Bank report accuracy evaluation method and device, equipment and medium

Abstract

The invention relates to the technical field of financial information evaluation. The bank report accuracy evaluation method comprises the steps that context semantic embedding processing is carried out on original material information and customer pre-loan full-duty investigation information through a BERTScore model, and then the original material information and the customer pre-loan full-duty investigation information are obtained; and performing cosine similarity calculation on the original material information embedded vector and the diliary investigation information embedded vector, executing greedy matching to obtain a semantic consistency score, adopting a PXL-BLEU algorithm to be combined with standard terms in a preset term specification library to obtain a term consistency verification result, and inputting the original information into a preset index calculation formula to obtain a term consistency verification result. And generating a corresponding reference value, constructing a risk feature vector, inputting the risk feature vector into the trained risk level evaluation 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. The method has the effect of improving the accuracy and consistency of report auditing.

Description

Technical Field

[0001] The present invention relates to the technical field of financial information evaluation, and in particular to a method, device, equipment and medium for evaluating the accuracy of bank reports. Background Art

[0002] Traditionally, manual review is often used to ensure the accuracy of bank reports. Reviewers carefully review the client's original documentation and pre-loan due diligence information, drawing on their expertise and experience to compare and assess each, including semantics, terminology, numerical accuracy, and risk assessment. Simple rule engines are also used to assist with review, checking specific content within the report according to pre-set rules. For example, they can verify the completeness of certain fixed-format information and whether specific numerical ranges meet requirements.

[0003] However, these existing audit methods have obvious flaws. Manual audits are not only time-consuming and labor-intensive, but also prone to omissions and misjudgments due to subjective factors and individual differences among auditors, making it difficult to guarantee the accuracy of audit results. Summary of the Invention

[0004] In order to improve the accuracy and consistency of report audits, the present application provides a method, apparatus, device and medium for evaluating the accuracy of bank reports.

[0005] The above-mentioned invention objective of this application is achieved through the following technical solutions: A method for evaluating the accuracy of bank reports, comprising: Obtain original material information of target bank customers; Utilize artificial intelligence models to obtain pre-loan due diligence information on customers corresponding to the source material information; Using the BERTScore model, contextual semantic embedding is performed on the original material information and the customer's pre-loan due diligence information to generate an original material information embedding vector and a due diligence information embedding vector; Calculate 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; If the semantic consistency score meets 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 standard terms in a preset terminology standard library to perform a normative comparison on the financial terminology information to obtain a terminology consistency verification result; The financial indicator information identified from the terminology consistency verification result is extracted from the original material information based on the financial indicator information, and the original information used for calculating the financial indicator information is input into a preset indicator calculation formula to generate a corresponding reference value. The indicator description value is obtained from the customer pre-loan due diligence information, and the reference value is compared with the description value to obtain a numerical consistency verification result. If abnormal financial indicator information exists in the numerical consistency verification result, then identifying risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information, extracting corresponding risk feature information from the original material information based on the risk judgment statement information, 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. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain a risk conclusion consistency verification result; Based on the semantic consistency score and the term 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 an entity identification consistency verification result.

[0006] By adopting the above technical solutions, artificial intelligence models and semantic modeling technology 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 comparison at the semantic level is achieved, and the PXL-BLEU algorithm is used to perform normative matching of financial terms. At the same time, the numerical consistency analysis of financial data is performed in combination with the preset indicator calculation formula, and the consistency of risk conclusions is verified with the help of a trained risk level assessment model. On this basis, the accuracy of the subject information is further guaranteed through entity recognition consistency verification. Finally, the verification results of each dimension are summarized as accuracy scores and correction suggestions. It completely replaces the traditional manual item-by-item review and rule engine-assisted judgment methods, effectively improving review efficiency, reducing labor costs, reducing the risk of misjudgment and omissions, and enhancing the accuracy of the bank report review process.

[0007] In a preferred example, the present application may be further configured as follows: using the BERTScore model to perform contextual semantic embedding processing on the original material information and the customer's pre-loan due diligence information to generate an original material information embedding vector and a due diligence information embedding vector, 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, and performs vector aggregation processing 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, and performs the vector aggregation process 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, effectively capturing the deep semantic relationships in the text, avoiding semantic misjudgments caused by language expression differences, ambiguous word meanings, or context fragmentation in manual review, and providing a structurally unified, dimensionally corresponding, and semantically rich vector foundation for subsequent semantic consistency calculations, thereby improving the automation and accuracy of semantic comparison and enhancing the intelligence and consistency of the evaluation process.

[0009] In a preferred example, the present application may be further configured as follows: performing cosine similarity calculation on the original material information embedding vector and the due diligence information embedding vector and performing greedy matching to obtain a semantic consistency score, including: Dividing the original material information embedding vector into a first sub-vector set, and dividing the due diligence information embedding vector into a second sub-vector set; Combining the first sub-vector set and the second sub-vector set, performing the cosine similarity calculation between each pair of sub-vectors, and generating a similarity score matrix; Based on the similarity score matrix, adopting the greedy matching, in each matching round, selecting a sub-vector pair with a high similarity score from the first sub-vector set that has not participated in the matching and the second sub-vector set that has not participated in the matching; Marking the sub-vector pair as matched, and adding the sub-vector pair to a semantic matching set; After all the sub-vector pairs are selected, an average calculation is performed on the similarity scores of the sub-vector pairs in the semantic matching set to generate the semantic consistency score.

[0010] By adopting the above technical solution, it is possible to substructure the embedded vectors of the original material information and the customer's pre-loan due diligence information, and by constructing a similarity score matrix for each sub-vector combination, combined with a greedy matching strategy to dynamically screen sub-vector pairs with highly corresponding semantics, effectively achieve fine-grained semantic hierarchical matching, avoiding the problem of difficulty in aligning key content due to large amounts of information and wide semantic spans in manual review. At the same time, by averaging the similarity scores of the matching sub-vector pairs, the degree of semantic consistency between the original material and the due diligence information is quantified, thereby improving the automated accuracy and objectivity of the semantic comparison and providing a reliable scoring basis for subsequent accuracy assessment.

[0011] In a preferred example, the present application may be further configured as follows: obtaining financial terminology information from the customer's pre-loan due diligence information, and using the PXL-BLEU algorithm in combination with standard terms in a preset terminology standard library to perform a standard comparison on the financial terminology information to obtain a terminology consistency verification result, including: Constructing a set of terms to be compared based on the financial term information; Using the PXL-BLEU algorithm, similarity scores are calculated for the terms to be compared in the set of terms to be compared and the standard terms in the preset term specification library to obtain a corresponding term matching score set; Based on the term matching score set, it is determined whether each of the to-be-compared terms has a matching relationship with any of the standard terms in the term specification library with a matching score higher than a preset similarity threshold; if so, the to-be-compared term is marked as a canonical matching state; if not, the to-be-compared term is marked as an unmatched state; Summarizing the to-be-compared terms corresponding to the unmatched states to obtain a set of term replacement candidates; Calculating similarity scores between the terms to be compared in the term replacement candidate set and the standard term using the PXL-BLEU algorithm, and updating the matching status corresponding to the terms to be compared to obtain an updated matching status; Merging the updated matching status with the standard matching status to obtain final matching status information; The term matching accuracy is calculated according to the final matching status information, and the term consistency verification result is generated based on the term matching accuracy.

[0012] By adopting the above technical solution, the financial terminology information in the customer's pre-loan due diligence information can be structured and organized into a set of terms to be compared. By introducing the PXL-BLEU algorithm, multiple rounds of similarity scoring calculations are performed on each term to be compared with the standard terms in the preset term specification library. Combined with the preset similarity threshold, it is determined whether there is a standard matching relationship between the terms, and term replacement candidates are further generated for unmatched terms and secondary scoring corrections are performed, effectively improving the accuracy and coverage of term matching. At the same time, by calculating the term matching accuracy in the final matching state, the compliance of the term expression is quantified, avoiding the problem of misjudgment caused by differences in term expression and insufficient industry experience in manual review, thereby realizing the automation and objectivity of the term standard comparison process.

[0013] In a preferred example, the present application may 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 a preset risk level classification rule, 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 vectors extracted from the original material information based on the trained risk level assessment model, use the attention mechanism to dynamically capture the correlation between different risk factors, and perform weight adjustment in combination with the time evolution trend of the 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 misjudgments caused by the lack of systematic modeling of complex risk factors in manual assessments.

[0015] In a preferred example, the present application may be further configured as follows: matching the client entity identification information with the subject identification information in the original material information to obtain an entity identification consistency verification result includes: Acquire target entity identification information from the client entity identification information, and acquire 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 identification information and the reference entity identification 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, it is possible to extract entity fields with identification functions from the customer's pre-loan due diligence information and original material information respectively, and adopt a field-level joint matching strategy to perform field standardization, alignment and similarity calculation on the target entity identification information and the reference entity identification information, and construct an entity matching score matrix. On this basis, by setting a threshold to compare the score strength, it is determined whether there is a consistent matching relationship between the two sets of entity identification information, which effectively solves the problem of subject misjudgment caused by differences in field naming, inconsistent expression methods or information omissions in traditional manual recognition, realizes consistency verification and automatic 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 example, the present application may be further configured as follows: the method for evaluating the accuracy of bank reports further includes: Summarizing the semantic consistency score, the term consistency verification result, the numerical consistency verification result, the risk conclusion consistency verification result, and the entity recognition consistency verification result to obtain a multi-dimensional consistency indicator set; Inputting the multidimensional consistency index set into the accuracy assessment model to obtain an accuracy score result; According to the accuracy scoring results, the accuracy of the customer's pre-loan due diligence information is determined to obtain accuracy assessment information Based on the accuracy evaluation information, corresponding correction suggestion information is generated.

[0018] By adopting the above technical solution, the semantic consistency score, terminology consistency verification results, numerical consistency verification results, risk conclusion consistency verification results and entity identification consistency verification results can be summarized in multiple dimensions to construct a unified consistency indicator set. This indicator set can be input into the accuracy assessment model for comprehensive scoring, thereby quantitatively determining the overall accuracy of the customer's pre-loan due diligence information. On this basis, structured accuracy assessment information can be further generated based on the scoring results, and targeted correction suggestion information can be automatically generated in combination with various sources of inconsistency, realizing a closed-loop processing logic from verification, scoring to feedback, effectively reducing the risk of subjective intervention in manual judgment, improving the systematicity, 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 object of the present invention is achieved through the following technical solutions: A device for evaluating the accuracy of bank reports, comprising: The original material information acquisition module is used to obtain the original material information of the target bank customers; A due diligence generation module, configured to utilize an artificial intelligence model to obtain pre-loan due diligence information on the customer corresponding to the source material information; A semantic embedding processing module, configured to perform contextual semantic embedding processing on the source material information and the customer pre-loan due diligence information using a BERTScore model, to generate an embedding vector for the source material information and an embedding vector for the due diligence information; a semantic consistency calculation module, configured to calculate 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; A terminology consistency verification module is configured to obtain financial terminology information from the customer's pre-loan due diligence information if the semantic consistency score meets a preset credibility condition, and perform a normative comparison on the financial terminology information using the PXL-BLEU algorithm combined with standard terminology in a preset terminology standard library to obtain a terminology consistency verification result; a numerical consistency verification module configured to extract, from the original material information, original information used for calculating the financial indicator information identified from the terminology consistency verification result, the financial indicator information, based on the financial indicator information, input the original information into a preset indicator calculation formula to generate a corresponding reference value, obtain an indicator description value from the customer's pre-loan due diligence information, compare the reference value with the description value, and obtain a numerical consistency verification result; a risk conclusion consistency verification module, configured to, if abnormal financial indicator information is present in the numerical consistency verification result, identify risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information, extract corresponding risk feature information from the original material information based on the risk judgment statement information, construct a risk feature vector based on the risk feature information, input the risk feature vector into a trained risk level assessment model to obtain a predicted risk level, compare the predicted risk level with the risk conclusion in the due diligence information, and obtain a risk conclusion consistency verification result; An 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 term consistency verification result, match the customer entity identification information with the subject identification information in the original material information, and obtain an entity recognition consistency verification result.

[0020] By adopting the above technical solutions, artificial intelligence models and semantic modeling technology 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 comparison at the semantic level is achieved, and the PXL-BLEU algorithm is used to perform normative matching of financial terms. At the same time, the numerical consistency analysis of financial data is performed in combination with the preset indicator calculation formula, and the consistency of risk conclusions is verified with the help of a trained risk level assessment model. On this basis, the accuracy of the subject information is further guaranteed through entity recognition consistency verification. Finally, the verification results of each dimension are summarized as accuracy scores and correction suggestions. It completely replaces the traditional manual item-by-item review and rule engine-assisted judgment methods, effectively improving review efficiency, reducing labor costs, reducing the risk of misjudgment and omissions, and enhancing the accuracy of the bank report review process.

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

[0022] The fourth objective of this application is achieved through the following technical solutions: A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for evaluating 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 realizes automatic comparison at the semantic level. It uses the PXL-BLEU algorithm to perform normative matching of financial terms. At the same time, it combines the preset indicator calculation formula to conduct numerical consistency analysis of financial data. It uses the 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 accuracy scores and correction suggestions. It completely replaces the traditional manual item-by-item review and rule engine-assisted judgment methods, effectively improving review efficiency, reducing labor costs, reducing the risk of misjudgment and omissions, and enhancing the accuracy of the bank report review process. 2. It can structure the financial terminology information in the customer's pre-loan due diligence information into a set of terms to be compared. By introducing the PXL-BLEU algorithm, it performs multiple rounds of similarity scoring calculations on each term to be compared with the standard terms in the preset term specification library. It then uses the preset similarity threshold to determine whether the terms have a standard matching relationship. For unmatched terms, it further generates term replacement candidates and performs secondary scoring corrections, effectively improving the accuracy and coverage of term matching. At the same time, by calculating the term matching accuracy rate in the final matching state, it quantifies the compliance of term expression and avoids misjudgments caused by differences in term expression and insufficient industry experience during manual review, thereby achieving automation and objectivity in the term standard comparison process. 3. It can summarize the semantic consistency scores, terminology consistency verification results, numerical consistency verification results, risk conclusion consistency verification results and entity identification consistency verification results in multiple dimensions to build a unified consistency indicator set, and input this indicator set into the accuracy assessment model for comprehensive scoring, so as to quantitatively determine the overall accuracy of the customer's pre-loan due diligence information. On this basis, it further generates structured accuracy assessment information based on the scoring results, and automatically generates targeted correction suggestion information based on various sources of inconsistency, realizing a closed-loop processing logic from verification, scoring to feedback, effectively reducing the risk of subjective intervention in manual judgment, improving the systematicity, transparency and interpretability of the assessment results, and enhancing the intelligence level and correction efficiency of the bank's report review process. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a schematic diagram of a method for evaluating the accuracy of bank reports in an embodiment of the application; Figure 2 This is a flowchart for implementing step S30 in a method for evaluating the accuracy of bank reports in one embodiment of the present application; Figure 3 This is a flowchart for implementing step S40 in a method for evaluating the accuracy of bank reports in one embodiment of the present application; Figure 4 This is a flowchart for implementing step S50 in a method for evaluating the accuracy of bank reports in one embodiment of the present application; Figure 5 This is a flowchart for implementing step S70 in a method for evaluating the accuracy of bank reports in one embodiment of the present application; Figure 6 This is a flowchart for implementing step S80 in a method for evaluating the accuracy of bank reports in one embodiment of the present application; Figure 7 This is a flowchart for implementing the method after step S80 in the bank report accuracy assessment method in one embodiment of the present application; Figure 8 This is a principle block diagram of a device for evaluating the accuracy of bank reports in one embodiment of the present application; Figure 9 It is a schematic diagram of a device in one embodiment of the present application. DETAILED DESCRIPTION

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

[0026] In one embodiment, if Figure 1 As shown, the present application discloses a method for evaluating the accuracy of bank reports, which specifically includes the following steps: S10: Obtain original material information of target bank customers.

[0027] Specifically, it is possible to receive a loan application request sent from a bank business processing system or a third-party credit service platform, which carries at least information fields such as a customer unique identifier, a project number or a project application form number, and connect to data sources such as a project database, an enterprise knowledge base, and a financial information center interface. The interface is called with the customer identifier or project number as a query index to obtain the associated original material information data. The original material information generally includes the project name, industry, estimated investment amount, proposed financing amount, implementing entity, project background, fund use description, historical operating data, project financial structure, cooperation unit information, and project cycle plan. After the acquisition is completed, the original material information can be preliminarily packaged in JSON structure or XML format and stored in a processing buffer to obtain original material information with a unified format structure and field label definition.

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

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

[0030] Specifically, the original material information is input into the artificial intelligence model according to field classification. The original material information includes corporate financial statements, credit records, project description documents and industrial and commercial registration information in different formats. Before input, the structured data 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 the sentence and word vector model. All of the above content is then combined into a unified data representation form and sent to the artificial intelligence model. The artificial intelligence model completes the syntactic organization and paragraph arrangement through the trained hierarchical content generation structure, and outputs due diligence sentence fragments with complete descriptions, clear structure and semantic compliance through the natural language generation module. Each sentence fragment is spliced ​​into a complete report text covering structural parts such as project overview, operating conditions, financial indicators, compliance information and risk warnings according to the preset template, so as to obtain customer pre-loan due diligence information corresponding to the semantics of the original material information.

[0031] S30: Use the BERTScore model to perform contextual semantic embedding on the original material information and the customer's pre-loan due diligence information, generating an embedding vector for the original material information and an embedding vector for the due diligence information. In this embodiment, the BERTScore model refers to a deep language understanding model used for text semantic embedding and similarity analysis. It is built based on the BERT structure, can contextually encode input sentences and generate word-level vector representations, supports semantic alignment and precise similarity scoring between two text fragments, and is used for context embedding and vector similarity calculations in the semantic consistency scoring process.

[0032] Specifically, sentence segmentation operations can be performed on the original material information, and the original material information can be organized into a sequence of semantic units. Word segmentation, word vector retrieval and encoding can be performed on each semantic unit. The encoded semantic unit vector sequence is input into the pre-trained BERTScore model for context modeling. The BERTScore model models the context dependency relationship 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 processing is performed. The vector aggregation processing includes performing average pooling on the embedding representations of all semantic units or selecting representative position vectors to form a unified vector to complete the construction of the original material information embedding vector. The same processing process is applied to the customer's pre-loan due diligence information. The semantic unit segmentation, encoding, modeling and aggregation are completed for the customer's pre-loan due diligence information, and the customer's pre-loan due diligence information embedding vector is constructed to obtain the original material information embedding vector and the customer's 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 building a semantic matching relationship, gradually selects the pair of unmatched sub-vectors with the highest score based on the similarity score matrix as the matching result of the current round, immediately marks the sub-vector pair as matched, and then continues to execute the same strategy from the remaining unmatched sub-vectors until all matching sub-vectors are processed. This algorithm does not backtrack 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 to split the original material information embedding vector into a first sub-vector set. At the same time, 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, and a cosine similarity calculation operation is performed between each pair of sub-vectors. The cosine similarity calculation measures the degree of similarity by calculating the cosine value of the angle between the sub-vectors, and the similarity scores of all sub-vector pairs are combined into a similarity score matrix. After the score matrix is ​​constructed, a greedy matching operation is performed. The greedy matching operation selects the sub-vector pair corresponding to the current maximum unmatched score from the score matrix and records the sub-vector pair as a semantic matching pair. The process is repeated until all sub-vectors in the first sub-vector set and the second sub-vector set are matched. Each round of matching ensures that the matched sub-vectors are not repeatedly selected. Finally, the similarity scores in all semantic matching pairs are weighted or equally weighted averaged to obtain a semantic consistency score.

[0036] S50: If the semantic consistency score meets 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 on the financial terminology information to obtain the terminology consistency verification result.

[0037] In this implementation, the preset credibility condition refers to the numerical threshold used to determine the reliability of the semantic consistency score, typically expressed as a floating-point number between 0 and 1. It is used to filter content with high semantic similarity to trigger subsequent terminology verification and consistency processing. Financial terminology information refers to financial-specific terminology extracted from pre-loan due diligence information. It includes financial terms, risk terms, indicator names, reporting expressions, and specific financial expression structures. Common examples include "asset-liability ratio," "capital adequacy ratio," and "project payback period." The PXL-BLEU algorithm is a term similarity calculation method that combines traditional BLEU scoring with a position-aware mechanism. It uses a comprehensive calculation based on multiple dimensions, including n-gram overlap, semantic embedding similarity, and a position offset penalty factor, to measure the degree of match between candidate terms and standard terms. Standard terms are standardized terminology items stored in a terminology standard library. They feature clear semantics, standard formatting, and industry-wide usage. They serve as a comparison benchmark to determine the compliance of terminology in client reports, such as "return on equity" and "short-term debt-paying ability."

[0038] Specifically, a term extraction operation can be performed on the customer's pre-loan due diligence information when the semantic consistency score meets the preset credibility conditions. The term extraction operation includes the joint processing of high-frequency matching recognition based on the financial field dictionary and the proper noun recognition algorithm based on the part-of-speech tagging model. The identified term set is sorted into a set of terms to be compared. Each term in the set of terms to be compared is used as an input item to perform similarity calculation 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 the term pairs and constructs word fragments The structure is aligned, and then the word sense vector is combined to perform sentence sense fusion and output the similarity score. After obtaining the term matching score set, each term is judged whether it is matched successfully according to the 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 aggregated to form a set of alternative candidate terms. The terms in the alternative candidate term set are then compared with the standard terms to perform PXL-BLEU scoring and the corresponding matching status is updated. Finally, the matching status of all terms is combined to form a complete term state mapping, and the term matching accuracy is calculated based on the mapping to obtain the term consistency verification result.

[0039] S60: Based on the financial indicator information identified from the term consistency verification results, extract the original information used for calculating the financial indicator information from the original material information, input the original information into the preset indicator calculation formula, 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, and obtain the numerical consistency verification result.

[0040] In this embodiment, the preset indicator calculation formula refers to a 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" and so on.

[0041] Specifically, based on the set of terms that have been confirmed to have semantic matching relationships in the term consistency verification results, terms with financial meaning and stable structure are screened as the set of financial indicator names. The mapping relationship between the field labels and the relevant field labels in the original material information is located item by item in the financial indicator name set. Based on the mapping relationship, the original numerical information used 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 cash flow from operating activities. The above original numerical values ​​are input into the preset indicator calculation formula according to the calculation logic specified in the indicator definition. The indicator calculation formula includes ratio, growth, and composite indicator calculation methods specified by the financial department or industry. After the calculation operation is performed, a reference value corresponding to each indicator item is generated. The description position of the financial indicator item is then located from the customer's pre-loan due diligence information and the indicator expression value corresponding to the position is extracted. A numerical comparison operation is performed between the reference value and the expression value. The comparison operation includes three steps: floating threshold control, unit normalization, and decimal precision alignment. The comparison result is used to determine whether the numerical values ​​are consistent to obtain the numerical consistency verification result.

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

[0043] In this city's case, risk assessment statements refer to natural language sentences with risk-expressing implications identified from pre-loan due diligence information on customers. These sentences typically include expressions expressing judgments, speculations, tendencies, or conclusions, such as "there is insufficient operating cash flow" or "debt-paying ability may decline." A trained risk rating assessment model is a deep learning model trained based on historical multi-source risk data. It receives risk feature vectors and outputs predicted risk rating labels. The model's internal structure includes an attention mechanism or a time evolution trend modeling module, supporting the scoring fusion of multi-dimensional risk factors.

[0044] Specifically, financial indicator items with differences exceeding the preset deviation threshold are identified in the numerical consistency verification results, and each abnormal financial indicator item is input as a keyword into the sentence set of the customer's pre-loan due diligence information to perform keyword matching and syntactic dependency recognition operations, and sentences containing risk judgment meanings are located from the syntactic structure. The sentence usually includes expression features such as "may lead to", "existence of risks", "liquidity tension", and "decline in debt repayment ability". These sentences are organized into a risk judgment sentence information set, and risk clue words that can be used for reasoning are extracted from each risk judgment sentence. The risk clue words are used to perform cross-segment query and field comparison on the original material information, and the risk clue words directly related to the risk theme are extracted from the original material information. Associated quantitative or qualitative data, including asset change trends, liability structure, short-term debt repayment sources, historical default records and credit rating results, are constructed into a risk feature set after field standardization and numerical normalization. The risk feature set is input into the vector generation module for encoding to generate a risk feature vector with unified dimensions. The risk feature vector is input into the trained risk grade assessment model. The risk grade assessment model analyzes the combination relationship between various risk factors through the attention weight mechanism and completes the grade score output. After the score result is mapped to the risk grade label, it is compared item by item with the original risk conclusion label in the customer's pre-loan due diligence information to obtain the risk conclusion consistency verification result.

[0045] S80: Based on the semantic consistency score and term consistency verification results, extract the 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 the set of key fields extracted from the customer's pre-loan due diligence information that identify the customer, including the company name, project name, legal representative, unified social credit code, registered address, etc., which can be used for subject comparison with the original materials. Subject identification information refers to the set of fields extracted from the original material information that have legal or business identification validity. Structurally, it corresponds to the customer entity identification information and is used to establish a subject consistency verification relationship with the customer's pre-loan due diligence information.

[0047] Specifically, the scope of content with business credibility in the text is determined based on the semantic consistency score and terminology consistency verification results. Named entity recognition is then performed on the customer pre-loan due diligence information within this credibility range. The named entity recognition model decodes the sentence label sequence based on financial domain word vectors and context recognition rules, extracting information fields representing customer 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 person name, registered address, unified social credit code, and contact information. After extraction, format standardization is performed. The standardization process includes removing spaces, unifying uppercase and lowercase letters, replacing common abbreviations, and cleaning punctuation. The corresponding subject identification fields are simultaneously extracted from the original material information and formatted in the same way. The two field sets are then input into the field-level joint matching module for field alignment. Field alignment adopts a sequence-independent full-data comparison strategy. For each pair of field combinations, semantic similarity calculation based on word vectors and edit distance calculation based on strings are performed. The calculation results are weighted and fused to form an entity matching score matrix. The entity matching score matrix is ​​then subjected to a scoring threshold determination operation. The matching status of all field levels is output and a comprehensive judgment is performed to obtain the entity recognition consistency verification result.

[0048] In one embodiment, if Figure 2 As shown, in step S30, the original material information and the customer's pre-loan due diligence information are respectively subjected to contextual semantic embedding processing using the BERTScore model to generate an embedding vector for the original material information and an embedding vector for the due diligence information, including: S301: 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, and performs vector aggregation processing on the original material information embedding representation sequence to generate an original material information embedding vector.

[0049] Specifically, the original material information is divided according to sentence boundaries and segmented according to natural language processing rules. The segmented results are input into the BERTScore model. The BERTScore model processes the semantic dependency between each word and its context in turn through a multi-layer Transformer encoding structure. The attention mechanism in the model scores the association between each word and the words in other positions in the sentence and generates a weighted combined context representation to form an embedded representation sequence of the original material information. After obtaining the embedded representation sequence, vector aggregation processing is performed. The aggregation processing includes performing an average pooling operation on all word representations or selecting the CLS vector as the sentence-level representation method. 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 the customer's pre-loan due diligence information to obtain a due diligence information embedding representation sequence, and performs vector aggregation processing on the due diligence information embedding representation sequence to generate a due diligence information embedding vector.

[0051] Specifically, the customer's pre-loan due diligence information is segmented at the sentence level and the word segmentation encoding process is completed based on the language model vocabulary. The encoded word sequence is input into the BERTScore model. The BERTScore model extracts the semantic relationship of each word in the context in turn through a multi-layer Transformer structure and establishes a global dependency representation. The intermediate representation of each layer of encoding output continues to iterate semantic aggregation and feature propagation in the next layer until the entire semantic modeling process is completed. The embedded representation sequence output by the model is composed of the contextual semantic representation of each word. After obtaining the complete representation sequence, vector aggregation processing is performed on all word representations. The vector aggregation processing includes performing an average pooling operation on the representation sequence to calculate the sentence-level embedding vector, or selecting the semantic representation of the CLS tag position as the global vector expression of the sentence. After the aggregation is completed, the due diligence information embedding vector is output. In one embodiment, if Figure 3 As shown, in step S40, the cosine similarity calculation is performed on the original material information embedding vector and the due diligence information embedding vector and 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 division 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 division 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 sequentially extracted along the vector dimension direction to form multiple sub-vectors to obtain a first sub-vector set. Similarly, window division is performed on the due diligence information embedding vector according to the same dimensional rule to generate a 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, pairwise combination operations are performed on the first sub-vector set and the second sub-vector set, where 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 calculation is performed on each pair of sub-vectors. The cosine similarity calculation calculates the dot product of the two sub-vectors after L2 norm normalization. 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 in the order of the combination of the first sub-vector set and the second sub-vector set. Each row of the matrix corresponds to a sub-vector in the first sub-vector set, and each column corresponds to a sub-vector in the second sub-vector set. Each element of the matrix is ​​the similarity score between the corresponding row and column sub-vector pairs to construct a complete similarity score matrix.

[0054] S403: Based on the similarity score matrix, greedy matching is adopted. In each matching round, a sub-vector pair with a high similarity score is selected from the first sub-vector set that has not participated in the matching and the second sub-vector set that has not participated in the matching.

[0055] Specifically, all candidate sub-vector pairs and their corresponding similarity scores are extracted from the similarity score matrix. Before each round of matching, two unmatched marker sets are constructed, and the index positions of the sub-vectors that have not been selected in the first sub-vector set and the second sub-vector set are marked respectively. The sub-vector pairs at the corresponding positions in the current two unmatched marker sets and their similarity scores are screened out from the similarity score matrix. The sub-vector pair with the highest similarity score is selected as the matching result of the current round. The corresponding two index positions in the sub-vector pair are removed from the unmatched marker set and added to the matching marker set respectively. After updating, the next round of matching is continued. The above operation is repeated until all sub-vectors in any sub-vector set have participated in the matching.

[0056] S404: Mark the sub-vector pair as matched, and add the sub-vector pair to the semantic matching set; In this embodiment, the semantic matching set refers to a set that records all sub-vector pairs marked as having high similarity during the cosine similarity calculation and greedy matching of the embedded vectors, and is used for statistical semantic consistency scoring. It is an intermediate structure of the similarity calculation output.

[0057] Specifically, after completing the similarity score screening of the current pair of sub-vectors, the index position of the sub-vector pair in the first sub-vector set and the second sub-vector set are recorded respectively, the index position is marked as occupied in the corresponding unmatched mark set, and the matching state mapping table is updated synchronously. After the marking is completed, the index position pair of the sub-vector pair and its corresponding similarity score form a matching pair structure, and the matching pair structure is added to the semantic matching set.

[0058] S405: After all sub-vector pairs are selected, average calculation is performed 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 of all sub-vector pairs, the semantic matching set is traversed, and the similarity score values ​​in each matching pair structure are read one by one. All score values ​​are stored in the scoring vector list in the original order. A numerical sum operation is performed on all elements in the scoring vector list to obtain the total similarity score, and the total number of elements in the scoring vector list is obtained as the number of matching pairs. The mean calculation operation is performed by dividing the total similarity score by the number of matching pairs. During the mean calculation process, the floating-point precision control can be set to unify the result format, and 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, if 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 combination with standard terms in the preset terminology standard library to perform a normative comparison on the financial terminology information to obtain terminology consistency verification results, including: S501: Construct a set of terms to be compared based on financial term information.

[0061] Specifically, based on the financial terminology information extracted from the customer's pre-loan due diligence information, the term collection is screened and sorted according to the naming conventions and domain part-of-speech characteristics of financial terms. First, the term entries with proprietary, indexical or categorization characteristics are identified through part-of-speech tagging and entity recognition algorithms. Then, deduplication, normalization and format standardization operations are performed on the term entries. The deduplication operation processes semantically equivalent expressions through exact string matching or synonym rule library. The normalization operation expands the abbreviation to the full name and unifies the term granularity. The format standardization operation includes unifying uppercase and lowercase letters, removing symbols, standardizing units, etc. Finally, the processed term entries are organized into a structured list, and each element in the list is a term to be compared, so as to construct a term collection to be compared for subsequent term standardization verification.

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

[0063] Specifically, each term to be compared in the set of terms to be compared is taken as an input item in turn, and a one-to-many similarity score calculation is performed with all the standard terms registered in the preset term specification library. In each scoring process, the n-gram sequence of the term to be compared and the standard term are first segmented 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. On the basis of the basic score, the position information weight coefficient of the PXL-BLEU algorithm is introduced to penalize the position offset. At the same time, the character components of the term and its semantic embedding representation vector are input into the semantic similarity evaluation module to further supplement the semantic layer correction. Finally, a weighted fusion is performed on multiple scoring factors to generate the final matching score between each term pair. The matching score is stored in the scoring record list together with the identifier of the term to be compared and the corresponding standard term. After completing the scoring item by item, all matching results are sorted out 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 standard library with a matching score higher than a preset similarity threshold. If so, mark the term to be compared as a standard matching state; if not, mark the term to be compared as an unmatched state.

[0065] Specifically, the matching records of each term to be compared in the term matching score set are traversed one by one. The matching records contain the similarity scoring results between the current term to be compared and all standard terms in the preset term specification library. For each set of scoring 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 divide semantic similarity and dissimilarity. 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 it is marked as an unmatched state.

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

[0067] Specifically, a status screening operation is performed on all the terms to be compared in the term status mapping structure. The screening condition is the term items whose matching status is marked as unmatched. All the terms to be compared that meet the condition are extracted and rearranged according to the original order or field priority. After the arrangement is completed, a preliminary set of alternative terms is formed. Each term in the preliminary set of alternative terms is further subjected to spelling similarity filtering, synonym expansion query and context semantic review processing. Spelling similarity filtering is used to identify similar terms caused by formatting or input errors. Synonym expansion query is based on the financial term vocabulary to retrieve equivalent expressions or industry aliases. Context semantic review ensures consistency in the term usage context by associating the context sentence content. After completing the above processing, a structured list is formed and stored as a set of term replacement candidates.

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

[0069] Specifically, each term to be compared is read from the set of term replacement candidates in turn, and the current term to be compared is used as input. A one-to-many similarity score calculation is performed 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 context nesting relationship of the term. The final similarity score is generated by modeling the spelling layer and semantic layer of the two terms respectively and weighted fusion. The score is compared with the preset similarity threshold. When the score value is higher than the threshold, the matching state of the term to be compared is updated from the unmatched state to the standard matching state. If it still does not reach the threshold, the original unmatched state is retained. After completing the re-scoring of all replacement candidate terms, a new state mapping structure is formed, and this structure is used as the updated matching state.

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

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

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

[0073] Specifically, all term records in the final matching status information are statistically processed item by item. 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 respectively. The number of standard matching terms is taken as the hit number, and the total number of terms to be compared is taken as the benchmark number. The term matching accuracy is obtained by dividing the hit number by the benchmark number. The term matching accuracy is expressed in decimal form and retains the specified number of digits of precision. After the accuracy calculation is completed, the value is mapped and judged with the set term consistency level standard. If the accuracy is higher than the high consistency threshold, it is judged as a 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, a low consistency level is output. The consistency level label and the accuracy value are written into the term consistency verification result together.

[0074] In one embodiment, if Figure 5 As shown, in step S70, based on the risk feature information, a risk feature vector is constructed, and the risk feature vector is input into the trained risk level assessment model to obtain a 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 based on the time evolution trend of the 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 that reflect the customer's risk status. These risk factors include financial indicator volatility, debt repayment ability indicator sequence, credit rating historical distribution, default record statistics, and liquidity status scores within the company's operating cycle. The risk level assessment model uses a multi-head attention mechanism to model the semantic association relationship between all risk factors, and measures the interaction intensity between factors through the attention weight matrix. The time evolution trend parameter is further introduced to adjust the weight structure during the modeling process, so that more attention is paid to risk factors with consistent historical change directions or significant trends. After completing the structural reasoning of each layer, the model maps the weighted fusion feature representation to a risk score value through the output layer to obtain the overall risk level score.

[0076] S702: Mapping the overall risk level score to a predicted risk level according to a preset risk level classification rule.

[0077] In this embodiment, the preset risk level classification rule refers to a segmentation rule for mapping the overall risk level score to a risk level label, which consists of multiple scoring 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 a text description.

[0078] Specifically, a set of risk level division rules is preset according to the numerical range of the overall risk level score. The rule divides the risk score interval into multiple non-overlapping level segments according to the numerical threshold. Each level segment corresponds to a risk level label. The risk level labels include five level gears: low risk, medium-low 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 the numerical value to the 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, if Figure 6 As shown, in step S80, the client entity identification information is matched with the subject identification information in the original material information to obtain the entity identification consistency verification result, including: S801: Obtain target entity identification information from client entity identification information, and obtain reference entity identification information from original material information.

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

[0081] Specifically, an entity field extraction operation is performed on the customer's pre-loan due diligence information, and a named entity recognition model is used to identify customer-related information fields with unique identification. The identified fields include company name, project name, legal representative name, registered address, unified social credit code, contact information and account opening bank, etc. These field information are combined according to predefined field standards to construct target entity identification information. Subsequently, structured field parsing and text content extraction are performed on the original material information. Data items corresponding to the above-mentioned 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 field combination is organized into reference entity identification information to obtain the target entity identification information and reference entity identification information for subsequent matching calculations.

[0082] S802: Using a field-level joint matching strategy, perform field alignment and similarity calculation on the target entity identification information and the reference entity identification information 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 basic matching units when performing a comparison between customer entity identification information and subject identification information in 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. Among them, field name standardization identifies semantically identical field names through synonym mapping or label unification rules. Field content format unification includes character encoding consistency, numeric format specification, and case unification. Field label alignment is used to establish a one-to-one matching mapping relationship. Field value similarity calculation comprehensively uses the 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 identification information and the reference entity identification information are grouped and mapped according to the field labels respectively. In the field-level joint matching strategy, a field alignment rule table is first constructed. The rule table uniformly maps the deformation, abbreviation and alias 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. 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. The edit distance calculation measures the degree of character deformation, word vector analysis is used to deal with semantic expression deviations, and spelling similarity evaluates abnormal abbreviations and symbol differences. The above results are weighted and fused to form a matching score for each pair of field values. Finally, the matching scores of all field pairs are filled in a two-dimensional matrix structure in rows and columns to obtain an entity matching score matrix.

[0085] S803: Determine the matching relationship 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 an entity recognition consistency verification result.

[0086] Specifically, the matching scores of each pair of fields in the entity matching score matrix are compared with the preset similarity threshold. The comparison method is to perform a greater than or equal judgment operation on the score value of each cell in the matrix. When 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 an unmatched field. After completing all field-level judgments, the number of matching fields is counted, and the proportion of matching fields to the total number of fields is calculated. The proportion value is used as the basic basis for the overall entity consistency, and interval matching is performed with the set consistency level rules. If the proportion value exceeds the standard threshold for consistency verification, the matching relationship between the target entity identification information and the reference entity identification information is marked as a consistent state, otherwise it is marked as an inconsistent state, thereby obtaining the entity recognition consistency verification result. In one embodiment, if Figure 7 As shown, after step S80, a method for evaluating the accuracy of bank reports further 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 multi-dimensional consistency indicator set.

[0087] Specifically, the semantic consistency score is used as the first dimension indicator, the term matching accuracy in the term consistency verification result is used as the second dimension indicator, the numerical deviation rate or numerical consistency label in the numerical consistency verification result is used as the third dimension indicator, the risk level comparison result in the risk conclusion consistency verification result is used as the fourth dimension indicator, and the entity matching status in the entity recognition consistency verification result 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 process includes unifying the decimal places of the score values, Boolean conversion of the comparison labels, mapping and encoding the classification labels, and writing them into the consistency indicator data container in a fixed field order to construct an indicator structure containing five categories of sub-items. Finally, the structure is output as a set of overall data objects to obtain a multidimensional consistency indicator set.

[0088] S805: Input the multidimensional consistency index set into the accuracy evaluation model to obtain an 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, which is usually a regression-type deep learning network that can evaluate the comprehensive accuracy level of pre-loan due diligence information through cross-feature modeling of input indicators in each dimension.

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

[0091] S806: Based on the accuracy scoring result, the accuracy of the customer's pre-loan due diligence information is determined to obtain accuracy assessment information.

[0092] Specifically, the numerical value corresponding to the accuracy score result is compared with the preset accuracy grade classification standard step by step. The accuracy grade classification standard is composed of multiple non-overlapping numerical intervals, and each interval corresponds to an accuracy grade label, including three grade levels of high accuracy, medium accuracy, and low accuracy. The accuracy score result is compared with the boundary value of each grade to make a position judgment, so as to determine the overall accuracy level of the current customer's pre-loan due diligence information. At the same time, combined with the consistency sub-results of the five dimensions of semantics, terminology, numerical value, risk and entity, the indicator items with abnormalities or deviations are marked and used as additional explanation content in the evaluation conclusion. By jointly sorting out the score levels and abnormal items, the accuracy grade label and abnormal indicator structure are constructed, and uniformly packaged into accuracy assessment information to obtain accuracy assessment information for subsequent correction suggestion generation.

[0093] S807: Generate corresponding correction suggestion information based on the accuracy evaluation information.

[0094] Specifically, the score value in the accuracy assessment information is jointly analyzed with the consistency results of each dimension. First, the trigger condition for generating correction suggestions is determined based on the accuracy score value. When the score value is lower than the preset accuracy threshold, the correction suggestion generation module is started. During the generation process, the paragraph content with semantic deviations in the original material information and due diligence information is located based on the sub-items with semantic consistency scores lower than the standard. For the case where the term matching accuracy does not reach the threshold, a list of unmatched terms is listed and corresponding standard terms are recommended for replacement. For numerical inconsistencies, the specific deviation between the reference value and the description 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, based on the entity matching failure field, the customer identification field content recommended for correction is listed. Finally, the above content is output as correction suggestion information in a structured format to obtain feedback suggestion text that can be checked, modified and manually intervened.

[0095] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean 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 evaluating the accuracy of bank reports is provided, and the device for evaluating the accuracy of bank reports corresponds one-to-one with the method for evaluating the accuracy of bank reports in the above embodiment. Figure 8 As shown, the device for evaluating the accuracy of bank reports includes a source 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 identification consistency verification module. Each functional module is described in detail below: The original material information acquisition module is used to obtain the original material information of the target bank customers; A due diligence generation module, which uses artificial intelligence models to obtain pre-loan due diligence information on customers that corresponds to the original material information; 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 to generate the original material information embedding vector and the due diligence information embedding vector; A 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 meets the preset credibility condition. It then uses the PXL-BLEU algorithm combined with standard terminology in the preset terminology standard library to perform a normative comparison on the financial terminology information to obtain the terminology consistency verification result; The numerical consistency verification module is used to extract the original information used for calculating the financial indicator information from the original material information based on the financial indicator information identified from the term consistency verification results, input the original information into the preset indicator calculation formula, 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, and obtain the numerical consistency verification result; The risk conclusion consistency verification module is used to identify the risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information if there is abnormal financial indicator information in the numerical consistency verification result. 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. The risk feature vector is input into the trained risk level assessment model to obtain a predicted risk level. The predicted risk level is compared with the risk conclusion 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 customer pre-loan due diligence information based on the semantic consistency score and term consistency verification results, match the customer entity identification information with the subject identification information in the original material information, and 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 for the BERTScore model to perform multi-layer encoding on the original material information to obtain a sequence of original material information embedding representations containing context-dependent features, and perform vector aggregation processing on the original material information embedding representation sequence to generate an original material information embedding vector; The due diligence information semantic embedding generation sub-module is used by the BERTScore model to perform multi-layer encoding on the customer's pre-loan due diligence information 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 a due diligence information embedding vector.

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

[0099] Optionally, the terminology consistency verification module includes: The term set to be matched submodule is used to construct a set of terms to be compared based on financial term information; The initial term similarity scoring submodule is used to calculate the similarity scores between the terms to be compared in the term set and the standard terms in the preset term specification library using the PXL-BLEU algorithm to obtain the corresponding term matching score set; The term matching status determination submodule is used to 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 based on the term matching score set. If so, the term to be compared is marked as a standard matching state; if not, the term to be compared is marked as an unmatched state; The term replacement candidate generation submodule is used to summarize the terms to be compared corresponding to the unmatched state to obtain a set of term replacement candidates; The term status update scoring submodule is used to calculate the similarity score between the to-be-compared terms in the term replacement candidate set and the standard terms using the PXL-BLEU algorithm, update the matching status of the corresponding to-be-compared terms, and obtain the updated matching status; The term matching status merging submodule is used to merge the updated matching status with the standard matching status to obtain the final matching status information; The term consistency verification generation submodule is used to calculate the term matching accuracy according to the final matching status information, and generate the term consistency verification result based on the term matching accuracy.

[0100] Optional, 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 trained risk level assessment model based on the attention mechanism, and perform weighted processing based on the time evolution trend of the 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 identification consistency verification module includes: The entity identification information extraction submodule is used to obtain the target entity identification information from the client entity identification information and obtain the reference entity identification information from the original material information; The entity field matching calculation submodule is used to use the field-level joint matching strategy to perform field alignment and similarity calculation on the target entity identification information and the reference entity identification information to 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 identification consistency verification module includes: The consistency index summary module is used to summarize the semantic consistency score, term 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; An accuracy score calculation module is used to input the multidimensional consistency index set into the accuracy evaluation model to obtain the accuracy score result; Due diligence information evaluation module is used to determine the accuracy of the customer's pre-loan due diligence information based on the accuracy scoring results and obtain accuracy evaluation information The correction suggestion information generation module is used to generate corresponding correction suggestion information based on the accuracy evaluation information.

[0103] The specific definitions of a device for evaluating the accuracy of bank reports can be found in the definitions of a method for evaluating the accuracy of bank reports described above and will not be further elaborated here. Each module in the device for evaluating the accuracy of bank reports can be implemented in whole or in part via software, hardware, or a combination thereof. Each of these modules can be embedded in or independent of a processor in a computer device in hardware form, or stored in a computer device memory in software form, so that the processor can call and execute the corresponding operations of each module.

[0104] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 9 As shown. The computer device includes a processor, memory, a network interface, and a database connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When executed by the processor, the computer program implements a method for evaluating 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. When the processor executes the computer program, the following steps are performed: Obtain original material information of target bank customers; Utilize artificial intelligence models to obtain pre-loan due diligence information on customers that corresponds to the original material information; Use the BERTScore model to perform contextual semantic embedding on the original material information and the customer's pre-loan due diligence information, generating an embedding vector for the original material information and an embedding vector for the due diligence information. Calculate the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and perform greedy matching to obtain the semantic consistency score; If the semantic consistency score meets the preset credibility condition, financial terminology information is obtained from the customer's pre-loan due diligence information. The PXL-BLEU algorithm is used in combination with standard terms in the preset terminology standard library to perform a normative comparison on the financial terminology information to obtain the terminology consistency verification result; Based on the financial indicator information identified from the terminology consistency verification results, extract the original information used for calculating the financial indicator information from the original material information. Enter 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; If there is abnormal financial indicator information in the numerical consistency verification result, 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. The risk feature vector is input into the trained risk level assessment model to obtain a predicted risk level. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain a risk conclusion consistency verification result; Based on the semantic consistency score and terminology consistency verification results, the customer entity identification information is extracted from the customer's 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.

[0106] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented: Obtain original material information of target bank customers; Utilize artificial intelligence models to obtain pre-loan due diligence information on customers that corresponds to the original material information; Use the BERTScore model to perform contextual semantic embedding on the original material information and the customer's pre-loan due diligence information, generating an embedding vector for the original material information and an embedding vector for the due diligence information. Calculate the cosine similarity between the original material information embedding vector and the due diligence information embedding vector and perform greedy matching to obtain the semantic consistency score; If the semantic consistency score meets the preset credibility condition, financial terminology information is obtained from the customer's pre-loan due diligence information. The PXL-BLEU algorithm is used in combination with standard terms in the preset terminology standard library to perform a normative comparison on the financial terminology information to obtain the terminology consistency verification result; Based on the financial indicator information identified from the terminology consistency verification results, extract the original information used for calculating the financial indicator information from the original material information. Enter 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; If there is abnormal financial indicator information in the numerical consistency verification result, 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. The risk feature vector is input into the trained risk level assessment model to obtain a predicted risk level. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain a risk conclusion consistency verification result; Based on the semantic consistency score and terminology consistency verification results, the customer entity identification information is extracted from the customer's 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.

[0107] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may 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 many forms such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double 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 of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by 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 the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present application, and should all be included in the scope of protection of the present application.

Claims

1. A method for evaluating the accuracy of bank reports, characterized in that: The method for evaluating the accuracy of bank reports includes: Obtain original material information of target bank customers; Utilize artificial intelligence models to obtain pre-loan due diligence information on customers corresponding to the source material information; Using the BERTScore model, contextual semantic embedding is performed on the original material information and the customer's pre-loan due diligence information to generate an original material information embedding vector and a due diligence information embedding vector; Calculate 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; If the semantic consistency score meets 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 standard terms in a preset terminology standard library to perform a normative comparison on the financial terminology information to obtain a terminology consistency verification result; The financial indicator information identified from the terminology consistency verification result is extracted from the original material information based on the financial indicator information, and the original information used for calculating the financial indicator information is input into a preset indicator calculation formula to generate a corresponding reference value. The indicator description value is obtained from the customer pre-loan due diligence information, and the reference value is compared with the description value to obtain a numerical consistency verification result. If abnormal financial indicator information exists in the numerical consistency verification result, then identifying risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information, extracting corresponding risk feature information from the original material information based on the risk judgment statement information, 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. The predicted risk level is compared with the risk conclusion in the due diligence information to obtain a risk conclusion consistency verification result; Based on the semantic consistency score and the term 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 an entity identification consistency verification result.

2. A method for evaluating the accuracy of bank reports according to claim 1, characterized in that: 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 to generate an original material information embedding vector and a due diligence information embedding vector, 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, and performs vector aggregation processing 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, and performs the vector aggregation process on the due diligence information embedding representation sequence to generate the due diligence information embedding vector.

3. A method for evaluating the accuracy of bank reports according to claim 1, characterized in that: The calculating 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: Dividing the original material information embedding vector into a first sub-vector set, and dividing the due diligence information embedding vector into a second sub-vector set; Combining the first sub-vector set and the second sub-vector set, performing the cosine similarity calculation between each pair of sub-vectors, and generating a similarity score matrix; Based on the similarity score matrix, adopting the greedy matching, in each matching round, selecting a sub-vector pair with a high similarity score from the first sub-vector set that has not participated in the matching and the second sub-vector set that has not participated in the matching; Marking the sub-vector pair as matched, and adding the sub-vector pair to a semantic matching set; After all the sub-vector pairs are selected, an average calculation is performed on the similarity scores of the sub-vector pairs in the semantic matching set to generate the semantic consistency score.

4. A method for evaluating the accuracy of bank reports according to claim 1, characterized in that: The 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 on the financial terminology information to obtain a terminology consistency verification result, including: Constructing a set of terms to be compared based on the financial term information; Using the PXL-BLEU algorithm, similarity scores are calculated for the terms to be compared in the set of terms to be compared and the standard terms in the preset term specification library to obtain a corresponding term matching score set; Based on the term matching score set, it is determined whether each of the to-be-compared terms has a matching relationship with any of the standard terms in the term specification library with a matching score higher than a preset similarity threshold; if so, the to-be-compared term is marked as a canonical matching state; if not, the to-be-compared term is marked as an unmatched state; Summarizing the to-be-compared terms corresponding to the unmatched states to obtain a set of term replacement candidates; Calculating similarity scores between the terms to be compared in the term replacement candidate set and the standard term using the PXL-BLEU algorithm, and updating the matching status corresponding to the terms to be compared to obtain an updated matching status; Merging the updated matching status with the standard matching status to obtain final matching status information; The term matching accuracy is calculated according to the final matching status information, and the term consistency verification result is generated based on the term matching accuracy.

5. A method for evaluating the accuracy of bank reports according to claim 1, characterized in that: The risk feature vector is constructed based on the risk feature information, and the risk feature vector is input 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 a preset risk level classification rule, the overall risk level score is mapped to the predicted risk level.

6. A method for evaluating the accuracy of bank reports according to claim 1, characterized in that: The matching of the client entity identification information with the subject identification information in the original material information to obtain an entity identification consistency verification result includes: Acquire target entity identification information from the client entity identification information, and acquire 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 identification information and the reference entity identification 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 method for evaluating the accuracy of bank reports according to claim 1, characterized in that: The method for evaluating the accuracy of bank reports further includes: Summarizing the semantic consistency score, the term consistency verification result, the numerical consistency verification result, the risk conclusion consistency verification result, and the entity recognition consistency verification result to obtain a multi-dimensional consistency indicator set; Inputting the multidimensional consistency index set into the accuracy assessment model to obtain an accuracy score result; Determining the accuracy of the customer's pre-loan due diligence information based on the accuracy scoring result to obtain accuracy assessment information; Based on the accuracy evaluation information, corresponding correction suggestion information is generated.

8. A device for evaluating the accuracy of bank reports, characterized in that: The device for evaluating the accuracy of bank reports comprises: The original material information acquisition module is used to obtain the original material information of the target bank customers; A due diligence generation module, configured to utilize an artificial intelligence model to obtain pre-loan due diligence information on the customer corresponding to the source material information; A semantic embedding processing module, configured to perform contextual semantic embedding processing on the source material information and the customer pre-loan due diligence information using a BERTScore model, to generate an embedding vector for the source material information and an embedding vector for the due diligence information; a semantic consistency calculation module, configured to calculate 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; A terminology consistency verification module is configured to obtain financial terminology information from the customer's pre-loan due diligence information if the semantic consistency score meets a preset credibility condition, and perform a normative comparison on the financial terminology information using the PXL-BLEU algorithm combined with standard terminology in a preset terminology standard library to obtain a terminology consistency verification result; a numerical consistency verification module configured to extract, from the original material information, original information used for calculating the financial indicator information identified from the terminology consistency verification result, the financial indicator information, based on the financial indicator information, input the original information into a preset indicator calculation formula to generate a corresponding reference value, obtain an indicator description value from the customer's pre-loan due diligence information, compare the reference value with the description value, and obtain a numerical consistency verification result; a risk conclusion consistency verification module, configured to, if abnormal financial indicator information is present in the numerical consistency verification result, identify risk judgment statement information associated with the abnormal financial indicator information in the customer's pre-loan due diligence information, extract corresponding risk feature information from the original material information based on the risk judgment statement information, construct a risk feature vector based on the risk feature information, input the risk feature vector into a trained risk level assessment model to obtain a predicted risk level, compare the predicted risk level with the risk conclusion in the due diligence information, and obtain a risk conclusion consistency verification result; An 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 term consistency verification result, match the customer entity identification information with the subject identification information in the original material information, and obtain an entity recognition consistency verification result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for evaluating the accuracy of bank reports according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of a method for evaluating the accuracy of bank reports as claimed in any one of claims 1 to 7 are implemented.

Images