessment results, and enhancing the intelligence level and correction efficiency of the bank's report review process.

[0019] The second objective of this invention is achieved through the following technical solution: An apparatus for assessing the accuracy of bank reports, the apparatus comprising: The original material information acquisition module is used to acquire the original material information of the target bank's customers; The due diligence generation module is used to obtain customer pre-loan due diligence information corresponding to the original material information using an artificial intelligence model; The semantic embedding processing module is used to perform contextual semantic embedding processing on the original material information and the customer pre-loan due diligence information using the BERTScore model, and generate the original material information embedding vector and the due diligence information embedding vector. The semantic consistency calculation module is used to calculate the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and perform greedy matching to obtain a semantic consistency score. The terminology consistency verification module is used to obtain financial terminology information from the customer's pre-loan due diligence information if the semantic consistency score reaches the preset credibility condition, and to perform a normative comparison of the financial terminology information using the PXL-BLEU algorithm combined with standard terms in the preset terminology standard library to obtain the terminology consistency verification result. The numerical consistency verification module is used to identify financial indicator information from the terminology consistency verification results, extract original information for calculating the financial indicator information from the original material information based on the financial indicator information, input the original information into the preset indicator calculation formula, generate corresponding reference values, obtain indicator description values ​​from the customer pre-loan due diligence information, compare the reference values ​​with the description values, and obtain numerical consistency verification results. The risk conclusion consistency verification module is used to identify risk judgment statements related to abnormal financial indicators in the customer's pre-loan due diligence information if abnormal financial indicator information exists in the numerical consistency verification results. Based on the risk judgment statements, the module extracts corresponding risk feature information from the original material information, constructs a risk feature vector based on the risk feature information, and inputs the risk feature vector into a trained risk level assessment model to obtain a predicted risk level. The module then compares the predicted risk level with the risk conclusions in the due diligence information to obtain a risk conclusion consistency verification result. The entity recognition consistency verification module is used to extract customer entity identification information from the customer pre-loan due diligence information based on the semantic consistency score and the terminology consistency verification result, and match the customer entity identification information with the subject identification information in the original material information to obtain the entity recognition consistency verification result.

[0020] By adopting the above technical solutions, artificial intelligence models and semantic modeling techniques can be applied to the content consistency verification process between original material information and customer pre-loan due diligence information. By introducing the BERTScore model for contextual semantic embedding and vector matching, automatic semantic comparison is achieved. The PXL-BLEU algorithm is used for normative matching of financial terms. At the same time, the financial data is analyzed for numerical consistency by combining preset indicator calculation formulas. The consistency of risk conclusions is verified by using a trained risk level assessment model. On this basis, entity recognition consistency verification is further used to ensure the accuracy of subject information. Finally, the verification results of each dimension are summarized into an accuracy score and correction suggestions, which completely replaces the traditional method of manual item-by-item review and rule engine-assisted judgment, effectively improving review efficiency, reducing labor costs, reducing the risk of misjudgment and omission, and enhancing the accuracy of the bank report review process.

[0021] The above-mentioned objective three of this application is achieved through the following technical solution: A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the aforementioned method for assessing the accuracy of bank reports.

[0022] The fourth objective of this application is achieved through the following technical solution: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method for assessing the accuracy of bank reports.

[0023] In summary, this application includes at least one of the following beneficial technical effects: 1. It can apply artificial intelligence models and semantic modeling technology to the content consistency verification process between original material information and customer pre-loan due diligence information. By introducing the BERTScore model for contextual semantic embedding and vector matching, it achieves automatic semantic comparison. It uses the PXL-BLEU algorithm to perform normative matching of financial terms, and combines preset indicator calculation formulas to perform numerical consistency analysis of financial data. It also uses a trained risk level assessment model to verify the consistency of risk conclusions. On this basis, it further ensures the accuracy of subject information through entity recognition consistency verification. Finally, it summarizes the verification results of each dimension into an accuracy score and correction suggestions, which completely replaces the traditional method of manual item-by-item review and rule engine-assisted judgment, effectively improving review efficiency, reducing labor costs, reducing the risk of misjudgment and omission, and enhancing the accuracy of the bank report review process. 2. It can structure and organize financial terminology information from pre-loan due diligence into a set of terms to be compared. By introducing the PXL-BLEU algorithm, it performs multiple rounds of similarity scoring calculations between each term to be compared and the standard terms in the preset terminology specification library. Combined with the preset similarity threshold, it determines whether there is a standard matching relationship between the terms. For unmatched terms, it further generates alternative terminology candidates and performs secondary scoring correction, which effectively improves the accuracy and coverage of terminology matching. At the same time, by calculating the terminology matching accuracy rate in the final matching state, it quantifies the compliance of terminology expression and avoids the problem of incorrect judgment caused by differences in terminology expression and lack of industry experience in manual review, thereby realizing the automation and objectification of the terminology standard comparison process. 3. It can summarize semantic consistency scores, terminology consistency verification results, numerical consistency verification results, risk conclusion consistency verification results, and entity recognition consistency verification results from multiple dimensions to construct a unified consistency index set. This index set is then input into an accuracy assessment model for comprehensive scoring, thereby quantitatively determining the overall accuracy of the customer's pre-loan due diligence information. Based on this, it further generates structured accuracy assessment information according to the scoring results and automatically generates targeted correction suggestions based on various sources of inconsistency. This achieves a closed-loop processing logic from verification and scoring to feedback, effectively reducing the risk of subjective intervention in manual judgment, improving the systematicness, transparency, and interpretability of the assessment results, and enhancing the intelligence level and correction efficiency of the bank's report review process. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of a method for assessing the accuracy of bank reports in one embodiment of the application; Figure 2 This is a flowchart illustrating the implementation of step S30 in a method for assessing the accuracy of bank reports according to an embodiment of this application. Figure 3 This is a flowchart illustrating the implementation of step S40 in a method for assessing the accuracy of bank reports according to an embodiment of this application. Figure 4 This is a flowchart illustrating the implementation of step S50 in a method for assessing the accuracy of bank reports according to an embodiment of this application. Figure 5 This is a flowchart illustrating the implementation of step S70 in a method for assessing the accuracy of bank reports according to an embodiment of this application. Figure 6 This is a flowchart illustrating the implementation of step S80 in a method for assessing the accuracy of bank reports according to an embodiment of this application. Figure 7 This is a flowchart illustrating the implementation of step S80 in a method for assessing the accuracy of bank reports according to an embodiment of this application. Figure 8 This is a schematic block diagram of a bank report accuracy assessment device according to one embodiment of this application; Figure 9 This is a schematic diagram of a device according to one embodiment of this application. Detailed Implementation

[0025] The present application will be further described in detail below with reference to the accompanying drawings.

[0026] In one embodiment, such as Figure 1 As shown, this application discloses a method for assessing the accuracy of bank reports, which specifically includes the following steps: S10: Obtain the original material information of the target bank's customers.

[0027] Specifically, loan application requests can be received from banking business processing systems or third-party credit service platforms. These requests must include at least the customer's unique identifier, project number, or project application form number. By connecting to data sources such as project databases, enterprise knowledge bases, and financial information center interfaces, the relevant original material information can be retrieved using the customer identifier or project number as the query index. The original material information typically includes the project name, industry, estimated investment amount, proposed financing amount, implementing entity, project background, explanation of fund usage, historical operating data, project financial structure, information on cooperating units, and project timeline. After retrieval, the original material information can be initially encapsulated in JSON or XML format and stored in the processing buffer to obtain original material information with a unified format structure and field label definitions.

[0028] S20: Utilize artificial intelligence models to obtain customer pre-loan due diligence information corresponding to the original material information.

[0029] In this embodiment, the artificial intelligence model refers to a text generation model pre-trained based on large-scale language modeling capabilities. It has the ability to perform semantic understanding, language reconstruction, and contextual reasoning on input text information and is used for language modeling of original material information and generation of due diligence information, including but not limited to sequence generation models.

[0030] Specifically, the original material information is categorized by field and input into the artificial intelligence model. The original material information includes corporate financial statements, credit records, project description documents, and business registration information in various formats. Before input, the structured and unstructured data should be formatted and vectorized respectively. The structured data can be converted into feature vectors through field mapping rules, and the unstructured text can be embedded and encoded based on sentence segmentation and word vector models. Then, all the above content is combined into a unified data representation and fed into the artificial intelligence model. The artificial intelligence model completes the syntactic organization and paragraph arrangement through the trained hierarchical content generation structure. The natural language generation module outputs due diligence statement fragments that are complete, well-organized, and semantically compliant. Each statement fragment is spliced ​​together according to a preset template to form a complete report text covering structural parts such as project overview, operating conditions, financial indicators, compliance information, and risk warnings, so as to obtain customer pre-loan due diligence information that corresponds semantically to the original material information.

[0031] S30: Use the BERTScore model to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information, respectively, to generate the original material information embedding vector and the due diligence information embedding vector. In this embodiment, the BERT Score model refers to a deep language understanding model used for text semantic embedding and similarity analysis. It is built on the BERT structure, can perform context encoding on the input sentence and generate word-level vector representations, supports semantic alignment and accurate similarity scoring between two text segments, and is used for context embedding and vector similarity calculation in the semantic consistency scoring process.

[0032] Specifically, sentence segmentation can be performed on the original material information to organize it into a sequence of semantic units. Each semantic unit is then segmented, word vector retrieved, and encoded. The encoded semantic unit vector sequence is input into a pre-trained BERTScore model for context modeling. The BERTScore model models the contextual dependencies between semantic units through a multi-layer Transformer network and outputs an embedding representation sequence. The context representation vector corresponding to each semantic unit is selected from the embedding representation sequence and vector aggregation is performed. Vector aggregation includes performing average pooling on all semantic unit embedding representations or selecting representative position vectors to form a unified vector, thus completing the construction of the original material information embedding vector. The same process is applied to customer pre-loan due diligence information. Semantic unit segmentation, encoding, modeling, and aggregation are performed on the customer pre-loan due diligence information to construct the customer pre-loan due diligence information embedding vector, thus obtaining the original material information embedding vector and the customer pre-loan due diligence information embedding vector.

[0033] S40: Calculate the cosine similarity between the original material information embedding vector and the due diligence information embedding vector, and perform greedy matching to obtain a semantic consistency score.

[0034] In this embodiment, greedy matching refers to an iterative matching algorithm that, in the process of constructing semantic matching relationships, selects the pair with the highest score among the currently unmatched sub-vector pairs based on the similarity score matrix as the matching result of the current round, and immediately marks the sub-vector pair as matched. Then, the same strategy is continued from the remaining unmatched sub-vectors until all matchable sub-vectors have been processed. This algorithm does not backtrack on the selected pairs and emphasizes selecting the local optimum at each step to approximately obtain the overall best matching result.

[0035] Specifically, an equidistant partitioning operation is performed on the original material information embedding vector, splitting it into a first sub-vector set. Simultaneously, an equidistant partitioning operation is performed on the due diligence information embedding vector to generate a second sub-vector set. Each sub-vector in the first sub-vector set is combined with each sub-vector in the second sub-vector set to form a sub-vector pair. A cosine similarity calculation is performed between each sub-vector pair, measuring the degree of similarity by calculating the cosine of the angle between the sub-vectors. The similarity scores of all sub-vector pairs are then combined to form a similarity score matrix. After constructing the score matrix, a greedy matching operation is performed. This operation selects the sub-vector pair corresponding to the current maximum unmatched score from the score matrix and records this sub-vector pair as a semantic matching pair. This process is repeated until all sub-vectors in the first and second sub-vector sets are matched. Each round of matching ensures that matched sub-vectors are not selected repeatedly. Finally, the similarity scores of all semantic matching pairs are weighted or averaged to obtain a semantic consistency score.

[0036] S50: If the semantic consistency score reaches the preset credibility condition, financial terminology information is obtained from the customer's pre-loan due diligence information, and the PXL-BLEU algorithm is used in combination with the standard terms in the preset terminology standard library to perform a normative comparison of the financial terminology information to obtain the terminology consistency verification result.

[0037] In this implementation, the preset credibility condition refers to a numerical threshold used to judge the reliability of semantic consistency scores. It is typically represented by a floating-point number between 0 and 1 and is used to filter content with high semantic similarity to trigger subsequent terminology verification and consistency processing. Financial terminology information refers to terms with financial characteristics extracted from pre-loan due diligence information, including financial terms, risk terms, indicator names, report-type expressions, and specific financial expression structures. Common examples include "debt-to-equity ratio," "capital adequacy ratio," and "project payback period." The PXL-BLEU algorithm is a terminology similarity calculation method that integrates traditional BLEU scoring ideas with a position-aware mechanism. It comprehensively calculates multiple dimensions such as n-gram overlap rate, semantic embedding similarity, and position offset penalty factor to measure the degree of matching between candidate terms and standard terms. Standard terms refer to standardized terminology entries stored in a terminology specification library. They possess characteristics such as clear semantics, standardized format, and industry universality, and serve as a comparison benchmark to determine whether the terminology expressions in customer reports are compliant. Examples include "return on equity" and "short-term solvency."

[0038] Specifically, when the semantic consistency score meets the preset credibility condition, term extraction can be performed on the customer's pre-loan due diligence information. The term extraction operation includes joint processing of high-frequency matching recognition based on financial domain dictionaries and proper noun recognition algorithms based on part-of-speech tagging models. The identified term set is organized into a term set to be compared. Each term in the term set to be compared is used as an input item to calculate similarity with all standard terms in the term specification library. The PXL-BLEU algorithm is used to perform similarity score calculation. The PXL-BLEU algorithm first performs position-sensitive matching processing on the n-gram semantic fragments between term pairs and constructs word fragments. The structure is aligned, and then semantic vectors are combined to perform semantic layer fusion and output similarity scores. After obtaining a set of term matching scores, each term is judged to be successfully matched according to a preset similarity threshold. If the matching score is higher than the threshold, it is marked as a matched state; if it is lower than the threshold, it is marked as an unmatched state. All unmatched terms are summarized to form a set of alternative candidate terms. Then, the terms in the alternative candidate term set are compared with the standard terms again to perform PXL-BLEU scoring and update the corresponding matching status. Finally, the matching status of all terms is combined to form a complete term status mapping, and the term matching accuracy is calculated based on the mapping to obtain the term consistency verification result.

[0039] S60: Financial indicator information identified from the terminology consistency verification results; based on the financial indicator information, extract the original information for calculating the financial indicator information from the original material information; input the original information into the preset indicator calculation formula to generate the corresponding reference value; obtain the indicator description value from the customer's pre-loan due diligence information; compare the reference value with the description value to obtain the numerical consistency verification result.

[0040] In this embodiment, the preset indicator calculation formula refers to the pre-configured standard financial indicator calculation expression, which is used to convert the numerical fields in the original material information into standard indicator values ​​that can be compared with the values ​​described in the customer report. Common forms include "asset-liability ratio = total liabilities / total assets", etc.

[0041] Specifically, based on the set of terms with confirmed semantic matching relationships in the terminology consistency verification results, terms with financial meaning and stable structure are selected as the set of financial indicator names. The mapping relationship between each financial indicator name and the relevant field labels in the original material information is located in this set. Based on this mapping relationship, the original numerical information for indicator calculation is extracted from the original material information. The original numerical information includes structured fields such as total assets, total liabilities, net profit, and operating cash flow. The above original values ​​are combined according to the calculation logic specified in the indicator definition and input into the preset indicator calculation formula. The indicator calculation formula includes ratio-type, growth-type, and composite indicator calculation methods stipulated by the finance department or industry. After the calculation operation is performed, a reference value corresponding to each indicator item is generated. Then, the description position of the financial indicator item is located from the customer's pre-loan due diligence information, and the corresponding indicator description value is extracted. A numerical comparison operation between the reference value and the description value is performed. This comparison operation includes three steps: floating threshold control, unit normalization, and decimal precision alignment. The numerical consistency verification result is obtained by judging whether the values ​​are consistent through the comparison results.

[0042] S70: If abnormal financial indicator information exists in the numerical consistency verification results, identify the risk judgment statement information related to the abnormal financial indicator information in the customer's pre-loan due diligence information. Based on the risk judgment statement information, extract the corresponding risk feature information from the original material information. Based on the risk feature information, construct a risk feature vector and input the risk feature vector into the trained risk level assessment model to obtain the predicted risk level. Compare the predicted risk level with the risk conclusion in the due diligence information to obtain the risk conclusion consistency verification result.

[0043] In this city's case, risk assessment statements refer to natural language sentences with risk-expressing meanings identified from pre-loan due diligence information. These typically include expressions of judgment, speculation, tendency, or conclusion, such as "insufficient operating cash flow" or "potential decline in debt repayment ability." A well-trained risk level assessment model is a deep learning model trained on historical multi-source risk data. It receives risk feature vectors and outputs predicted risk level labels. The model's internal structure includes attention mechanisms or time-evolution trend modeling modules, supporting the fusion of scores from multiple risk factors.

[0044] Specifically, financial indicators with discrepancies exceeding a preset deviation threshold are identified in the numerical consistency verification results. Each abnormal financial indicator is then input as a keyword into the statement set of the client's pre-loan due diligence information. Keyword matching and syntactic dependency identification operations are performed to locate statements containing risk judgment meanings from the syntactic structure. These statements typically include expressive features such as "may lead to," "risk exists," "liquidity stress," and "decline in debt repayment ability." These statements are compiled into a risk judgment statement information set, and risk clue words that can be used for reasoning are extracted from each risk judgment statement. These risk clue words are then used to perform cross-segment queries and field comparisons on the original material information to extract information directly related to the risk theme. The associated quantitative or qualitative data, including asset change trends, liability structure, short-term debt repayment sources, historical default records, and credit rating results, are standardized and normalized to form a risk feature set. This risk feature set is then encoded into a vector generation module to generate a risk feature vector with uniform dimensions. The risk feature vector is then input into a pre-trained risk level assessment model. The model analyzes the combination relationships between various risk factors through an attention weight mechanism and outputs a level score. The score is mapped to a risk level label and compared item by item with the original risk conclusion labels in the customer's pre-loan due diligence information to obtain a consistency verification result for the risk conclusion.

[0045] S80: Based on the semantic consistency score and terminology consistency verification results, extract customer entity identification information from the customer's pre-loan due diligence information, match the customer entity identification information with the subject identification information in the original material information, and obtain the entity identification consistency verification result.

[0046] In this embodiment, customer entity identification information refers to a set of key fields extracted from the customer's pre-loan due diligence information that identify the customer's identity, including company name, project name, legal representative, unified social credit code, registered address, etc., which can be used to compare the entity with the original materials. Entity identification information refers to a set of fields extracted from the original material information that have legal or business identification effect, structurally corresponding to the customer entity identification information, and used to construct an entity consistency verification relationship with the customer's pre-loan due diligence information.

[0047] Specifically, based on the semantic consistency score and terminology consistency verification results, the scope of content with business credibility in the text is determined. Within this credibility scope, named entity recognition is performed on the customer's pre-loan due diligence information. The named entity recognition model decodes the sentence's tag sequence based on financial domain word vectors and context recognition rules, extracting information fields representing the customer's uniqueness as customer entity identification information. Customer entity identification information includes, but is not limited to, a set of fields such as project name, company name, legal representative's name, registered address, unified social credit code, and contact information. After extraction, format standardization processing is performed, including removing spaces, unifying capitalization, replacing common abbreviations, and cleaning punctuation. Simultaneously, the corresponding subject identification fields are extracted from the original material information and processed in the same format. Then, the two field sets are input into the field-level joint matching module to perform field alignment. Field alignment adopts a sequentially independent full comparison strategy. Each pair of field combinations performs semantic similarity calculation based on word vectors and edit distance calculation based on strings. The calculation results are weighted and fused to form an entity matching score matrix. Then, a scoring threshold determination operation is performed on the entity matching score matrix, outputting all field-level matching statuses and making a comprehensive judgment to obtain the entity recognition consistency verification result.

[0048] In one embodiment, such as Figure 2 As shown, in step S30, the BERTScore model is used to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information, generating embedding vectors for the original material information and due diligence information, including: S301: The BERTScore model performs multi-layer encoding on the original material information to obtain an embedded representation sequence of the original material information containing context-dependent features. Vector aggregation processing is then performed on the original material information embedded representation sequence to generate the original material information embedded vector.

[0049] Specifically, the original material information is divided according to sentence boundaries and segmented into words according to natural language processing rules. The segmented results are input into the BERTScore model. The BERTScore model processes the semantic dependency relationship between each word and its context sequentially through a multi-layer Transformer encoding structure. The attention mechanism in the model associates and scores each word with words in other positions in the sentence and generates a weighted combination of context representations, forming an embedding representation sequence of the original material information. After obtaining the embedding representation sequence, vector aggregation processing is performed. The aggregation processing includes performing average pooling on all word representations or selecting CLS vectors as sentence-level representations. The aggregated representation is used as the global semantic embedding vector of the original material information under the current encoding structure to generate the original material information embedding vector.

[0050] S302: The BERTScore model performs multi-layer encoding on customer pre-loan due diligence information to obtain a sequence of due diligence information embedding representations. Vector aggregation processing is then performed on the sequence of due diligence information embedding representations to generate due diligence information embedding vectors.

[0051] Specifically, the pre-loan due diligence information is segmented at the sentence level and encoded using a language model vocabulary. The encoded word sequence is then input into the BERTScore model. The BERTScore model uses a multi-layer Transformer structure to extract the semantic relationships of each word in its context and establish a global dependency representation. The intermediate representation output from each layer continues to iterate semantic aggregation and feature propagation in the next layer until the entire semantic modeling process is completed. The model outputs an embedding representation sequence composed of the contextual semantic representations of each word. After obtaining the complete representation sequence, vector aggregation is performed on all word representations. Vector aggregation includes performing average pooling on the representation sequence to calculate sentence-level embedding vectors, or selecting the semantic representation at the CLS marker position as the global vector representation of the sentence. After aggregation, the due diligence information embedding vector is output. In one embodiment, such as Figure 3 As shown, in step S40, the cosine similarity of the original material information embedding vector and the due diligence information embedding vector is calculated and a greedy matching is performed to obtain a semantic consistency score, including: S401: Divide the original material information embedding vector into a first sub-vector set, and divide the due diligence information embedding vector into a second sub-vector set. Specifically, a fixed window sliding partitioning process is performed on the original material information embedding vector according to the dimensional structure of the original material information embedding vector. During the sliding partitioning process, the dimensional length of each sub-vector is set to a preset constant and the step size is set to a non-overlapping mode or a semi-overlapping mode. Continuous dimensions are extracted sequentially along the vector dimensional direction to form multiple sub-vectors, resulting in the first sub-vector set. Similarly, the due diligence information embedding vector is partitioned according to the same dimensional rules to generate the second sub-vector set.

[0052] S402: Combine the first sub-vector set and the second sub-vector set, perform cosine similarity calculation between each pair of sub-vectors, and generate a similarity score matrix.

[0053] Specifically, a pairwise combination operation is performed on the first sub-vector set and the second sub-vector set. Each combination consists of a sub-vector from the first sub-vector set and a sub-vector from the second sub-vector set. Cosine similarity is calculated for each pair of sub-vectors. The cosine similarity calculation is performed by normalizing the two sub-vectors using the L2 norm and then calculating their vector dot product. The dot product result is the similarity score of the current sub-vector pair. The similarity scores of all sub-vector pairs are filled into a two-dimensional matrix structure according to the combination order of the first and second sub-vector sets. Each row of the matrix corresponds to a sub-vector from the first sub-vector set, each column corresponds to a sub-vector from the second sub-vector set, and each element of the matrix is ​​the similarity score between the corresponding row and column sub-vector pairs, thus constructing the completed similarity score matrix.

[0054] S403: Based on the similarity score matrix, a greedy matching algorithm is used. In each round of matching, a pair of sub-vectors with high similarity scores is selected from the first set of sub-vectors that have not yet participated in the matching and the second set of sub-vectors that have not yet participated in the matching.

[0055] Specifically, all candidate subvector pairs and their corresponding similarity scores are extracted from the similarity score matrix. Before each round of matching begins, two unmatched marker sets are constructed, marking the index positions of unselected subvectors in the first and second subvector sets respectively. Subvector pairs and their similarity scores at corresponding positions in the current two unmatched marker sets are selected from the similarity score matrix. The subvector pair with the highest similarity score is selected as the matching result for the current round. The two corresponding index positions of this subvector pair are removed from the unmatched marker set and added to the matched marker set. After updating, the next round of matching is performed. The above operation is repeated until all subvectors in any subvector set have participated in the matching.

[0056] S404: Mark the subvector pair as matched and add the subvector pair to the semantic match set; In this embodiment, the semantic matching set refers to the set of all sub-vector pairs marked as having high similarity during the process of performing embedded vector cosine similarity calculation and greedy matching. It is used to calculate semantic consistency scores and is an intermediate structure of similarity calculation output.

[0057] Specifically, after completing the similarity score filtering of the current pair of subvectors, the index position of the subvector pair in the first subvector set and the second subvector set is recorded respectively. The index position is marked as occupied in the corresponding unmatched tag set, and the matching status mapping table is updated synchronously. After the marking is completed, the index position of the subvector pair and its corresponding similarity score form a matching pair structure, and the matching pair structure is appended to the semantic matching set.

[0058] S405: After selecting all sub-vector pairs, perform an average calculation on the similarity scores of the sub-vector pairs in the semantic matching set to generate a semantic consistency score.

[0059] Specifically, after completing the greedy matching operation for all sub-vector pairs, the semantic matching set is traversed, and the similarity score value in the structure of each matching pair is read one by one. All score values ​​are stored in the scoring vector list in the original order. The total similarity score is obtained by performing a numerical summation operation on all elements in the scoring vector list. The total number of elements in the scoring vector list is taken as the number of matching pairs. The mean is calculated by dividing the total similarity score by the number of matching pairs. Floating-point precision control can be set during the mean calculation to unify the result format. The obtained mean is used as the semantic consistency score between the current original material information and the customer's pre-loan due diligence information.

[0060] In one embodiment, such as Figure 4 As shown, in step S50, financial terminology information is obtained from the customer's pre-loan due diligence information, and the PXL-BLEU algorithm is used in conjunction with standard terms in a preset terminology standardization library to perform a standardization comparison of the financial terminology information, obtaining terminology consistency verification results, including: S501: Construct a set of terms to be compared based on financial terminology information.

[0061] Specifically, based on the financial terminology information extracted from the customer's pre-loan due diligence information, the term set is screened and organized according to the naming conventions and domain part-of-speech features of financial terms. First, part-of-speech tagging and entity recognition algorithms are used to identify terms that are proprietary, indicative, or categorized. Then, deduplication, normalization, and format standardization operations are performed on the term entries. The deduplication operation uses precise string matching or a thesaurus to process semantically equivalent expressions. The normalization operation expands abbreviations to full names and unifies the terminology granularity. The format standardization operation includes unifying capitalization, removing symbols, and standardizing units. Finally, the processed term entries are organized into a structured list, where each element is a term to be compared, in order to construct a set of terms to be compared for subsequent terminology standardization verification.

[0062] S502: Using the PXL-BLEU algorithm, the terms to be compared in the term set to be compared are compared with the standard terms in the preset term specification library to calculate the similarity score and obtain the corresponding term matching score set.

[0063] Specifically, each term in the set of terms to be compared is used as an input item in turn, and a one-to-many similarity score is calculated with all standard terms registered in the preset term specification library. In each scoring process, the term to be compared and the standard term are first segmented into n-gram sequences and an n-gram set is constructed. The basic overlap rate score is obtained by calculating the overlap and semantic distance between the two n-gram sets. Based on the basic score, the position information weight coefficient of the PXL-BLEU algorithm is introduced to penalize and adjust the position offset. At the same time, the term character components and their semantic embedding representation vectors are input into the semantic similarity evaluation module to further supplement the semantic layer correction. Finally, a weighted fusion of multiple scoring factors is performed to generate the final matching score between each term pair. The matching score, along with the identifiers of the term to be compared and the corresponding standard term, is stored in the scoring record list. After scoring is completed item by item, all matching results are sorted and summarized into a term matching score set.

[0064] S503: Based on the term matching score set, determine whether each term to be compared has a matching relationship with any standard term in the term specification library with a matching score higher than the preset similarity threshold. If it exists, mark the term to be compared as a standard matching state; otherwise, mark the term to be compared as an unmatched state.

[0065] Specifically, the matching records of each term to be compared are iterated through in the term matching score set. Each matching record contains the similarity score results between the current term to be compared and all standard terms in the preset term specification library. For each set of score results, the maximum matching score value is extracted and compared with the preset similarity threshold. The similarity threshold is a numerical boundary set by the system to distinguish between semantic similarity and non-similarity. When the maximum matching score value is higher than the threshold, the current term to be compared is marked as a standard matching state, indicating that the term has found a semantic matching item in the term specification library. If the maximum score does not reach the threshold, it is determined that the current term has failed to establish a matching relationship with any standard term and is marked as an unmatched state.

[0066] S504: Summarize the terms to be compared corresponding to the unmatched states to obtain a set of term replacement candidates.

[0067] Specifically, a state filtering operation is performed on all terms to be compared in the terminology state mapping structure. The filtering condition is that the matching state is marked as an unmatched state. All terms to be compared that meet the condition are extracted and rearranged according to the original order or domain priority. After the rearrangement, a preliminary set of alternative terms is formed. For each term in the preliminary set of alternative terms, spelling similarity filtering, synonym expansion query and context semantic verification are further performed. Spelling similarity filtering is used to identify similar terms due to format or input errors. Synonym expansion query is based on the financial terminology database to retrieve equivalent expressions or industry alternatives. Context semantic verification ensures that the term usage context is consistent by associating with the content of the context sentences. After the above processing is completed, a structured list is formed and stored as a set of candidate alternative terms.

[0068] S505: Using the PXL-BLEU algorithm, the similarity score is calculated between the terms to be compared in the candidate term replacement set and the standard terms, and the matching status of the corresponding terms to be compared is updated to obtain the updated matching status.

[0069] Specifically, each term to be compared is read sequentially from the set of candidate term replacements. The current term to be compared is used as input and a one-to-many similarity score is calculated with all standard terms in the preset term specification library. The scoring method calls the PXL-BLEU algorithm. The PXL-BLEU algorithm performs position-sensitive n-gram fragment matching and semantic representation fusion based on the character combination structure and contextual nesting relationship of the terms. By modeling the spelling layer and semantic layer of the two terms respectively and weighting and fusing them, the final similarity score is generated. The score is compared with a preset similarity threshold. When the score is higher than the threshold, the matching status of the term to be compared is updated from unmatched to standard matching status. If the score is still lower than the threshold, the original unmatched status is retained. After re-scoring all replacement candidate terms, a new state mapping structure is formed, and this structure is used as the updated matching status.

[0070] S506: Merge the updated matching status with the standard matching status to obtain the final matching status information.

[0071] Specifically, after completing the similarity scoring and status update of the term substitution candidates, the updated matching status structure and the initial standard matching status structure are read respectively. The two structures are indexed and compared according to the term to be compared as the primary key. For duplicate terms that exist in both structures, the updated matching status is retained first. For terms that exist only in the standard matching status structure, the original status is directly retained. After merging all term statuses, a unified status mapping table is constructed. Each record in this mapping table corresponds to a term to be compared and its final matching status mark. The mark value is standard match or no match. The status merging operation achieves data coverage and completion through key-value association and conflict update strategies. Finally, all merging results are organized and output according to the original term order to obtain the final matching status information.

[0072] S507: Calculate the term matching accuracy based on the final matching status information, and generate a term consistency verification result based on the term matching accuracy.

[0073] Specifically, each term record in the final matching status information is statistically processed. During the statistical process, the number of terms marked as standard matching status and the total number of all terms to be compared are calculated separately. The number of standard matching terms is taken as the hit count, and the total number of terms to be compared is taken as the baseline count. The term matching accuracy is obtained by dividing the hit count by the baseline count. The term matching accuracy is expressed in decimal form and retains the specified number of precision digits. After the accuracy is calculated, the value is mapped to the set term consistency level standard. If the accuracy is higher than the high consistency threshold, it is judged as high consistency level. If it is in the middle range, it is marked as medium consistency. If it is lower than the low consistency threshold, it is output as low consistency level. The consistency level label and the accuracy value are written into the term consistency verification result.

[0074] In one embodiment, such as Figure 5 As shown, in step S70, a risk feature vector is constructed based on the risk feature information, and the risk feature vector is input into the trained risk level assessment model to obtain the predicted risk level, including: S701: The trained risk level assessment model models the correlation between risk factors in the risk feature vector based on the attention mechanism, and performs weighted processing in combination with the time evolution trend of risk factors. After completing the modeling of the risk feature vector, it outputs the overall risk level score.

[0075] Specifically, the constructed risk feature vector is input into the trained risk level assessment model. The risk feature vector contains multiple factor dimensions reflecting the customer's risk status. These risk factors include financial indicator volatility, debt repayment capacity indicator sequence, historical distribution of credit rating, default record statistics, and liquidity status score within the enterprise's operating cycle. The risk level assessment model uses a multi-head attention mechanism to model the semantic relationship between all risk factors. The interaction strength between factors is measured through an attention weight matrix. Furthermore, a time evolution trend parameter is introduced to adjust the weight structure during the modeling process, so that the attention ratio is increased on risk factors with consistent historical change direction or significant trend. After completing the inference of each layer structure, the model maps the weighted and fused feature representation into a risk score value through the output layer to obtain the overall risk level score.

[0076] S702: Based on the preset risk level classification rules, the overall risk level score is mapped to the predicted risk level.

[0077] In this embodiment, the preset risk level classification rule refers to the segmentation rule used to map the overall risk level score to a risk level label. It consists of multiple score intervals and corresponding risk levels, such as: [0–0.3] is low risk, [0.3–0.6] is medium risk, and [0.6–1] is high risk. The label can be configured as a level code or text description.

[0078] Specifically, a set of risk level classification rules is preset based on the numerical range of the overall risk level score. These rules divide the risk score interval into multiple non-overlapping level segments according to the numerical threshold. Each level segment corresponds to a risk level label, which includes five levels: low risk, low-medium risk, medium risk, medium-high risk, and high risk. The overall risk level score is compared with the boundaries of each segment one by one to determine the interval position to which the score value belongs, thereby completing the mapping operation from numerical value to level label. After the mapping is completed, the risk level label is returned as the model output result to obtain the predicted risk level.

[0079] In one embodiment, such as Figure 6 As shown, in step S80, the customer entity identification information is matched with the entity identification information in the original material information to obtain the entity identification consistency verification result, including: S801: Obtain target entity identification information from customer entity identification information and reference entity identification information from original material information.

[0080] In this embodiment, the target entity identification information refers to the set of entity identification fields extracted and standardized from the customer's pre-loan due diligence information. This set serves as the comparison target for matching and verifying entity information in the original materials, and has a complete field structure and naming conventions. The reference entity identification information refers to the set of fields related to customer identification extracted from the original material information. This set serves as the comparison object for the target entity identification information, used to identify information consistency and ensure the accurate attribution of the reporting entity.

[0081] Specifically, entity field extraction is performed on customer pre-loan due diligence information. A named entity recognition model is used to identify customer-related information fields with unique identifiers. The identified fields include company name, project name, legal representative's name, registered address, unified social credit code, contact information, and bank of account. These fields are combined according to predefined field standards to construct target entity identifier information. Subsequently, the original material information is subjected to structured field parsing and text content extraction. Data items corresponding to the above customer-related fields are located from the original material information, and the extraction process of the same field name and data value is completed. The extracted fields are combined and organized into reference entity identifier information to obtain target entity identifier information and reference entity identifier information used for subsequent matching calculations.

[0082] S802: Using a field-level joint matching strategy, the target entity identifier information and the reference entity identifier information are aligned and similarity is calculated to generate an entity matching score matrix.

[0083] In this embodiment, the field-level joint matching strategy refers to a matching method using fields as the basic matching unit when comparing customer entity identification information with subject identification information in the original material information. Specifically, it includes multiple operation steps such as field name standardization, field content format unification, field label alignment, and field value similarity calculation. Field name standardization identifies semantically identical field names through synonym mapping or label unification rules. Field content format unification includes character encoding consistency, number format standardization, and case unification. Field label alignment is used to establish a one-to-one matching mapping relationship. Field value similarity calculation comprehensively uses edit distance algorithm, word vector similarity analysis, and spelling similarity scoring. Finally, the matching results of each field are aggregated into an overall entity matching score to support entity recognition consistency verification.

[0084] Specifically, the target entity identifier information and the reference entity identifier information are grouped and mapped according to field labels. In the field-level joint matching strategy, a field alignment rule table is first constructed. This rule table uniformly maps the variations, abbreviations and aliases of common field names. For example, "enterprise name" and "company name", "unified social credit code" and "credit code" are processed in a one-to-one correspondence manner. After completing the field label alignment, a similarity calculation operation is performed on each pair of field values. The similarity calculation includes the fusion of three algorithms: edit distance calculation, word vector similarity analysis and spelling similarity scoring. Edit distance calculation measures the degree of character deformation, word vector analysis is used to handle semantic expression deviation, and spelling similarity assesses the difference between abnormal abbreviations and symbols. The above results are weighted and fused to form the matching score of each pair of field values. Finally, the matching scores of all field pairs are filled into a two-dimensional matrix structure in a row and column manner to obtain the entity matching score matrix.

[0085] S803: By performing a threshold judgment operation on the entity matching score matrix, the matching relationship status between the target entity identification information and the reference entity identification information is determined, and the entity recognition consistency verification result is obtained.

[0086] Specifically, the matching score of each pair of fields in the entity matching score matrix is ​​compared with a preset similarity threshold. The comparison method is to perform a greater than or equal to check on the score value of each cell in the matrix. If the score of a single field is higher than the similarity threshold, the field is marked as a matching field; otherwise, it is marked as a non-matching field. After completing the field-level judgment, the number of matching fields is counted, and the proportion of matching fields to the total number of fields is calculated. This proportion is used as the basis for overall entity consistency and is matched against the set consistency level rules. If the proportion exceeds the standard threshold for passing consistency verification, the matching relationship between the target entity identifier information and the reference entity identifier information is marked as consistent; otherwise, it is marked as inconsistent, thus obtaining the entity recognition consistency verification result. In one embodiment, such as Figure 7 As shown, after step S80, i.e., a method for assessing the accuracy of bank reports, it also includes; S804: Summarize the semantic consistency score, terminology consistency verification results, numerical consistency verification results, risk conclusion consistency verification results, and entity recognition consistency verification results to obtain a multidimensional consistency index set.

[0087] Specifically, semantic consistency score is used as the first dimension indicator, term matching accuracy in terminology consistency verification results is used as the second dimension indicator, numerical deviation rate or numerical consistency label in numerical consistency verification results is used as the third dimension indicator, risk level comparison result in risk conclusion consistency verification results is used as the fourth dimension indicator, and entity matching status in entity recognition consistency verification results is used as the fifth dimension indicator. The results of the above five dimensions are normalized and formatted according to a unified data structure. The processing includes unifying the decimal places of the score value, performing Boolean conversion on the comparison label, mapping and encoding the classification label, and writing it into the consistency indicator data container in a fixed field order to construct an indicator structure containing five sub-items. Finally, this structure is output as a set of overall data objects to obtain a multi-dimensional consistency indicator set.

[0088] S805: Input the multidimensional consistency index set into the accuracy assessment model to obtain the accuracy score result.

[0089] In this embodiment, the accuracy assessment model refers to a model structure that receives a set of multidimensional consistency indicators and outputs an overall accuracy score. It is usually a regression-type deep learning network that can evaluate the comprehensive accuracy level of pre-loan due diligence information by modeling the cross-features of input indicators from various dimensions.

[0090] Specifically, a set of multidimensional consistency indicators is input into the accuracy scoring model according to a predetermined field order. The accuracy scoring model is a scoring inference model trained based on a multilayer perceptron structure. The model input vector contains five types of feature fields: semantic consistency score, term matching accuracy, numerical consistency status identifier, risk conclusion consistency status identifier, and entity recognition consistency status identifier. During the input phase, the model performs unified normalization and embedding transformation operations on the above input features. The standardized indicator features are passed to the hidden layer through the input layer for combined modeling. A weight allocation mechanism is introduced in the hidden layer to express the degree of interdependence between the consistency indicators. The overall semantic and structural deviation risk is calculated by weighted aggregation of information from different dimensions. Finally, the comprehensive modeling result is mapped to a single numerical score value through the output layer to obtain the accuracy score information.

[0091] S806: Based on the accuracy score, determine the accuracy of the customer's pre-loan due diligence information and obtain accuracy assessment information.

[0092] Specifically, the accuracy score results are compared step-by-step with the preset accuracy level classification standards. The accuracy level classification standards consist of multiple non-overlapping numerical intervals, each interval corresponding to an accuracy level label, including three levels: high accuracy, medium accuracy, and low accuracy. The accuracy score results are compared with the boundary values ​​of each level to determine the overall accuracy level of the current customer pre-loan due diligence information. At the same time, combined with the consistency sub-results of the five dimensions of semantics, terminology, numerical values, risk, and entities, the indicators that are abnormal or deviated are marked and used as supplementary explanations in the evaluation conclusion. By jointly organizing the score levels and abnormal items, an accuracy level label and abnormal indicator structure are constructed and uniformly packaged into accuracy evaluation information to obtain the accuracy evaluation information used for generating subsequent correction suggestions.

[0093] S807: Based on the accuracy assessment information, generate corresponding correction suggestions.

[0094] Specifically, the accuracy assessment scores are jointly analyzed with the consistency results of each dimension. First, the triggering conditions for generating correction suggestions are determined based on the accuracy scores. When the score is lower than the preset accuracy threshold, the correction suggestion generation module is activated. During the generation process, the paragraphs with semantic discrepancies between the original material information and the due diligence information are located based on the sub-items with semantic consistency scores lower than the standard. For cases where the terminology matching accuracy rate does not reach the threshold, a list of unmatched terms is compiled and corresponding standard terms are recommended for replacement. For numerical inconsistencies, the specific deviation between the reference value and the descriptive value is attached. For risk conclusion inconsistencies, the difference level between the predicted risk level and the original conclusion is displayed. At the same time, the customer identification field content for suggested correction is listed based on the entity matching failure field. Finally, the above content is output in a structured format as correction suggestion information to obtain feedback suggestion text that can be checked, modified, and manually intervened.

[0095] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0096] In one embodiment, a device for assessing the accuracy of bank reports is provided, which corresponds one-to-one with a method for assessing the accuracy of bank reports described in the above embodiments. For example... Figure 8 As shown, this device for assessing the accuracy of bank reports includes a raw material information acquisition module, a due diligence generation module, a semantic embedding processing module, a semantic consistency calculation module, a terminology consistency verification module, a numerical consistency verification module, a risk conclusion consistency verification module, and an entity recognition consistency verification module. Detailed descriptions of each functional module are as follows: The original material information acquisition module is used to acquire the original material information of the target bank's customers; The due diligence generation module is used to obtain customer pre-loan due diligence information corresponding to the original material information using an artificial intelligence model; The semantic embedding processing module is used to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information using the BERTScore model, and generate the original material information embedding vector and the due diligence information embedding vector. The semantic consistency calculation module is used to calculate the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and perform greedy matching to obtain a semantic consistency score. The terminology consistency verification module is used to obtain financial terminology information from the customer's pre-loan due diligence information if the semantic consistency score reaches the preset credibility condition. Then, it uses the PXL-BLEU algorithm combined with the standard terms in the preset terminology standard library to perform a normative comparison of the financial terminology information and obtain the terminology consistency verification result. The numerical consistency verification module is used to identify financial indicator information from the terminology consistency verification results. Based on the financial indicator information, it extracts the original information for calculating the financial indicator information from the original material information, inputs the original information into the preset indicator calculation formula, generates the corresponding reference value, obtains the indicator description value from the customer's pre-loan due diligence information, compares the reference value with the description value, and obtains the numerical consistency verification result. The risk conclusion consistency verification module is used to identify risk judgment statements related to abnormal financial indicators in the customer's pre-loan due diligence information if abnormal financial indicator information is found in the numerical consistency verification results. Based on the risk judgment statements, the module extracts the corresponding risk feature information from the original material information, constructs a risk feature vector based on the risk feature information, and inputs the risk feature vector into the trained risk level assessment model to obtain the predicted risk level. The predicted risk level is then compared with the risk conclusions in the due diligence information to obtain the risk conclusion consistency verification results. The entity recognition consistency verification module is used to extract customer entity identification information from the customer's pre-loan due diligence information based on semantic consistency scoring and terminology consistency verification results, and match the customer entity identification information with the subject identification information in the original material information to obtain the entity recognition consistency verification result.

[0097] Optionally, the semantic embedding processing module includes: The original material information semantic embedding generation submodule is used by the BERTScore model to encode the original material information in multiple layers to obtain the original material information embedding representation sequence containing context-dependent features. Vector aggregation processing is performed on the original material information embedding representation sequence to generate the original material information embedding vector. The due diligence information semantic embedding generation submodule is used by the BERTScore model to encode customer pre-loan due diligence information in multiple layers to obtain a due diligence information embedding representation sequence. Vector aggregation processing is then performed on the due diligence information embedding representation sequence to generate due diligence information embedding vectors.

[0098] Optionally, the semantic consistency calculation module includes: The embedded vector substructure partitioning module is used to divide the original material information embedded vector into a first sub-vector set and the due diligence information embedded vector into a second sub-vector set. The sub-vector similarity calculation submodule is used to combine the first sub-vector set and the second sub-vector set, perform cosine similarity calculation between each pair of sub-vectors, and generate a similarity score matrix. The greedy matching execution submodule is used to perform greedy matching based on the similarity score matrix. In each round of matching, it selects sub-vector pairs with high similarity scores from the first sub-vector set that has not yet participated in the matching and the second sub-vector set that has not yet participated in the matching. The semantic matching pair marking submodule is used to mark subvector pairs as matched and add the subvector pairs to the semantic matching set; The semantic consistency score generation submodule is used to perform an average calculation of the similarity scores of the sub-vector pairs in the semantic matching set after all sub-vector pairs have been selected, and generate a semantic consistency score.

[0099] Optionally, the terminology consistency verification module includes: The term matching set construction submodule is used to construct a term set to be compared based on financial term information; The initial term similarity scoring submodule is used to calculate the similarity score between the terms to be compared in the term set to be compared and the standard terms in the preset term specification library using the PXL-BLEU algorithm, and obtain the corresponding term matching score set. The term matching status determination submodule is used to determine, based on the term matching score set, whether each term to be compared has a matching relationship with any standard term in the term specification library with a matching score higher than a preset similarity threshold. If such a relationship exists, the term to be compared is marked as a specification matching state; otherwise, it is marked as an unmatched state. The term substitution candidate generation submodule is used to summarize the terms to be compared corresponding to the unmatched states and obtain a set of term substitution candidates; The term status update scoring submodule is used to calculate the similarity score between the terms to be compared in the candidate term replacement set and the standard terms using the PXL-BLEU algorithm, update the matching status of the corresponding terms to be compared, and obtain the updated matching status. The terminology matching status merging submodule is used to merge the updated matching status with the specification matching status to obtain the final matching status information; The terminology consistency verification generation submodule is used to calculate the terminology matching accuracy based on the final matching status information, and generate the terminology consistency verification result based on the terminology matching accuracy.

[0100] Optionally, the risk conclusion consistency verification module includes: The risk score generation submodule is used to model the correlation between risk factors in the risk feature vector based on the attention mechanism of the trained risk level assessment model, and to perform weighted processing in combination with the time evolution trend of risk factors. After completing the modeling of the risk feature vector, it outputs the overall risk level score. The risk level mapping submodule is used to map the overall risk level score to the predicted risk level according to the preset risk level classification rules.

[0101] Optionally, the entity recognition consistency verification module includes: The entity identification information extraction submodule is used to obtain target entity identification information from customer entity identification information and reference entity identification information from original material information; The entity field matching calculation submodule is used to perform field alignment and similarity calculation on the target entity identifier information and the reference entity identifier information using a field-level joint matching strategy, and generate an entity matching score matrix. The entity consistency judgment submodule is used to determine the matching relationship status between the target entity identification information and the reference entity identification information by performing a threshold judgment operation on the entity matching score matrix, and obtain the entity recognition consistency verification result.

[0102] Optionally, the entity recognition consistency verification module includes the following: The consistency index aggregation module is used to aggregate semantic consistency scores, terminology consistency verification results, numerical consistency verification results, risk conclusion consistency verification results, and entity recognition consistency verification results to obtain a multi-dimensional consistency index set. The accuracy score calculation module is used to input a multidimensional consistency index set into the accuracy evaluation model to obtain the accuracy score result; The due diligence information assessment module is used to determine the accuracy of the client's pre-loan due diligence information based on the accuracy score results, and to obtain accuracy assessment information. The correction suggestion information generation module is used to generate corresponding correction suggestion information based on the accuracy assessment information.

[0103] For specific limitations regarding a device for assessing the accuracy of bank reports, please refer to the limitations regarding a method for assessing the accuracy of bank reports described above, which will not be repeated here. The various modules in the aforementioned device for assessing the accuracy of bank reports can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0104] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-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 stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for assessing the accuracy of bank reports.

[0105] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain original material information from the target bank's customers; Utilize artificial intelligence models to obtain customer pre-loan due diligence information corresponding to the original material information; The BERTScore model is used to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information, respectively, to generate the original material information embedding vector and the due diligence information embedding vector. The cosine similarity of the original material information embedding vector and the due diligence information embedding vector is calculated and greedy matching is performed to obtain the semantic consistency score. If the semantic consistency score reaches the preset credibility condition, then financial terminology information is obtained from the customer's pre-loan due diligence information, and the PXL-BLEU algorithm is used in combination with the standard terms in the preset terminology standard library to perform a normative comparison of the financial terminology information to obtain the terminology consistency verification result. The financial indicator information identified from the terminology consistency verification results is used to extract the original information for calculating the financial indicator information from the original material information. The original information is then input into the preset indicator calculation formula to generate the corresponding reference value. The indicator description value is obtained from the customer's pre-loan due diligence information. The reference value and the description value are compared to obtain the numerical consistency verification result. If abnormal financial indicator information is found in the numerical consistency verification results, the risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information is identified. Based on the risk judgment statement information, the corresponding risk feature information is extracted from the original material information. Based on the risk feature information, a risk feature vector is constructed and input into the trained risk level assessment model to obtain the predicted risk level. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain the risk conclusion consistency verification result. Based on the semantic consistency score and terminology consistency verification results, customer entity identification information is extracted from the customer's pre-loan due diligence information. The customer entity identification information is then matched with the subject identification information in the original material information to obtain the entity identification consistency verification results.

[0106] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: Obtain original material information from the target bank's customers; Utilize artificial intelligence models to obtain customer pre-loan due diligence information corresponding to the original material information; The BERTScore model is used to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information, respectively, to generate the original material information embedding vector and the due diligence information embedding vector. The cosine similarity of the original material information embedding vector and the due diligence information embedding vector is calculated and greedy matching is performed to obtain the semantic consistency score. If the semantic consistency score reaches the preset credibility condition, then financial terminology information is obtained from the customer's pre-loan due diligence information, and the PXL-BLEU algorithm is used in combination with the standard terms in the preset terminology standard library to perform a normative comparison of the financial terminology information to obtain the terminology consistency verification result. The financial indicator information identified from the terminology consistency verification results is used to extract the original information for calculating the financial indicator information from the original material information. The original information is then input into the preset indicator calculation formula to generate the corresponding reference value. The indicator description value is obtained from the customer's pre-loan due diligence information. The reference value and the description value are compared to obtain the numerical consistency verification result. If abnormal financial indicator information is found in the numerical consistency verification results, the risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information is identified. Based on the risk judgment statement information, the corresponding risk feature information is extracted from the original material information. Based on the risk feature information, a risk feature vector is constructed and input into the trained risk level assessment model to obtain the predicted risk level. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain the risk conclusion consistency verification result. Based on the semantic consistency score and terminology consistency verification results, customer entity identification information is extracted from the customer's pre-loan due diligence information. The customer entity identification information is then matched with the subject identification information in the original material information to obtain the entity identification consistency verification results.

[0107] 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 may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of 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.

[0108] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0109] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application, and should all be included within the protection scope of this application.

Claims

1. A method for assessing the accuracy of bank reports, characterized in that, The method includes: Obtain original material information from the target bank's customers; Artificial intelligence models are used to obtain customer pre-loan due diligence information corresponding to the original material information; The BERTScore model is used to perform contextual semantic embedding processing on the original material information and the customer pre-loan due diligence information to generate the original material information embedding vector and the due diligence information embedding vector. The cosine similarity of the original material information embedding vector and the due diligence information embedding vector is calculated and greedy matching is performed to obtain a semantic consistency score. If the semantic consistency score reaches the preset credibility condition, then financial terminology information is obtained from the customer's pre-loan due diligence information, and the PXL-BLEU algorithm is used in combination with the standard terms in the preset terminology standard library to perform a normative comparison of the financial terminology information to obtain the terminology consistency verification result. The financial indicator information identified from the terminology consistency verification results is used to extract the original information for calculating the financial indicator information from the original material information. The original information is then input into the preset indicator calculation formula to generate the corresponding reference value. The indicator description value is obtained from the customer pre-loan due diligence information. The reference value and the description value are compared to obtain the numerical consistency verification result. If abnormal financial indicator information is found in the numerical consistency verification result, the risk judgment statement information associated with the abnormal financial indicator information in the customer pre-loan due diligence information is identified. Based on the risk judgment statement information, the corresponding risk feature information is extracted from the original material information. Based on the risk feature information, a risk feature vector is constructed and input into the trained risk level assessment model to obtain the predicted risk level. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain the risk conclusion consistency verification result. Based on the semantic consistency score and the terminology consistency verification result, customer entity identification information is extracted from the customer pre-loan due diligence information, and the customer entity identification information is matched with the subject identification information in the original material information to obtain the entity identification consistency verification result. The method further includes; The semantic consistency score, the terminology consistency verification result, the numerical consistency verification result, the risk conclusion consistency verification result, and the entity recognition consistency verification result are summarized to obtain a multidimensional consistency index set; The multidimensional consistency index set is input into the accuracy evaluation model to obtain the accuracy score result; Based on the accuracy score results, the accuracy of the customer's pre-loan due diligence information is determined, and accuracy assessment information is obtained; Based on the accuracy assessment information, corresponding correction suggestions are generated.

2. The method for assessing the accuracy of bank reports according to claim 1, characterized in that, The step involves using the BERTScore model to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information, generating embedding vectors for the original material information and due diligence information, including: The BERTScore model performs multi-layer encoding on the original material information to obtain an original material information embedding representation sequence containing context-dependent features. Vector aggregation processing is then performed on the original material information embedding representation sequence to generate the original material information embedding vector. The BERTScore model performs multi-layer encoding on the customer's pre-loan due diligence information to obtain a due diligence information embedding representation sequence. The vector aggregation process is then performed on the due diligence information embedding representation sequence to generate the due diligence information embedding vector.

3. The method for assessing the accuracy of bank reports according to claim 1, characterized in that, The step of calculating the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and performing greedy matching to obtain a semantic consistency score includes: The original material information embedding vector is divided into a first sub-vector set, and the due diligence information embedding vector is divided into a second sub-vector set; The first sub-vector set and the second sub-vector set are combined, and the cosine similarity calculation is performed between each pair of sub-vectors to generate a similarity score matrix; Based on the similarity score matrix, the greedy matching is used to select, in each round of matching, sub-vector pairs with high similarity scores from the first sub-vector set that has not yet participated in matching and the second sub-vector set that has not yet participated in matching; The subvector pairs are marked as matched, and the subvector pairs are added to the semantic matching set; After all the sub-vector pairs have been selected, the similarity scores of the sub-vector pairs in the semantic matching set are averaged to generate the semantic consistency score.

4. The method for assessing the accuracy of bank reports according to claim 1, characterized in that, The step involves obtaining financial terminology information from the customer's pre-loan due diligence information and using the PXL-BLEU algorithm combined with standard terms from a pre-defined terminology standardization library to perform a standardization comparison of the financial terminology information, obtaining terminology consistency verification results, including: Based on the aforementioned financial terminology information, a set of terms to be compared is constructed; Using the PXL-BLEU algorithm, the terms to be compared in the term set to be compared are compared with the standard terms in the preset term specification library to calculate the similarity score, and the corresponding term matching score set is obtained. Based on the term matching score set, determine whether each term to be compared has a matching relationship with any standard term in the term standard library with a matching score higher than a preset similarity threshold. If it exists, mark the term to be compared as a standard matching state; otherwise, mark the term to be compared as an unmatched state. The terms to be compared corresponding to the unmatched states are summarized to obtain a set of term replacement candidates; The PXL-BLEU algorithm is used to calculate the similarity score between the term to be compared in the candidate term replacement set and the standard term, and update the matching status of the corresponding term to be compared to obtain the updated matching status. The updated matching status is merged with the standard matching status to obtain the final matching status information; Based on the final matching status information, the term matching accuracy is calculated, and based on the term matching accuracy, the term consistency verification result is generated.

5. The method for assessing the accuracy of bank reports according to claim 1, characterized in that, The step of constructing a risk feature vector based on the risk feature information and inputting the risk feature vector into a trained risk level assessment model to obtain a predicted risk level includes: The trained risk level assessment model models the correlation between risk factors in the risk feature vector based on the attention mechanism, and performs weighted processing in combination with the time evolution trend of the risk factors. After completing the modeling of the risk feature vector, it outputs an overall risk level score. According to the preset risk level classification rules, the overall risk level score is mapped to the predicted risk level.

6. The method for assessing the accuracy of bank reports according to claim 1, characterized in that, The step of matching the customer entity identification information with the entity identification information in the original material information to obtain the entity identification consistency verification result includes: Obtain target entity identification information from the customer entity identification information, and obtain reference entity identification information from the original material information; Using a field-level joint matching strategy, field alignment and similarity calculation are performed on the target entity identifier information and the reference entity identifier information to generate an entity matching score matrix; By performing a threshold judgment operation on the entity matching score matrix, the matching relationship status between the target entity identification information and the reference entity identification information is determined, and the entity recognition consistency verification result is obtained.

7. A device for assessing the accuracy of bank reports, characterized in that, The device includes: The original material information acquisition module is used to acquire the original material information of the target bank's customers; The due diligence generation module is used to obtain customer pre-loan due diligence information corresponding to the original material information using an artificial intelligence model; The semantic embedding processing module is used to perform contextual semantic embedding processing on the original material information and the customer pre-loan due diligence information using the BERTScore model, and generate the original material information embedding vector and the due diligence information embedding vector. The semantic consistency calculation module is used to calculate the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and perform greedy matching to obtain a semantic consistency score. The terminology consistency verification module is used to obtain financial terminology information from the customer's pre-loan due diligence information if the semantic consistency score reaches the preset credibility condition, and to perform a normative comparison of the financial terminology information using the PXL-BLEU algorithm combined with standard terms in the preset terminology standard library to obtain the terminology consistency verification result. The numerical consistency verification module is used to identify financial indicator information from the terminology consistency verification results, extract original information for calculating the financial indicator information from the original material information based on the financial indicator information, input the original information into the preset indicator calculation formula, generate corresponding reference values, obtain indicator description values ​​from the customer pre-loan due diligence information, compare the reference values ​​with the description values, and obtain numerical consistency verification results. The risk conclusion consistency verification module is used to identify risk judgment statements related to abnormal financial indicators in the customer's pre-loan due diligence information if abnormal financial indicator information exists in the numerical consistency verification results. Based on the risk judgment statements, the module extracts corresponding risk feature information from the original material information, constructs a risk feature vector based on the risk feature information, and inputs the risk feature vector into a trained risk level assessment model to obtain a predicted risk level. The module then compares the predicted risk level with the risk conclusions in the due diligence information to obtain a risk conclusion consistency verification result. The entity recognition consistency verification module is used to extract customer entity identification information from the customer pre-loan due diligence information based on the semantic consistency score and the terminology consistency verification result, and match the customer entity identification information with the subject identification information in the original material information to obtain the entity recognition consistency verification result. The device further includes; The semantic consistency score, the terminology consistency verification result, the numerical consistency verification result, the risk conclusion consistency verification result, and the entity recognition consistency verification result are summarized to obtain a multidimensional consistency index set; The multidimensional consistency index set is input into the accuracy evaluation model to obtain the accuracy score result; Based on the accuracy score results, the accuracy of the customer's pre-loan due diligence information is determined, and accuracy assessment information is obtained; Based on the accuracy assessment information, corresponding correction suggestions are generated.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of a method for assessing the accuracy of bank reports as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of a method for assessing the accuracy of bank reports as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Personal credit risk evaluation method and system

    CN107993140A

  • Clinical examination result auditing method and system based on artificial intelligence and big data

    CN118629571A