Electronic invoice management method based on artificial intelligence
By using an AI-based electronic invoice management method, the problem of low invoice screening efficiency in the existing system has been solved, and precise matching of invoices with bidding requirements has been achieved, thereby improving the efficiency and reliability of bidding and tendering processes.
Patent Information
- Application Number
- CN202511815151.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-03
AI Technical Summary
The existing electronic invoice management system is inefficient in bidding and tendering processes and lacks intelligent matching capabilities, resulting in low invoice screening efficiency and a high risk of errors.
An AI-based electronic invoice management method is adopted. Invoice data is collected and preprocessed from multiple data sources, multi-dimensional features are extracted to generate feature vectors, a pre-trained intelligent matching model is used to calculate matching scores, and the model is optimized by combining user feedback to provide a recommended list of invoices.
It improves the efficiency and accuracy of invoice screening, reduces the workload and error rate of manual operations, enhances the reliability and security of data, and supports the smooth conduct of bidding and tendering business.
Smart Images
Figure CN121597891A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of electronic invoice technology, and in particular to an electronic invoice management method based on artificial intelligence. Background Technology
[0002] In bidding processes, suppliers need to sift through a large number of historical invoices to find those that meet the bidding requirements. Traditional manual screening methods are inefficient and prone to errors.
[0003] While existing electronic invoice management systems have achieved digital storage, they lack the ability to intelligently match data for bidding and tendering scenarios, thus failing to effectively solve the aforementioned problems. Summary of the Invention
[0004] This application provides an artificial intelligence-based electronic invoice management method to solve the problems of low invoice screening efficiency and poor matching accuracy in the prior art.
[0005] This application provides an artificial intelligence-based electronic invoice management method, including: S101, collect electronic invoice data and tender document data from at least one data source, and preprocess them to generate a structured invoice data set and tender requirement data; S102, based on the invoice data set and the bidding demand data, extract multi-dimensional features to obtain the invoice feature vector and the bidding demand feature vector respectively; S103, input each invoice feature vector and bidding requirement feature vector into the pre-trained intelligent matching model, and output the matching score of the invoice; S104: Sort the electronic invoices based on the matching score and output a recommended invoice list to the front-end user interface.
[0006] Preferably, the electronic invoice data includes product information, financial information, time information, and supplier information; The tender document data includes material requirements, financial requirements, time requirements, and qualification requirements.
[0007] Preferably, in step S102, the invoice feature vector specifically includes: A1. Extract product attribute features, including product name, specifications, and technical parameters, denoted as... ; A2. Extract financial attribute features, including invoice amount and tax rate, denoted as... ; A3. Extract time attribute features, including the invoice issuance date, calculate the difference between the invoice date and the current date, and apply an exponential decay function to convert it into a scalar value. ; A4. Extract supplier attribute features, including reputation score, denoted as... ; A5. The product attribute features, financial attribute features, time attribute features, and supplier attribute features extracted from the invoice data set are combined into an invoice feature vector. .
[0008] Preferably, the bidding requirement feature vector specifically includes: B1. Extract commodity attribute features, including the name, specifications, and technical parameters of the tendered goods, denoted as... ; B2. Extract financial attribute features, including the tender budget amount and required tax rate, denoted as... ; B3. Extract time attribute features, including the earliest issuance date within the time range required by the tender, calculate the difference between the date and the current date, and convert it to a scalar value using an exponential decay function. ; B4. The commodity attribute features, financial attribute features, and time attribute features extracted from the bidding demand data are combined into a bidding demand feature vector. .
[0009] Preferably, the pre-trained intelligent matching model specifically comprises: S201 uses historical bidding data with labeled winning bids as the training set, where invoices that were successfully won are used as positive samples and invoices that were not selected are used as negative samples. A pre-selected neural network is used as the model architecture, and the model parameters are continuously optimized. S202 calculates the matching score S using a preset matching model, specifically as follows:
[0010] in, The product similarity score. For financial matching score, Match scores to time. Match scores to suppliers. , , and This indicates the preset weighting coefficients, which are based on preset settings by domain experts.
[0011] Preferably, the product similarity score The cosine similarity is calculated using the following formula:
[0012] in, The product attribute features that represent the invoice feature vector. The product attribute features represent the feature vector of bidding demand.
[0013] Preferably, the financial matching score The following formula is used for calculation:
[0014] in, The invoice amount represents the financial attribute feature of the invoice feature vector. This represents the bidding budget amount within the financial attribute features of the bidding demand feature vector. Indicates the tax rate on the invoice. This indicates the tax rate required for the tender.
[0015] Preferably, the time matching score The following formula is used for calculation:
[0016] in, Indicates the current date. Indicates the date the invoice was issued. This represents the attenuation coefficient, which can be preset according to the actual situation.
[0017] Preferably, the supplier matching score The following formula is used for calculation:
[0018] in, This indicates the supplier's reputation score. This indicates the highest historical reputation rating.
[0019] Preferably, S104 specifically includes: All invoices are sorted in descending order based on matching scores, and the top quartile invoices and their feature vectors are recommended to the front-end user interface for filtering. The system records user feedback data on recommended invoices in real time, including whether the invoice was selected or not. This feedback data is then added to the training dataset as new label data to retrain the intelligent matching model and achieve adaptive optimization of the model.
[0020] One or more technical solutions provided in this application have at least the following technical effects or advantages: This paper provides a comprehensive and effective solution to the problems of inefficiency and error-proneness in supplier invoice screening during bidding processes, as well as the lack of intelligent matching capabilities in existing electronic invoice management systems. Starting with data collection and preprocessing, it ensures data quality and usability; through multi-dimensional feature extraction and feature vector construction, it lays the foundation for intelligent matching; utilizing a pre-trained intelligent matching model to output matching scores, it achieves accurate quantitative matching of invoices with bidding requirements; based on matching score ranking and user feedback mechanisms, it provides users with recommended invoices and optimizes the model. This improves the efficiency and accuracy of invoice screening, reduces the workload and error rate of manual operations, and enhances data reliability and security, effectively supporting the smooth operation of bidding processes. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating an artificial intelligence-based electronic invoice management method according to an embodiment of the present invention. Detailed Implementation
[0022] To facilitate understanding of the present invention, a more complete description of this application will be given below with reference to the accompanying drawings, which illustrate preferred embodiments of the invention. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to enable a more thorough and complete understanding of the disclosure of the present invention.
[0023] It should be noted that the terms "vertical," "horizontal," "up," "down," "left," "right," and similar expressions used in this article are for illustrative purposes only and do not represent the only possible implementation.
[0024] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.
[0025] Example 1: Figure 1 This is a flowchart illustrating an artificial intelligence-based electronic invoice management method according to an embodiment of the present invention.
[0026] like Figure 1 As shown, an artificial intelligence-based electronic invoice management method includes the following steps: S101: Collect electronic invoice data and tender document data from at least one data source, and preprocess them to generate a structured invoice data set and tender requirement data.
[0027] Collecting e-invoice data: Obtain e-invoice data from data sources such as enterprise ERP systems, tax platforms, and cloud storage interfaces through API interfaces. The formats include PDF, images, JSON, etc. The invoice data fields include product information (product name, specifications, technical parameters), financial information (unit price, amount, tax rate, invoice type), time information (issue date, authentication time), and supplier information (seller name, taxpayer identification number).
[0028] Collect bidding document data: Extract key requirements information from the bidding platform or document, including material requirements (material name, specifications, technical parameters), financial requirements (budget amount, tax rate requirements), time requirements (invoice issuance time range), and qualification requirements (supplier qualification conditions).
[0029] This includes preprocessing electronic invoice data and tender document data, including: OCR (Optical Character Recognition) technology is used to parse images or PDF invoices and convert them into structured text. For example, an open-source OCR engine (such as Tesseract) can be used, and the invoice template can be configured to improve recognition accuracy. Use natural language processing (NLP) techniques (such as named entity recognition) to extract key entities from tender documents. For example, use pre-trained models (such as BERT-based NER models) to identify entities such as material names and specifications. Data cleaning: Remove duplicate invoices, correct format errors (e.g., unify dates to YYYY-MM-DD), and unify units (e.g., convert currency to RMB); Data standardization: Construct a thesaurus or ontology library, normalize product names and specifications (e.g., unify "laptop" as "notebook"), and train the thesaurus based on industry standards (such as GB / T standards) or historical data.
[0030] S102, based on the invoice data set and the bidding demand data, extract multi-dimensional features to obtain the invoice feature vector and the bidding demand feature vector respectively.
[0031] Specifically, based on each invoice dataset, multi-dimensional features are extracted to obtain an invoice feature vector, including: A1. Extract product attribute features, including product name, specifications, and technical parameters, and convert them into vector representations using a word embedding model, denoted as... This is used to represent the semantic features of products, ensuring that the similarity between different products is computable. It should be noted that the word embedding model is pre-trained based on historical invoice text to ensure that semantic similarity is computable; however, this will not be elaborated upon further in this invention.
[0032] A2. Extract financial attribute features, including invoice amount and tax rate. Numerical features (such as amount) are directly normalized to the range [0,1]. Tax rate features: use one-hot encoding, supporting tax rate types such as 6%, 9%, and 13%, denoted as... .
[0033] A3. Extract time attribute features, including the invoice issuance date, calculate the difference between the invoice date and the current date (in days), and apply an exponential decay function to convert it into a scalar value. , , Indicates the date difference. The attenuation coefficient (with a value of 0.001) is used to reflect time sensitivity.
[0034] A4. Extract supplier attribute features, including seller name, number of historical collaborations, and reputation score (a weighted evaluation based on the number of collaborations and fulfillment rate; the two indicators are normalized and weighted to calculate the reputation score). The reputation score is obtained from the enterprise database, normalized to the range [0,1], and denoted as... .
[0035] For example, the credit score formula is: Credit Score = w_c × Normalized Number of Cooperations + w_p × Normalized Performance Rate, where w_c and w_p are weighting coefficients (0.5 each by default), normalized to the range of [0,1]; the number of cooperations and performance rate are queried from the enterprise database.
[0036] A5. The product attribute features, financial attribute features, time attribute features, and supplier attribute features extracted from the invoice database are combined into an invoice feature vector:
[0037] Specifically, based on the bidding demand data, multi-dimensional features are extracted to obtain a bidding demand feature vector, including: B1. Extract product attribute features, including the name, specifications, and technical parameters of the tendered goods, and convert them into vectors of the same dimension using the same word embedding model as the invoice. .
[0038] B2. Extract financial attribute features, including the tender budget amount and required tax rate. Numerical features (such as the budget amount) are normalized to the range [0,1]. Categorical features are converted into vectors of the same dimension using the same one-hot encoding scheme as invoices. .
[0039] B3. Extract time attribute features, including the earliest issuance date within the time range required by the tender, calculate the difference between the date and the current date, and convert it to a scalar value using the same exponential decay function. .
[0040] B4. The commodity attribute features, financial attribute features, and time attribute features extracted from the bidding demand data are combined into a bidding demand feature vector:
[0041] Therefore, by using the same feature extraction method, vectorization model, and encoding scheme, it is ensured that the commodity, financial, and time attributes in the invoice feature vector completely correspond to the corresponding attributes in the bidding requirement feature vector in terms of dimension, scope, and semantics. For example, commodity attribute vectors are all generated using the same word embedding model, and financial attribute vectors all use the same one-hot encoding mapping table.
[0042] S103: Input each invoice feature vector and bidding requirement feature vector into the pre-trained intelligent matching model, and output the matching score of the invoice.
[0043] In some embodiments, the pre-trained intelligent matching model specifically includes: S201 uses historical bidding data with labeled winning bids as the training set, where invoices that were successfully won are used as positive samples and invoices that were not selected are used as negative samples. Gradient boosting decision trees or neural networks are used as the model architecture, and the training objective is to minimize the loss function.
[0044] For example, the training set is divided by time (e.g., 70% training, 20% validation, 10% testing) to avoid data leakage. The model architecture adopts gradient boosting decision tree or neural network. The gradient boosting decision tree is implemented using XGBoost, and the neural network is a multilayer perceptron (with 3 hidden layers and ReLU activation function). The training objective is to minimize the loss function and continuously optimize the model parameters to obtain the final intelligent matching model. The principle architecture of the model can be described with reference to relevant existing technologies, and this invention will not elaborate on it.
[0045] S202 calculates the matching score S using a preset matching model, specifically as follows:
[0046] in, The product similarity score. For financial matching score, Match scores to time. Match scores to suppliers. , , and This represents the preset weight coefficients, which are based on preset settings by domain experts and can be dynamically adjusted later according to model training to optimize matching accuracy. For example, they can be set to 0.4, 0.3, 0.2, and 0.1 respectively.
[0047] Specifically, product similarity score The cosine similarity is calculated using the following formula:
[0048] in, Indicates the product attribute characteristics of the invoice. It represents the product attribute characteristics of the bidding requirements and is used to measure the degree of matching of product attributes.
[0049] Specifically, financial matching score Calculated based on amount and tax rate:
[0050] in, This refers to the invoice amount, which is one of the financial attributes of an invoice. This represents the bidding budget amount within the financial attribute characteristics of the bidding requirements. Indicates the tax rate on the invoice. This indicates the tax rate required for the tender, used to assess financial compliance.
[0051] Specifically, time matching score Calculated based on the invoice date:
[0052] in, Indicates the current date. Indicates the date the invoice was issued. This represents the attenuation coefficient, which is preset according to the actual situation. This score is used to reflect the time effectiveness.
[0053] Specifically, supplier matching score Calculated based on reputation score:
[0054] in, This indicates the supplier's reputation score. This represents the highest historical reputation score, which is used to assess supply chain reliability.
[0055] S104 sorts electronic invoices based on matching scores, outputs a recommended invoice list to the front-end user interface, and records user feedback to update the training dataset for adaptive learning of the intelligent matching model.
[0056] Specifically, all invoices are sorted in descending order based on their matching scores. The top quartile (top 25% of invoices) and their feature vectors are recommended to the front-end user interface for user filtering. User feedback on the recommended invoices, including whether they were selected or not, is recorded and added as new labels to the training dataset to retrain the intelligent matching model, achieving adaptive optimization. The front-end interface displays the invoices in tabular form, including key invoice fields and matching scores for user filtering.
[0057] S105 automatically generates a matching degree analysis report between the electronic invoices in the submitted tender documents and the tender requirements, and verifies the authenticity of the invoices by calling an external system through an interface, and outputs the verification results and the matching degree analysis report together.
[0058] For example, the system automatically compares invoices submitted by suppliers with bidding requirements, generates a visual report highlighting matching points (consistent product specifications) and mismatches (such as exceeding the budget). The report automatically categorizes invoices based on matching scores and preset matching thresholds (e.g., S > 0.7 for a match). It connects to the tax platform via API to verify the authenticity of invoices, uses blockchain technology to store invoice hash values to ensure data immutability, provides judges with a one-click verification function, and automatically sorts out the matching of invoice details with bidding requirements, reducing manual verification time.
[0059] In summary, invoice data was obtained from multiple data sources, covering different formats to ensure data integrity. Key requirement information was extracted from bidding platforms or documents, with clear screening criteria. During preprocessing, OCR technology was used to parse image or PDF invoices to improve data usability, and NLP technology was used to extract key entities from bidding documents to accurately obtain requirements. Data cleaning removed duplicates, corrected errors, and standardized units to ensure data quality. Data standardization normalized product names and specifications to facilitate subsequent comparative analysis. Invoice features were extracted from multiple dimensions such as product, financial, time, and supplier to comprehensively reflect invoice attributes. Similarly, bidding requirement features were extracted from the product, financial, and time dimensions using the same feature extraction methods, vectorization models, and coding schemes to ensure the accuracy of the data. The feature vectors of invoices and bidding requirements correspond in dimension, scope, and semantics, improving matching accuracy. Historical bidding data with labeled winning bids is used to train the model, with successfully won invoices as positive samples and unselected invoices as negative samples. A gradient boosting decision tree or neural network architecture is employed, with the training objective being to minimize the loss function, thus improving model accuracy. Matching scores are calculated using a preset formula, comprehensively considering product similarity, financial matching, time matching, and supplier matching. Weight coefficients can be dynamically adjusted to make the matching results more closely reflect actual needs. Invoices are sorted in descending order of matching scores, and the top quartile invoices are recommended to the front-end interface for easy user filtering. User feedback data is recorded and added to the training dataset, and the model is retrained to achieve adaptive optimization, improving subsequent matching accuracy.
[0060] The technical solutions described in the embodiments of this application have at least the following technical effects or advantages: Intelligent matching mechanism with multi-dimensional feature fusion: Traditional electronic invoice management is mostly based on a single condition (such as amount or date) for screening. By introducing features of four dimensions, namely goods, finance, time and supplier, collaborative analysis and vectorization are carried out, and a comprehensive mathematical model is used to calculate the matching score. The multi-dimensional cross-validation mechanism can more comprehensively and accurately evaluate the fit between invoices and bidding requirements, and solve the technical problems of single perspective and low efficiency of manual screening.
[0061] The dynamic adaptive intelligent matching model does not use static, fixed rules for matching. Instead, it constructs an intelligent matching model based on machine learning and designs a mathematical model that includes dynamic weight parameters. It can not only be trained with historical data to optimize performance, but also learn adaptively based on user feedback during application, continuously iterating and optimizing. This makes the system's recommendation results increasingly accurate, has self-evolution capabilities, and significantly surpasses fixed logic based on preset rules.
[0062] The integrated business closed-loop design throughout the entire process: Traditional technologies usually only solve the problem of invoice storage or single query. By seamlessly embedding invoice management into the entire bidding process, it extends from "intelligent screening and recommendation" on the supplier side to "automated review support" on the judge side, forming a complete solution. In particular, by combining the authenticity verification interface with intelligent matching analysis reports, it solves the two key problems of "accurate selection" and "authentic verification" in one process, which greatly improves the overall efficiency and reliability of the bidding process.
[0063] Therefore, this paper provides a comprehensive and effective solution to the problems of low efficiency and error-proneness in supplier invoice screening during bidding processes, as well as the lack of intelligent matching capabilities in existing electronic invoice management systems. Starting with data collection and preprocessing, it ensures data quality and usability; through multi-dimensional feature extraction and feature vector construction, it lays the foundation for intelligent matching; using a pre-trained intelligent matching model to output matching scores, it achieves accurate quantitative matching of invoices with bidding requirements; based on matching score ranking and user feedback mechanisms, it provides users with recommended invoices and optimizes the model; finally, it automatically generates matching degree analysis reports and verifies the authenticity of invoices, providing decision-making basis for judges. This improves the efficiency and accuracy of invoice screening, reduces the workload and error rate of manual operations, and enhances data reliability and security, effectively supporting the smooth conduct of bidding processes.
[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. For those skilled in the art, the present invention can have various modifications and variations. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An electronic invoice management method based on artificial intelligence, characterized in that, include: S101, collect electronic invoice data and tender document data from at least one data source, and preprocess them to generate a structured invoice data set and tender requirement data; S102, based on the invoice data set and the bidding demand data, extract multi-dimensional features to obtain the invoice feature vector and the bidding demand feature vector respectively; S103, input each invoice feature vector and bidding requirement feature vector into the pre-trained intelligent matching model, and output the matching score of the invoice; S104: Sort the electronic invoices based on the matching score and output a recommended invoice list to the front-end user interface.
2. The electronic invoice management method based on artificial intelligence as described in claim 1, characterized in that, The electronic invoice data includes product information, financial information, time information, and supplier information; The tender document data includes material requirements, financial requirements, time requirements, and qualification requirements.
3. The electronic invoice management method based on artificial intelligence as described in claim 2, characterized in that, In step S102, the invoice feature vector specifically includes: A1. Extract product attribute features, including product name, specifications, and technical parameters, denoted as... ; A2. Extract financial attribute features, including invoice amount and tax rate, denoted as... ; A3. Extract time attribute features, including the invoice issuance date, calculate the difference between the invoice date and the current date, and apply an exponential decay function to convert it into a scalar value. ; A4. Extract supplier attribute features, including reputation score, denoted as... ; A5. The product attribute features, financial attribute features, time attribute features, and supplier attribute features extracted from the invoice data set are combined into an invoice feature vector. .
4. The electronic invoice management method based on artificial intelligence as described in claim 3, characterized in that, The bidding requirement feature vector specifically includes: B1. Extract commodity attribute features, including the name, specifications, and technical parameters of the tendered goods, denoted as... ; B2. Extract financial attribute features, including the tender budget amount and required tax rate, denoted as... ; B3. Extract time attribute features, including the earliest issuance date within the time range required by the tender, calculate the difference between the date and the current date, and convert it to a scalar value using an exponential decay function. ; B4. The commodity attribute features, financial attribute features, and time attribute features extracted from the bidding demand data are combined into a bidding demand feature vector. .
5. The artificial intelligence-based electronic invoice management method as described in claim 4, characterized in that, The pre-trained intelligent matching model is specifically as follows: S201 uses historical bidding data with labeled winning bids as the training set, where invoices that were successfully won are used as positive samples and invoices that were not selected are used as negative samples. A pre-selected neural network is used as the model architecture, and the model parameters are continuously optimized. S202 calculates the matching score S using a preset matching model, specifically as follows: in, The product similarity score. For financial matching score, Match scores to time. Match scores to suppliers. , , and This represents the preset weighting coefficients, which are pre-set based on domain experts.
6. The electronic invoice management method based on artificial intelligence as described in claim 5, characterized in that, Product similarity score The cosine similarity is calculated using the following formula: in, The product attribute features that represent the invoice feature vector. The product attribute features represent the feature vector of bidding demand.
7. The electronic invoice management method based on artificial intelligence as described in claim 6, characterized in that, The financial matching score The following formula is used for calculation: in, The invoice amount represents the financial attribute feature of the invoice feature vector. This represents the bidding budget amount within the financial attribute features of the bidding demand feature vector. Indicates the tax rate on the invoice. This indicates the tax rate required for the tender.
8. The electronic invoice management method based on artificial intelligence as described in claim 7, characterized in that, The time-matching score The following formula is used for calculation: in, Indicates the current date. Indicates the date the invoice was issued. This represents the attenuation coefficient, which can be preset according to the actual situation.
9. The electronic invoice management method based on artificial intelligence as described in claim 8, characterized in that, Supplier matching score The following formula is used for calculation: in, This indicates the supplier's reputation score. This indicates the highest historical reputation rating.
10. The electronic invoice management method based on artificial intelligence as described in claim 1, characterized in that, S104 specifically includes: All invoices are sorted in descending order based on matching scores, and the top quartile invoices and their feature vectors are recommended to the front-end user interface for filtering. The system records user feedback data on recommended invoices in real time, including whether the invoice was selected or not. This feedback data is then added to the training dataset as new label data to retrain the intelligent matching model and achieve adaptive optimization of the model.