A bill risk assessment method, device, equipment and medium

Abstract

The present application relates to the technical field of data processing, and discloses a kind of bill risk assessment method, device, equipment and medium, comprising: obtaining the content data of bill to be evaluated, and the business data associated with bill to be evaluated;Content data is carried out multi-modal feature extraction, obtains the first multi-modal feature of content data, and the business data is carried out multi-modal feature extraction, obtains second multi-modal feature;Based on preset processing rule base, matching target processing rule corresponding to the first multi-modal feature and second multi-modal feature;Based on target processing rule, risk assessment is carried out to bill to be evaluated, and risk assessment result is obtained.The present application can be applied to the data processing scene in the field of financial technology and medical health, improves the risk assessment efficiency and risk assessment accuracy of bill.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, equipment and medium for assessing the risk of invoices. Background Technology

[0002] In the process of digitalizing corporate financial management, invoice risk assessment, as a key link in the financial process, bears the core responsibility of confirming the authenticity, compliance, and accuracy of expenditures and accounting records. Currently, mainstream financial systems and some automated tools have implemented some invoice risk assessment functions, such as preliminary comparison by matching purchase orders with contracts and automatic push of approval workflows.

[0003] However, existing technologies still have significant limitations and systemic defects. On the one hand, existing systems generally adopt static rule-driven evaluation logic, which is difficult to cope with complex and ever-changing business scenarios. Once business rules change or new regulatory policies are added, scripts or configurations need to be manually modified, resulting in long response cycles, high error rates, and a lack of adaptability. On the other hand, most systems can only extract structured fields from invoices and cannot understand the business semantics behind them, leading to a high misjudgment rate and requiring extensive manual review, indicating weak recognition and semantic understanding capabilities. Furthermore, existing systems cannot identify potential risks such as counterfeit invoices, duplicate reimbursements, abnormal amounts, or related-party transactions. They lack proactive early warning mechanisms for high-risk scenarios, leaving room for financial fraud and compliance loopholes, and their intelligent decision-making and risk warning capabilities are insufficient. Therefore, there is an urgent need for a method for invoice risk assessment that can achieve multimodal feature extraction and intelligent risk assessment. Summary of the Invention

[0004] This invention provides a method, apparatus, equipment, and medium for bill risk assessment to solve the technical problems of insufficient feature extraction, poor adaptability of risk assessment, and low detection efficiency in the prior art.

[0005] Firstly, this disclosure provides a method for assessing the risk of negotiable instruments, including: Obtain the content data of the invoice to be evaluated, as well as the business data associated with the invoice; Multimodal feature extraction is performed on the content data to obtain the first multimodal feature of the content data, and multimodal feature extraction is performed on the business data to obtain the second multimodal feature; Based on a preset processing rule library, target processing rules are matched with the first multimodal feature and the second multimodal feature; Based on the target processing rules, a risk assessment is conducted on the invoices to be evaluated, and the risk assessment results are obtained.

[0006] Secondly, this disclosure provides a bill risk assessment device, comprising: The data acquisition module is used to acquire the content data of the invoice to be evaluated, as well as the business data associated with the invoice to be evaluated; The feature extraction module is used to perform multimodal feature extraction on content data to obtain the first multimodal feature of the content data, and to perform multimodal feature extraction on business data to obtain the second multimodal feature; The rule determination module is used to match the target processing rule corresponding to the first multimodal feature and the second multimodal feature based on a preset processing rule library; The risk assessment module is used to perform risk assessments on the invoices to be assessed based on the target processing rules, and obtain the risk assessment results.

[0007] Thirdly, this disclosure provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned bill risk assessment method.

[0008] Fourthly, this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described bill risk assessment method.

[0009] In the aforementioned scheme implemented by the bill risk assessment method, device, equipment, and medium, the following steps are taken: First, the content data of the bill to be assessed, and the business data associated with the bill to be assessed, are acquired; second, multimodal features are extracted from the content data to obtain a first multimodal feature, and multimodal features are extracted from the business data to obtain a second multimodal feature; third, target processing rules corresponding to the first and second multimodal features are matched based on a preset processing rule library; and fourth, a risk assessment is performed on the bill to be assessed based on the target processing rules to obtain a risk assessment result. This invention achieves the linked acquisition of invoice data and related business data by obtaining the content data and associated business data of the invoice to be evaluated, providing a comprehensive data foundation for subsequent risk assessment and avoiding assessment bias caused by a single data source. By extracting multimodal features from both content and business data, it enables feature modeling of invoice and business information from multiple dimensions, achieving semantic association between invoice content and business context, and improving the richness and accuracy of feature expression. Through matching target processing rules corresponding to multimodal features with a preset processing rule base, it achieves intelligent matching of rules and features, enabling the risk assessment logic to adapt to different invoice types and business scenarios, improving the versatility and flexibility of the solution. Finally, by conducting risk assessment on the invoice to be evaluated based on target processing rules, it achieves rule-driven intelligent risk assessment, improving the efficiency and accuracy of risk assessment. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic diagram of a scenario for a bill risk assessment method according to an embodiment of the present invention; Figure 2 This is a flowchart of a bill risk assessment method according to an embodiment of the present invention; Figure 3 This is another flowchart of the bill risk assessment method in one embodiment of the present invention; Figure 4 This is another flowchart of the bill risk assessment method in one embodiment of the present invention; Figure 5 This is another flowchart of the bill risk assessment method in one embodiment of the present invention; Figure 6 This is another flowchart of the bill risk assessment method in one embodiment of the present invention; Figure 7 This is another flowchart of the bill risk assessment method in one embodiment of the present invention; Figure 8 This is another flowchart of the bill risk assessment method in one embodiment of the present invention; Figure 9 This is a schematic diagram of a bill risk assessment device according to an embodiment of the present invention; Figure 10 This is a schematic block diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0012] The bill risk assessment method provided in this invention can be applied to, for example... Figure 1In this application environment, the client communicates with the server via a network. The server can receive the invoice to be evaluated uploaded by the client, obtain the content data and associated business data of the invoice; extract multimodal features from the content data and business data respectively; match target processing rules based on a preset processing rule base; perform risk assessment based on the target processing rules, obtain the risk assessment result, and return the risk assessment result to the client. This invention can be applied to invoice risk assessment scenarios in the fields of fintech and healthcare, and can improve the accuracy, efficiency, and adaptability of risk assessment. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0013] In one embodiment, such as Figure 2 As shown, a method for assessing the risk of bills is provided. Taking the application of this method on the server side as an example, it includes the following steps: S101, Obtain the content data of the invoice to be evaluated, as well as the business data associated with the invoice to be evaluated; S102, perform multimodal feature extraction on the content data to obtain the first multimodal feature of the content data, and perform multimodal feature extraction on the business data to obtain the second multimodal feature; S103, based on the preset processing rule library, match the target processing rule corresponding to the first multimodal feature and the second multimodal feature; S104. Based on the target processing rules, a risk assessment is performed on the invoices to be evaluated, and the risk assessment results are obtained.

[0014] As an example, in step S101, after receiving the invoice to be evaluated, the server first obtains the content data of the invoice to be evaluated, as well as the business data associated with the invoice to be evaluated.

[0015] Among them, the invoices to be assessed refer to various invoices that require risk assessment, including but not limited to invoices, receipts, drafts, checks, etc.; content data refers to the data information contained in the invoice itself, including but not limited to the text, numbers, seals, layout, etc. on the invoice; business data refers to the business context data associated with the invoice, including but not limited to purchase contract data, payment application data, approval record data, budget usage data, etc.

[0016] In one optional embodiment, the content data of the invoice to be evaluated can be obtained by receiving the invoice file uploaded by the user. The user can submit the invoice to be evaluated through various methods such as taking a photo with a mobile phone, uploading from a computer, or system integration.

[0017] In an optional embodiment, business data associated with the invoice to be evaluated can be obtained by calling the application programming interface (API) of the business system, thereby enabling automatic acquisition of cross-system data.

[0018] For example, in the scenario of corporate expense reimbursement, after an employee uploads a travel and accommodation invoice, the server first obtains the content data of the invoice, including the invoice amount, invoice date, invoice issuer, accommodation details, and other information; then it obtains the business data associated with the invoice, including travel application forms, approval records, departmental budget usage, employee historical reimbursement records, and other information, providing comprehensive data support for subsequent risk assessment.

[0019] For example, in the scenario of medical expense reimbursement, after a patient uploads an outpatient bill, the server first obtains the content data of the bill, including the bill amount, charge items, date of visit, name of medical institution, etc.; then it obtains the business data associated with the bill, including medical insurance settlement records, medical records, drug list, patient's historical reimbursement records, etc., to provide comprehensive data support for subsequent risk assessment.

[0020] As an example, in step S102, after obtaining the content data and business data, multimodal feature extraction is performed on the content data and business data respectively to obtain the first multimodal feature and the second multimodal feature.

[0021] Among them, multimodal feature extraction refers to extracting feature information from data of multiple dimensions and types; the first multimodal feature refers to the multi-dimensional features extracted from invoice content data; the second multimodal feature refers to the multi-dimensional features extracted from business data.

[0022] In an alternative embodiment, a deep learning model can be used for multimodal feature extraction, such as extracting visual features through a convolutional neural network and extracting semantic features through a recurrent neural network or a Transformer model.

[0023] In an optional embodiment, a multimodal large model can be used for joint feature extraction to achieve interaction and fusion between features of different modalities.

[0024] For example, taking invoice risk assessment as an example, when performing multimodal feature extraction on invoice content data, we can extract the invoice's layout structure features, key field features, business semantic features, etc., to form the first multimodal feature; when performing multimodal feature extraction on related business data, we can extract contract amount features, approval process features, supplier association features, etc., to form the second multimodal feature.

[0025] As an example, in step S103, after obtaining the first multimodal feature and the second multimodal feature, the target processing rule corresponding to the first multimodal feature and the second multimodal feature is matched based on the preset processing rule library.

[0026] The preset processing rule base refers to a pre-configured set of rules containing multiple risk assessment rules; the target processing rule refers to the risk assessment rule that best matches the current invoice to be assessed and the business scenario. Each rule in the processing rule base can contain two parts: rule conditions and rule execution logic.

[0027] In one optional embodiment, rule matching can be performed by feature matching, comparing multimodal features with processing rules in the rule base one by one to filter out rules that meet the conditions.

[0028] In an optional embodiment, rule matching can be performed by similarity calculation, calculating the similarity between multimodal features and each processing rule, and selecting the rule with the highest similarity as the target processing rule.

[0029] For example, in the scenario of travel expense reimbursement, when the invoice content is detected as accommodation fee and the amount exceeds the department's budget, a high-risk warning rule is matched, which requires triggering a manual review process; when the invoice amount is small and meets the regular reimbursement standards, a low-risk exemption rule is matched, which allows the invoice to be directly reimbursed.

[0030] As an example, in step S104, after matching the target processing rule, a risk assessment is performed on the invoice to be evaluated based on the target processing rule to obtain the risk assessment result.

[0031] Risk assessment refers to judging and quantifying the risk level of bills based on target processing rules; risk assessment results refer to the output of risk assessment, including but not limited to risk scores, risk levels, risk descriptions and other information.

[0032] In one optional embodiment, the target processing rules can be executed through a rule engine to verify each invoice and obtain risk assessment results.

[0033] In one alternative embodiment, a joint risk assessment can be performed by combining a machine learning model with a rule engine to further improve the accuracy of the assessment.

[0034] For example, taking supplier risk assessment as an example, based on the matched "supplier abnormal behavior detection" rules, a comprehensive assessment is conducted on the supplier's recent invoicing frequency, amount distribution, cooperation duration, and other dimensions to calculate the supplier's risk score and risk level, and generate a corresponding risk description. For example, if the supplier's invoicing frequency has increased significantly in the past 3 months compared to the same period last year, it is recommended to continue to monitor the situation.

[0035] This invention provides a method for assessing the risk of a bill of exchange, comprising: acquiring content data of the bill to be assessed and business data associated with the bill to be assessed; performing multimodal feature extraction on the content data to obtain a first multimodal feature of the content data, and performing multimodal feature extraction on the business data to obtain a second multimodal feature; matching target processing rules corresponding to the first and second multimodal features based on a preset processing rule library; and performing a risk assessment on the bill to be assessed based on the target processing rules to obtain a risk assessment result. This method acquires the content data and associated business data of the invoices to be evaluated, enabling the linked acquisition of invoice and business data. This provides a comprehensive data foundation for subsequent risk assessment, avoiding assessment bias caused by a single data source. By extracting multimodal features from both content and business data, it achieves feature modeling of invoice and business information from multiple dimensions, realizing semantic association between invoice content and business context, and improving the richness and accuracy of feature expression. By matching target processing rules corresponding to multimodal features with a preset processing rule base, it achieves intelligent matching of rules and features, enabling the risk assessment logic to adapt to different invoice types and business scenarios, improving the versatility and flexibility of the solution. Finally, by conducting risk assessment on the invoices to be evaluated based on target processing rules, it achieves rule-driven intelligent risk assessment, improving the efficiency and accuracy of risk assessment.

[0036] In one embodiment, such as Figure 3 As shown, the steps in S101 above, which involve obtaining the content data of the invoice to be evaluated and the business data associated with the invoice, specifically include the following steps: S201, perform format identification on the invoice to be evaluated to determine the data format of the invoice to be evaluated; S202, based on the data format, parse the invoice to be evaluated and extract the content data of the invoice to be evaluated; S203, based on the identification information in the content data, obtain the corresponding business data.

[0037] As an example, in step S201, after receiving the invoice to be evaluated, the format of the invoice to be evaluated is first identified to determine the data format of the invoice to be evaluated.

[0038] The data format refers to the form in which the invoice data is presented; different data formats correspond to different parsing methods. Format recognition allows for the selection of appropriate processing methods for subsequent parsing steps, improving the accuracy and efficiency of the parsing process.

[0039] In an alternative embodiment, the data format can be identified through the file header information, for example, by reading the first few bytes of the file to determine the file format type.

[0040] In an alternative embodiment, the data format can be identified by the file extension, for example, by determining the corresponding format type based on different file extensions.

[0041] For example, after receiving the ticket file uploaded by the user, the server reads the file header information to identify that the file is in image format, and then selects the image recognition and parsing processing method.

[0042] As an example, in step S202, after determining the data format of the invoice to be evaluated, the invoice to be evaluated is parsed based on the data format to extract the content data of the invoice to be evaluated.

[0043] Parsing refers to the process of converting the raw data of a document into processable content data. Different data formats employ different parsing methods to ensure the accuracy of content data extraction.

[0044] In an optional embodiment, for image-formatted tickets, optical character recognition technology can be used for parsing to convert the text information in the image into text data.

[0045] In an optional embodiment, for tickets in a structured data format, the field information can be read directly to quickly extract the content data.

[0046] For example, for a scanned invoice, optical character recognition technology is used to parse it and extract data such as invoice number, invoice date, amount, tax, seller information, and buyer information, laying the foundation for subsequent feature extraction.

[0047] As an example, in step S203, after extracting the content data, the corresponding business data is obtained based on the identification information in the content data.

[0048] Identification information refers to information that uniquely identifies a bill or related transaction, such as bill number, contract number, and application number. Through identification information, corresponding business data can be accurately retrieved from the related business system.

[0049] In one optional embodiment, the corresponding expense records, payment records, and other business data can be retrieved from the financial system based on the invoice number as identification information.

[0050] In one optional embodiment, the contract number can be used as identification information to retrieve corresponding contract information, performance records, and other business data from the contract management system.

[0051] For example, after extracting the invoice number, the system uses that invoice number as a keyword to retrieve related business data such as purchase applications, approval records, payment information, and budget usage from multiple business systems, including the enterprise resource planning system, office automation system, and contract management system, thereby achieving automatic association between invoices and business data.

[0052] In one embodiment, such as Figure 4 As shown, step S102 above involves multimodal feature extraction of the content data to obtain the first multimodal feature, specifically including the following steps: S301, Perform character recognition on the content data to obtain the invoice text data; S302, Extract fields from the document text data to obtain document structural features; S303, Perform semantic analysis on the document text data to obtain the semantic features of the document; S304, the structural features and semantic features of the bill are fused to obtain the first multimodal feature.

[0053] As an example, in step S301, character recognition is performed on the content data to obtain the ticket text data.

[0054] Character recognition refers to the process of converting text, numbers, and other character information in content data into editable text data. Through character recognition, the visual information on tickets can be converted into machine-processable text, laying the foundation for subsequent field extraction and semantic analysis.

[0055] In one alternative embodiment, a deep learning-based optical character recognition model can be used for character recognition, such as an end-to-end recognition model based on convolutional neural networks and recurrent neural networks, to improve the recognition accuracy in complex scenarios.

[0056] In one optional embodiment, a dedicated character recognition model optimized for the ticket scenario can be used to recognize characters by combining the layout characteristics of the ticket, thereby further improving the recognition speed and accuracy.

[0057] For example, character recognition is performed on the image of the invoice to identify all the text, numbers, symbols, and other content on the invoice, resulting in complete invoice text data, including the invoice name, invoice code, invoice number, invoice date, buyer information, seller information, name of goods or taxable services, amount, tax amount, and all other text content.

[0058] As an example, in step S302, after obtaining the document text data, the document text data is processed to extract fields and obtain document structure features.

[0059] Field extraction refers to extracting structured key field information from invoice text data; invoice structural features refer to features composed of structured fields, used to express the core information of the invoice. Through field extraction, unstructured text data can be transformed into structured feature representations.

[0060] In one optional embodiment, a rule-based field extraction method can be used to extract the corresponding field values ​​based on the format characteristics and field positions of the document.

[0061] In an optional embodiment, named entity recognition technology can be used for field extraction to automatically identify various entity information in the text, such as amount, date, name, etc.

[0062] For example, key fields such as invoice number, invoice date, amount, tax, seller's name, and buyer's name are extracted from the invoice text data. These fields are then organized into key-value pairs to form invoice structure features, which are used for subsequent rule matching and risk assessment.

[0063] As an example, in step S303, semantic analysis is performed on the document text data to obtain the document semantic features.

[0064] Semantic analysis refers to understanding the semantics of invoice text data to determine the business meaning behind the text; invoice semantic features refer to the features that express the business semantics of the invoice. Through semantic analysis, deeper information such as the business type, expense category, and applicable scenarios of the invoice can be understood.

[0065] In an optional embodiment, a pre-trained language model can be used for semantic analysis, such as a dedicated language model trained on a large-scale financial corpus, to improve the semantic understanding capability in the context of financial instruments.

[0066] In one optional embodiment, a domain-adaptive natural language processing model can be used for semantic analysis to identify key semantic tags in the invoice, such as business type, expense category, and reimbursement scenario.

[0067] For example, semantic analysis is performed on an invoice containing keywords such as "accommodation" and "travel" to understand that the business type of the invoice is "travel expenses" and the expense category is "accommodation expenses". This semantic information is then encoded into invoice semantic features to assist in risk assessment.

[0068] For example, semantic analysis is performed on an outpatient bill containing keywords such as "registration fee," "consultation fee," and "medication fee" to understand that the business type of the bill is "outpatient treatment," the fee category is "medical service fee," and the applicable scenario is "medical insurance reimbursement." This semantic information is then encoded into the bill's semantic features to assist in risk assessment.

[0069] As an example, in step S304, the structural features and semantic features of the bill are fused to obtain the first multimodal features.

[0070] Feature fusion refers to integrating different types of features to form a unified multimodal feature representation. Through feature fusion, structured and semantic information can be comprehensively utilized to improve the expressive power and discriminative power of features.

[0071] In one optional embodiment, feature fusion can be performed by concatenation, which directly concatenates the structural feature vector and the semantic feature vector to form a unified multimodal feature vector.

[0072] In an alternative embodiment, an attention mechanism can be used for feature fusion to adaptively learn the weights of features of different modalities, so that more important features occupy a higher proportion in the fusion result.

[0073] For example, the structural features (amount, date, issuer, etc.) and semantic features (business type, expense category, etc.) of a bill are fused to form a unified first multimodal feature vector. This feature vector contains both the structured information and deep semantic information of the bill, enabling a more comprehensive expression of the bill's characteristics and providing richer evidence for subsequent rule matching and risk assessment.

[0074] In one embodiment, such as Figure 5 As shown, step S102 above involves multimodal feature extraction of the business data to obtain the second multimodal feature, specifically including the following steps: S401, Perform structural analysis on business data to obtain business structure characteristics; S402, Perform correlation analysis on business data to obtain business correlation characteristics; S403, perform semantic analysis on business data to obtain business semantic features; S404 integrates business structure features, business association features, and business semantic features to obtain the second multimodal feature.

[0075] As an example, in step S401, the business data is structurally analyzed to obtain the business structure characteristics.

[0076] Structural analysis refers to the analysis and extraction of structured information from business data; business structural features refer to the characteristics that express the structured attributes of business data. Through structural analysis, key structured indicators can be extracted from multi-source business data, providing a quantitative basis for risk assessment.

[0077] In one optional embodiment, statistical analysis can be performed on the numerical fields in the business data to extract structured indicators such as amount, quantity, and proportion, forming business structure features.

[0078] In one optional embodiment, status fields in the business data can be encoded to convert discrete states such as approval status and payment status into feature representations, thereby enriching the business structure features.

[0079] For example, structural analysis is performed on business data associated with invoices to extract structured indicators such as contract amount, amount already paid, remaining budget, number of approval nodes, and duration of cooperation with suppliers. These indicators are then encoded as business structure features to assess the compliance and risk level of the business.

[0080] As an example, in step S402, correlation analysis is performed on the business data to obtain business correlation characteristics.

[0081] Among them, correlation analysis refers to analyzing the relationships between different entities and processes in business data; business correlation characteristics refer to the features that express the relationships between business data. Through correlation analysis, the underlying network of relationships in business data can be identified, and potential correlation risks can be discovered.

[0082] In one optional embodiment, a business association graph can be constructed, with suppliers, employees, departments, contracts, invoices, etc. as nodes, and the relationships between them as edges, and association features can be extracted through graph analysis algorithms.

[0083] In an optional embodiment, the temporal correlation between different business processes can be analyzed, such as the interval between approval time and payment time, the interval between invoice issuance time and reimbursement time, etc., to extract temporal correlation features.

[0084] For example, a correlation analysis of the related business data of the supplier's transaction invoices revealed frequent small-amount transactions between the supplier and several employees within the company, even though these employees belonged to different departments. Graph analysis algorithms were used to extract the supplier's correlation characteristics, marking them as "suspected of diversified transactions," thus providing a basis for risk assessment.

[0085] As an example, in step S403, semantic analysis is performed on the business data to obtain business semantic features.

[0086] Semantic analysis refers to the deep semantic understanding of textual information in business data; business semantic features refer to the features that express the deeper meaning of business text. Through semantic analysis, the business meaning and potential risks in texts such as contract terms, approval opinions, and reimbursement instructions can be understood.

[0087] In one optional embodiment, semantic analysis can be performed on the contract text to extract semantic information such as key clauses, risk clauses, and special agreements in the contract, forming business semantic features.

[0088] In one optional embodiment, semantic analysis can be performed on the approval opinions to identify semantic information such as attitudes, concerns, and risk warnings in the approval opinions, thereby assisting in risk assessment.

[0089] For example, semantic analysis of a purchase contract text identifies special provisions such as "payment before delivery" and "no liability for breach of contract clauses." This semantic information is encoded as business semantic features and marked as "contract terms pose a risk," prompting its focused assessment during risk evaluation.

[0090] As an example, in step S404, the business structure features, business association features, and business semantic features are fused to obtain the second multimodal features.

[0091] Feature fusion refers to integrating business features from different dimensions to form a unified multimodal feature representation. Through feature fusion, structured information, relational information, and semantic information can be comprehensively utilized to fully characterize the risk features of a business scenario.

[0092] In an optional embodiment, feature concatenation can be used for fusion, whereby the three types of feature vectors are concatenated sequentially to form a second multimodal feature vector.

[0093] In an alternative embodiment, a gating mechanism can be used for feature fusion to adaptively adjust the contribution of different types of features, so that the features most relevant to risk assessment can play a greater role.

[0094] For example, business structure features (amount, budget, status, etc.), business association features (supplier association, employee association, time sequence association, etc.), and business semantic features (contract terms, approval opinions, etc.) are integrated to form a unified second multimodal feature vector. This feature vector comprehensively characterizes the features of the business scenario from three dimensions: structure, association, and semantics, and can provide multidimensional basis for risk assessment.

[0095] In one embodiment, such as Figure 6 As shown, step S103 above, based on a preset processing rule library, matches the target processing rules corresponding to the first multimodal feature and the second multimodal feature, specifically including the following steps: S501, Match the first multimodal feature and the second multimodal feature with the processing rules in the preset processing rule base respectively to obtain a set of candidate processing rules; S502, When the candidate processing rule set contains multiple candidate processing rules, determine the priority information of each candidate processing rule; S503, based on priority information, determines the target processing rule from the set of candidate processing rules.

[0096] As an example, in step S501, the first multimodal feature and the second multimodal feature are matched with the processing rules in the preset processing rule library to obtain a set of candidate processing rules.

[0097] Rule matching refers to comparing the extracted multimodal features with processing rules in the rule base to find rules that meet the conditions; the candidate processing rule set is the set of all processing rules that meet the matching conditions. Through rule matching, processing rules suitable for the current scenario can be selected from the rule base.

[0098] In an optional embodiment, rule matching can be performed using an exact matching method. When the multimodal features fully satisfy the processing rules, the rule is added to the candidate set.

[0099] In an optional embodiment, fuzzy matching can be used for rule matching, calculating the matching degree between multimodal features and processing rules. When the matching degree exceeds a preset threshold, the rule is added to the candidate set.

[0100] For example, the first multimodal feature of the invoice (large amount, business type is equipment purchase) and the second multimodal feature (insufficient budget remaining, supplier is a new cooperative supplier) are matched one by one with the processing rules in the preset processing rule base to filter out multiple rules such as "large purchase risk rule", "new supplier risk rule" and "budget overrun warning rule" to form a candidate processing rule set.

[0101] As an example, in step S502, when the candidate processing rule set contains multiple candidate processing rules, the priority information of each candidate processing rule is determined.

[0102] Priority information refers to information used to indicate the importance or order of application of rules. When multiple rules apply simultaneously, it is necessary to determine the final processing rule based on priority to avoid inconsistent evaluation results caused by rule conflicts.

[0103] In one alternative embodiment, priority can be determined based on the risk level of the processing rules, with rules having higher risk levels having higher priority, ensuring that high-risk rules are executed first.

[0104] In one alternative embodiment, priority can be determined based on the timeliness and regional applicability of processing rules, with the most recently published rules applicable to the current region having higher priority.

[0105] For example, the candidate processing rule set contains three processing rules, which are prioritized according to their risk level: the "budget overspending warning rule" (high risk) has the highest priority, the "new supplier risk rule" (medium risk) has the second highest priority, and the "large purchase risk rule" (low risk) has the lowest priority.

[0106] As an example, in step S503, the target processing rule is determined from the set of candidate processing rules based on priority information.

[0107] The target processing rule refers to the rule ultimately selected for risk assessment. Through priority screening, one or more appropriate processing rules can be selected from multiple applicable rules to ensure the consistency and accuracy of the risk assessment logic.

[0108] In an alternative embodiment, the candidate processing rule with the highest priority can be selected as the target processing rule to ensure that the most important rule is executed first.

[0109] In one optional embodiment, the processing rules in multiple candidate processing rule sets can be weighted and combined according to priority to form a comprehensive target processing rule, taking into account the evaluation logic of multiple rules; alternatively, if the processing rules do not conflict, multiple processing rules in the candidate processing rule set can be directly determined as the target processing rule at the same time.

[0110] In one optional embodiment, priority analysis can be performed on the processing rules in a set of multiple candidate processing rules according to their priority. When a high-priority processing rule and a low-priority processing rule are logically conflicting, the high-priority processing rule is selected as the target processing rule, thereby avoiding logical contradictions.

[0111] In one embodiment, such as Figure 7 As shown, in step S104 above, a risk assessment is performed on the invoice to be evaluated based on the target processing rules to obtain the risk assessment result. Specifically, this includes the following steps: S601, Based on the target processing rules, perform rule verification on the invoice to be evaluated and obtain the rule verification result; S602, Calculate the risk score of the bill to be evaluated based on the rule verification results; S603, based on risk scoring and preset risk level thresholds, determine the risk level of the bill to be evaluated; S604, Based on the rule validation results, generate a risk assessment description; S605 defines risk scores, risk levels, and risk assessment descriptions as risk assessment results.

[0112] As an example, in step S601, the bill to be evaluated is validated based on the target processing rules to obtain the rule validation result.

[0113] Rule verification refers to checking each item of the invoices to be evaluated and related business data according to the logic of the target processing rules; rule verification result refers to the output result of the verification, including the pass status of each verification.

[0114] In an optional embodiment, a rule engine can be used to execute target processing rules, and check whether the invoices and business data to be evaluated meet the requirements one by one according to the verification conditions defined in the rules.

[0115] In one optional embodiment, the rule verification results can be output in a structured form, recording information such as the name, result, and details of each verification, which facilitates subsequent score calculation and explanation generation.

[0116] For example, the purchase invoices are validated based on the "budget overspending warning rule". The rule checks are performed sequentially to see if the invoice amount exceeds the department's monthly budget threshold, whether the supplier is in the list of qualified suppliers, whether the purchase application has been fully approved, and whether there are any recent large transactions with the same supplier, thus obtaining the rule validation results.

[0117] As an example, in step S602, the risk score of the bill to be evaluated is calculated based on the rule verification results.

[0118] Risk score refers to the degree of risk of a bill of exchange as expressed in numerical form; generally, a higher score indicates a greater risk. Quantitative risk scoring provides a more intuitive way to measure the risk level of a bill of exchange.

[0119] In one optional embodiment, a risk score can be calculated by weighting the results of each verification item. Different verification items can be assigned different weights, with more important verification items receiving higher weights.

[0120] In one optional embodiment, a scorecard model can be used to calculate a risk score, with different scores corresponding to different risk factors, and the total score is obtained by accumulating the scores.

[0121] As an example, in step S603, the risk level of the bill to be evaluated is determined based on the risk score and the preset risk level threshold.

[0122] Here, risk level refers to the graded representation of the degree of risk; preset risk level threshold refers to the pre-set score threshold used to divide different risk levels. By dividing risk levels, quantitative scores can be converted into grade representations, so that risk scores within a preset range can be mapped to the same risk level, avoiding jumps in risk assessment results due to slight fluctuations in scores.

[0123] In one optional embodiment, the risk level can be divided into three levels: low risk, medium risk, and high risk, each corresponding to a different score range. More risk levels can also be set according to business needs to achieve a more refined risk classification.

[0124] As an example, in step S604, a risk assessment description can be generated based on the rule verification results, and then the risk score, risk level and risk assessment description can be determined as the risk assessment results.

[0125] The risk assessment description is a written explanation of the risk assessment results, used to explain the reasons, basis, and recommendations for the occurrence of the risk. The risk assessment description reflects the explainability of the risk assessment, helps users understand the assessment results, and increases their trust in the results.

[0126] In one optional embodiment, a structured risk assessment description can be automatically generated based on the rule validation results, including a description of the risk points, the basis for the risk, and handling suggestions.

[0127] In an optional embodiment, natural language generation technology can be used to convert the structured verification results into fluent natural language descriptions, thereby improving the readability of the evaluation descriptions.

[0128] In one optional embodiment, the risk assessment results can be presented in the form of a visual report, including a risk scoring dashboard, risk level indicators, detailed assessment instructions, etc., to improve the readability of the results.

[0129] In an optional embodiment, after determining the risk level, a corresponding invoice approval task can be generated based on the risk level. For example, based on the risk level in the risk assessment result, the corresponding approval process path can be matched from the preset approval path library; based on the business type tag in the semantic features of the invoice, the approval node configuration of the approval process path can be determined; the invoice to be evaluated, related business data and risk assessment results can be synchronously pushed to the corresponding approval node to generate a standardized approval task.

[0130] Among them, the preset approval path library is a set of pre-configured approval links corresponding to each risk level; the approval node configuration refers to the approval positions, approval order and single-node approval permissions included in the approval process.

[0131] Optionally, a three-tiered standardized approval path can be set: when the risk level is low, the system will automatically complete the invoice verification and confirmation without manual approval; when the risk level is medium, the system will match the regular approval path and push the application to the department head and financial specialist of the employee for approval in sequence; when the risk level is high, the system will match the special review path and push the application to the risk control specialist for further review.

[0132] Furthermore, different approval positions can be configured for different business types such as travel expenses, procurement expenses, and office expenses, under the same risk level. For example, procurement invoices need to be reviewed by the procurement position.

[0133] For example, if a travel accommodation expense invoice is assessed as low-risk and the business type is labeled "travel expenses", then the travel expense direct verification path will be matched, and the system will automatically complete the verification registration of the invoice and update the department's budget usage data simultaneously, without manual intervention.

[0134] In one embodiment, such as Figure 8 As shown, after step S104 above, there is also a step of pre-processing rule base self-learning update, which specifically includes the following steps: S701, Obtain the audit feedback information corresponding to the risk assessment results; S702, Based on the audit feedback information, generate rule update information; S703 updates the preset processing rule base based on the updated rule information.

[0135] As an example, in step S701, the audit feedback information corresponding to the risk assessment result is obtained.

[0136] The review feedback information refers to the feedback from human reviewers on the risk assessment results, including whether the assessment results are accurate, whether there are any misjudgments, and whether the risk level is reasonable.

[0137] In one optional embodiment, review feedback information can be obtained through the feedback interface of the review system. When the reviewer completes the review and submits comments, the feedback information is automatically obtained.

[0138] In one alternative embodiment, review feedback information can be exported in batches from the review system periodically for batch optimization of the rule base.

[0139] For example, after manually reviewing an invoice marked as high-risk, the finance staff confirmed it was a normal expense reimbursement, indicating a misjudgment by the system. The finance staff submitted feedback in the review system: "This invoice represents normal travel and accommodation expenses. Due to concentrated departmental reimbursements this month, the budget percentage is high. I suggest lowering the risk level." The system automatically retrieves this feedback information for subsequent rule optimization.

[0140] As an example, in step S702, rule update information is generated based on the audit feedback information.

[0141] Rule update information refers to information used to update the preset processing rule base, including adding, modifying, and deleting rules. By combining audit feedback and invoice characteristics, accurate rule update information can be generated, enabling the rule base to self-learn and optimize.

[0142] In one alternative embodiment, rule adjustment suggestions can be generated based on the deficiencies of the misjudgment case analysis rules, such as adjusting rule thresholds, adding exception conditions, or modifying rule priorities.

[0143] In one alternative embodiment, new risk patterns can be automatically identified based on audit feedback data, using machine learning methods to generate suggestions for new rules.

[0144] For example, based on the misjudgment case of "centralized reimbursement leading to a high budget ratio but no actual risk", combined with the characteristics of the invoice structure (travel and accommodation expenses, reasonable amount) and the characteristics of business semantics (centralized reimbursement at the end of the month, multiple people traveling on business at the same time), rule update information is generated: add an exception condition to the "budget overspending warning rule" when the invoice type is travel and accommodation expenses and the reimbursement time is at the end of the month and multiple people are traveling on business at the same time, appropriately reduce the risk weight of budget overspending.

[0145] For example, based on the misjudgment case of "chronic disease patients regularly purchasing medicines, resulting in high-frequency, low-amount transactions that are actually risk-free", and combining the characteristics of the invoice structure (medicine cost, reasonable amount) and the characteristics of the business semantics (chronic disease diagnosis, regular follow-up visits), rule update information is generated: an exception condition is added to the "high-frequency transaction risk rule" when the invoice type is medicine cost and the patient has a chronic disease diagnosis record and is in the regular follow-up visit cycle, the risk weight of high-frequency transactions is appropriately reduced.

[0146] As an example, in step S703, the preset processing rule base is updated based on the rule update information.

[0147] The pre-defined processing rule base update refers to adding, deleting, and modifying rules in the pre-defined processing rule base based on rule update information. Through rule base updates, the experience of manual review can be integrated into the pre-defined processing rule base, allowing it to continuously evolve and optimize.

[0148] In one optional embodiment, a rule versioning management approach can be adopted for updates, with each update generating a new rule version while retaining historical versions for easy tracing and rollback.

[0149] In one optional embodiment, a review mechanism for rule updates can be set up. Rule update information can only take effect after being reviewed and confirmed by the administrator, ensuring the security and accuracy of rule updates.

[0150] For example, based on the generated rule update information, the "Budget Overspending Warning Rule" is updated to include exceptions such as "travel and accommodation expenses, end of the month, and multiple people traveling on the same business trip." The updated processing rule will automatically identify these exception scenarios and reduce their risk weight when calculating budget overspending risk, thereby reducing misjudgments.

[0151] In an optional embodiment, subsequent risk assessment strategies can also be optimized based on the updated preset processing rule base.

[0152] In one optional embodiment, the effect of the rule update can be verified by A / B testing, comparing the evaluation accuracy, false positive rate and other indicators of the preset processing rule base before and after the update, to ensure that the rule update brings positive results.

[0153] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0154] In one embodiment, a bill risk assessment device is provided, which corresponds one-to-one with the bill risk assessment methods in the above embodiments. For example... Figure 9 As shown, the bill risk assessment device includes a data acquisition module 801, a feature extraction module 802, a rule determination module 803, and a risk assessment module 804. Detailed descriptions of each functional module are as follows: The data acquisition module 801 is used to acquire the content data of the invoice to be evaluated, as well as the business data associated with the invoice to be evaluated; The feature extraction module 802 is used to perform multimodal feature extraction on the content data to obtain the first multimodal feature of the content data, and to perform multimodal feature extraction on the business data to obtain the second multimodal feature; The rule determination module 803 is used to match the target processing rule corresponding to the first multimodal feature and the second multimodal feature based on a preset processing rule library; Risk assessment module 804 is used to perform risk assessment on the invoices to be assessed based on the target processing rules and obtain the risk assessment results.

[0155] In one embodiment, the data acquisition module 801 is further configured to perform format recognition on the invoice to be evaluated and determine the data format of the invoice to be evaluated; Based on the data format, the invoice to be evaluated is parsed to extract the content data of the invoice to be evaluated. Based on the identification information in the content data, obtain the corresponding business data.

[0156] In one embodiment, the feature extraction module 802 is further configured to perform character recognition on the content data to obtain invoice text data; Fields are extracted from the document text data to obtain the document structure features; Semantic analysis is performed on the document text data to obtain the semantic features of the document; The structural features and semantic features of the bill are fused to obtain the first multimodal feature.

[0157] In one embodiment, the feature extraction module 802 is further configured to perform structural analysis on the business data to obtain business structural features; Perform correlation analysis on business data to obtain business correlation characteristics; Perform semantic analysis on business data to obtain business semantic features; The second multimodal feature is obtained by fusing business structure features, business association features, and business semantic features.

[0158] In one embodiment, the rule determination module 803 is further configured to match the first multimodal feature and the second multimodal feature with the processing rules in the preset processing rule library to obtain a set of candidate processing rules; When the candidate processing rule set contains multiple processing rules, determine the priority information of the multiple processing rules; Based on priority information, the target processing rule is determined from the set of candidate processing rules.

[0159] In one embodiment, the risk assessment module 804 is further configured to perform rule verification on the invoice to be assessed based on the target processing rules, and obtain the rule verification result; Based on the rule verification results, calculate the risk score of the bill to be evaluated; The risk level of the bill to be evaluated is determined based on the risk score and the preset risk level threshold. Based on the rule validation results, risk scores, and risk levels, risk assessment results are generated.

[0160] In one embodiment, the risk assessment module 804 is further configured to obtain audit feedback information corresponding to the risk assessment result; Based on the review feedback information, generate rule update information; The preset processing rule base is updated based on the rule update information.

[0161] This invention provides a bill risk assessment device, comprising: a data acquisition module for acquiring content data of the bill to be assessed and business data associated with the bill to be assessed; a feature extraction module for performing multimodal feature extraction on the content data to obtain a first multimodal feature of the content data, and performing multimodal feature extraction on the business data to obtain a second multimodal feature; a rule determination module for matching target processing rules corresponding to the first and second multimodal features based on a preset processing rule library; and a risk assessment module for performing risk assessment on the bill to be assessed based on the target processing rules to obtain a risk assessment result. This device acquires the content data and associated business data of the invoices to be evaluated, enabling the linked acquisition of invoice and business data. This provides a comprehensive data foundation for subsequent risk assessment, avoiding assessment bias caused by a single data source. By extracting multimodal features from the content and business data respectively, it achieves feature modeling of invoice and business information from multiple dimensions, realizing semantic association between invoice content and business context, and improving the richness and accuracy of feature expression. By matching the target processing rules corresponding to the multimodal features based on a preset processing rule base, it achieves intelligent matching of rules and features, enabling the risk assessment logic to adapt to different invoice types and business scenarios, improving the versatility and flexibility of the solution. By performing risk assessment on the invoices to be evaluated based on target processing rules, it achieves rule-driven intelligent risk assessment, improving the efficiency and accuracy of risk assessment.

[0162] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database is used for data employed in the bill risk assessment method. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a bill risk assessment method.

[0163] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described bill risk assessment method.

[0164] In one embodiment, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described bill risk assessment method.

[0165] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0166] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0167] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0168] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for assessing the risk of negotiable instruments, characterized in that, include: Obtain the content data of the invoice to be evaluated, as well as the business data associated with the invoice to be evaluated; Multimodal feature extraction is performed on the content data to obtain the first multimodal feature of the content data, and multimodal feature extraction is performed on the business data to obtain the second multimodal feature; Based on a preset processing rule library, target processing rules corresponding to the first multimodal feature and the second multimodal feature are matched; Based on the target processing rules, a risk assessment is performed on the invoice to be evaluated, and the risk assessment result is obtained.

2. The method according to claim 1, characterized in that, The acquisition of the content data of the invoice to be evaluated, and the business data associated with the invoice to be evaluated, includes: The format of the invoice to be evaluated is identified to determine the data format of the invoice to be evaluated. Based on the data format, the invoice to be evaluated is parsed to extract the content data of the invoice to be evaluated; Based on the identification information in the content data, the corresponding business data is obtained.

3. The method according to claim 1, characterized in that, The step of performing multimodal feature extraction on the content data to obtain the first multimodal feature includes: Character recognition is performed on the content data to obtain the invoice text data; Fields are extracted from the document text data to obtain document structure features; Semantic analysis is performed on the document text data to obtain document semantic features; The structural features and semantic features of the bill are fused to obtain the first multimodal feature.

4. The method according to claim 1, characterized in that, The step of performing multimodal feature extraction on the business data to obtain the second multimodal feature includes: The business data is subjected to structural analysis to obtain business structure characteristics; Perform correlation analysis on the business data to obtain business correlation characteristics; Perform semantic analysis on the business data to obtain business semantic features; The business structure features, business association features, and business semantic features are fused to obtain the second multimodal features.

5. The method according to claim 1, characterized in that, The step of matching the target processing rules corresponding to the first multimodal feature and the second multimodal feature based on a preset processing rule library includes: The first multimodal feature and the second multimodal feature are matched with the processing rules in the preset processing rule base to obtain a set of candidate processing rules; When the candidate processing rule set contains multiple processing rules, the priority information of the multiple processing rules is determined; Based on the priority information, the target processing rule is determined from the set of candidate processing rules.

6. The method according to claim 1, characterized in that, The step of performing a risk assessment on the invoice to be assessed based on the target processing rules to obtain the risk assessment result includes: Based on the target processing rules, the invoice to be evaluated is subjected to rule verification to obtain the rule verification result; Based on the rule verification results, the risk score of the bill to be evaluated is calculated; Based on the risk score and the preset risk level threshold, the risk level of the bill to be evaluated is determined; The risk assessment result is generated based on the rule verification result, the risk score, and the risk level.

7. The method according to claim 1, characterized in that, After performing a risk assessment on the invoice to be evaluated based on the target processing rules and obtaining the risk assessment result, the process further includes: Obtain the audit feedback information corresponding to the risk assessment results; Based on the audit feedback information, rule update information is generated; Based on the rule update information, the preset processing rule base is updated.

8. A bill risk assessment device, characterized in that, include: The data acquisition module is used to acquire the content data of the invoice to be evaluated, as well as the business data associated with the invoice to be evaluated; The feature extraction module is used to perform multimodal feature extraction on the content data to obtain the first multimodal feature of the content data, and to perform multimodal feature extraction on the business data to obtain the second multimodal feature; The rule determination module is used to match the target processing rule corresponding to the first multimodal feature and the second multimodal feature based on a preset processing rule library; The risk assessment module is used to perform risk assessment on the invoice to be assessed based on the target processing rules, and obtain the risk assessment result.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the bill risk assessment method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the bill risk assessment method as described in any one of claims 1 to 7.

Images