Artificial intelligence-based document recognition management system

Abstract

The AI-based bill identification management system of the present application belongs to the field of artificial intelligence technology, and comprises a candidate paradigm generation unit for collecting image data of bills to be processed and extracting multi-modal features from the image data; the multi-modal features include text content, position coordinates and layout features; the candidate paradigm generation unit is also used to perform semantic analysis on the multi-modal features by using a pre-trained large language model to generate candidate paradigms and calculate the paradigm reasoning confidence of the candidate paradigms; a consistency checking unit is used to generate instantiated data according to the candidate paradigms and compare the instantiated data with a pre-constructed historical business knowledge graph to calculate a cross-document checking rate; the present application can stably process bills of various formats, achieving a substantial breakthrough in technology.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a document recognition management system based on artificial intelligence. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, automated document processing has become a key to improving efficiency and reducing costs for enterprises. However, existing document processing systems still face many challenges in practical applications. Traditional document recognition methods usually rely on pre-set templates, which are inefficient in processing unknown documents with various formats and difficult to cope with challenges brought by format changes. In addition, due to the complexity of document data sources, the extracted information may have format errors or be inconsistent with the actual business logic; traditional methods usually only focus on the verification of internal information of the document, and cannot effectively verify the rationality and consistency of the data in the cross-document business process.

[0003] In addition, existing systems usually lack the ability of continuous learning and self-evolution. Traditional verification and correction processes usually rely on manual intervention and heuristic rules, and it is difficult to form an automated and adaptive closed loop. When there is an error in document recognition, the system cannot accurately attribute the error, nor can it use feedback data from manual correction to optimize its own model. This leads to stagnation of system performance, and the system cannot become more intelligent and accurate as the processing business volume increases, thereby restricting the further improvement of the level of document processing automation.

[0004] The above problems lead to the fact that the existing technology has low recognition accuracy, high business logic error rate, and is difficult to cope with various formats, etc. when processing massive document data, and it is difficult to fully exert the advantages of automated processing.

[0005] In recent years, multi-modal feature fusion technology has gradually become a research hotspot in the field of document recognition. By fusing text, position coordinates and layout information, etc., more rich context information can be provided for semantic understanding of the document, thereby improving the accuracy of recognition.

[0006] In summary, in the existing technology, although deep learning models have shown great ability in processing multi-source data, due to the lack of efficient business logic verification mechanism and continuous adaptive learning ability, it is difficult to solve the problem of high-precision, end-to-end automated processing of unknown documents in zero-shot scenarios.

[0007] The above information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0008] The present application aims to provide an artificial intelligence-based document recognition management system to solve the problems raised in the background.

[0009] The technical solution of the present application is as follows:

[0010] A candidate schema generation unit is configured to collect image data of a document to be processed and extract multi-modal features from the image data, wherein the multi-modal features include text content, position coordinates and layout features; the candidate schema generation unit is further configured to perform semantic analysis on the multi-modal features using a pre-trained large language model to generate a candidate schema and calculate a schema reasoning confidence of the candidate schema;

[0011] A consistency checking unit is configured to generate instantiated data based on the candidate schema and compare the instantiated data with a pre-constructed historical business knowledge graph to calculate a cross-document verification rate;

[0012] An adaptive correction unit is configured to compare the schema reasoning confidence with the cross-document verification rate to determine a deviation type; when it is determined that the schema reasoning is incorrect, the adaptive correction unit is configured to calculate a schema correction update amount to update model parameters used for schema reasoning;

[0013] A hierarchical review unit is configured to fuse the schema reasoning confidence and the cross-document verification rate to generate a comprehensive confidence score; and trigger hierarchical artificial review according to the comprehensive confidence score and a preset threshold; the hierarchical review unit is further configured to feed back the result of the artificial review to the system to fine-tune the pre-trained large language model and the schema correction module.

[0014] Preferably, the process of calculating the schema reasoning confidence by the candidate schema generation unit includes:

[0015] Obtaining the total number of entities in the candidate schema;

[0016] For each entity in the candidate schema, the semantic consistency score and the layout proximity score between the key and the value are calculated respectively;

[0017] The semantic consistency scores and the layout proximity scores of all entities are combined for weighted summation, and then averaged according to the total number of entities to generate the schema reasoning confidence.

[0018] Preferably, the semantic consistency score is determined by inputting the key and the value into a pre-trained natural language reasoning model to obtain the probability of the implied relationship; and the layout proximity score is determined by calculating a normalized layout proximity value based on the geometric position of the text block of the key and the value.

[0019] Preferably, the process of calculating the cross-document verification rate by the consistency checking unit includes:

[0020] Obtaining the total number of core entities extracted from the new document and required for cross-document verification;

[0021] For each entity instance to be verified, the conditional probability under the given historical business knowledge graph is calculated;

[0022] The conditional probabilities of all entity instances to be verified are averaged to generate the cross-document verification rate.

[0023] Preferably, the process of determining the conditional probability comprises:

[0024] Querying the path or value related to the entity instance to be verified in the historical business knowledge graph;

[0025] If the path exists and the value meets the preset business rules, the conditional probability is determined as 1;

[0026] If the path does not exist or the value conflicts with the preset business rules, the conditional probability is determined as 0;

[0027] If the entity instance cannot be directly verified, the conditional probability is set as the estimated value based on similar entity inference.

[0028] Preferably, the process of determining the deviation type by the adaptive correction unit comprises:

[0029] When the paradigm reasoning confidence is high and the cross-document verification rate is low, the deviation type is determined as the business logic error of the data instance;

[0030] When both the paradigm reasoning confidence and the cross-document verification rate are low, the deviation type is determined as the paradigm reasoning error, and the paradigm correction process is triggered.

[0031] Preferably, the process of calculating the paradigm correction update amount by the adaptive correction unit comprises:

[0032] The verification error term is obtained by subtracting the cross-document verification rate from 1;

[0033] A loss function for measuring the degree of inconsistency between the current paradigm element and the historical business knowledge graph is constructed, and the gradient of the loss function with respect to the paradigm element is calculated;

[0034] The preset learning rate, the verification error term, and the gradient are multiplied to generate the paradigm correction update amount.

[0035] Preferably, the process of triggering the hierarchical artificial review by the hierarchical review unit comprises:

[0036] When the comprehensive confidence score is greater than a first preset threshold, the data is automatically processed and archived;

[0037] When the comprehensive confidence score is between the first preset threshold and a second preset threshold, a low confidence field is marked and pushed to artificial review;

[0038] When the comprehensive credibility score is less than a second preset threshold, the entire bill is pushed to a human for comprehensive review.

[0039] Preferably, the hierarchical review unit takes the result of the manual review as labeled data, and uses it to fine-tune the pre-trained large language model and the paradigm correction module, to form a closed-loop iterative optimization.

[0040] The present application provides an artificial intelligence-based bill identification management system, which has the following improvements and advantages compared with the prior art:

[0041] 1. The present application solves the problem that traditional template matching technology cannot adapt to the diversity and dynamic changes of bill formats; the system innovatively sets up a candidate paradigm generation unit, which does not require any preset template, acquires image data of the bill to be processed, and extracts multi-modal features including text content, position coordinates and layout features; with the help of a pre-trained large language model, the system performs deep semantic analysis on these multi-modal features that integrate rich contextual information, and can directly extract structured information from bills of any unknown format under zero-sample conditions, generating candidate paradigms; this method gives the system strong format generalization ability, enabling it to stably process bills of different formats, achieving a fundamental breakthrough in technology;

[0042] 2. The present application establishes a dual, cross-dimensional verification mechanism, significantly improving the accuracy of data recognition and the correctness of business logic; the system not only calculates the paradigm reasoning confidence through the candidate paradigm generation unit, combining the semantic consistency and layout proximity between keys and values, to evaluate the extraction quality from the structural analysis level; but also innovatively introduces a consistency verification unit; this unit deeply compares the instantiated data generated by the candidate paradigm generation unit with a pre-constructed historical business knowledge graph, and calculates the cross-document verification rate; this design goes beyond the verification scope of the integrity of internal information in the bill, and realizes deep logical verification of the association of data business processes, and can identify business logic errors hidden behind the data content that traditional technology cannot find;

[0043] 3.The present application constructs a closed-loop intelligent system capable of self-diagnosis, self-adaptive correction and continuous evolution; the self-adaptive correction unit in the system can accurately determine the deviation type by comparing the combination of the paradigm reasoning confidence and the cross-document verification rate, and intelligently distinguish whether it is a reasoning error of the paradigm itself or a business logic error of the data instance; when it is determined to be a paradigm reasoning error, the system will automatically calculate the paradigm correction update amount to update the model parameters used for paradigm reasoning, realizing the dynamic adjustment of the self-analysis ability; further, the hierarchical review unit takes the final result of manual review as high-quality labeled data, which is fed back to fine-tune the pre-trained large language model and the paradigm correction module, forming a complete closed-loop iterative optimization process of processing, review, feedback and learning, so that the performance of the system can be continuously improved with the growth of business volume;

[0044] 4.The present application designs an efficient and resource-optimized hierarchical artificial review strategy, which maximizes the efficiency of human-machine collaboration; the hierarchical review unit generates a comprehensive confidence score by combining the paradigm reasoning confidence and the cross-document verification rate; according to the relationship between the score and the preset threshold, the system can trigger different levels of processing procedures: high-confidence data is automatically processed and archived, medium-confidence data only pushes low-confidence fields to manual review, and low-confidence data submits the entire document to manual comprehensive review; this intelligent shunting mechanism liberates valuable artificial expert resources from massive, simple and repetitive review work, enabling them to focus on handling complex, ambiguous and high-risk business scenarios, maximizing the automation rate of document processing while strictly controlling key risk points to ensure the highest quality of the final data in the warehouse. BRIEF DESCRIPTION OF DRAWINGS

[0045] The present application will be further explained in conjunction with the accompanying drawings and examples:

[0046] Figure 1 is the flowchart of the system of the present application. DETAILED DESCRIPTION

[0047] To make the purpose, technical scheme and advantages of the present application clearer, the present application will be further explained in detail in conjunction with specific examples.

[0048] Example 1

[0049] Please refer to Figure 1 , the present application provides a document recognition management system based on artificial intelligence, which comprises:

[0050] The candidate schema generation unit is configured to collect image data of a to-be-processed document and extract multi-modal features from the image data, wherein the multi-modal features include text content, position coordinates and layout features; the candidate schema generation unit is further configured to perform semantic analysis on the multi-modal features by using a pre-trained large language model to generate a candidate schema and calculate a schema reasoning confidence of the candidate schema;

[0051] The consistency verification unit is configured to generate instantiated data according to the candidate schema and compare the instantiated data with a pre-constructed historical business knowledge graph to calculate a cross-document verification rate;

[0052] The adaptive correction unit is configured to compare the schema reasoning confidence with the cross-document verification rate to determine a deviation type; when it is determined that the schema reasoning is incorrect, a schema correction update amount is calculated to update model parameters used for schema reasoning;

[0053] The hierarchical review unit is configured to fuse the schema reasoning confidence and the cross-document verification rate to generate a comprehensive confidence score; and trigger hierarchical artificial review according to the comprehensive confidence score and a preset threshold; the hierarchical review unit is further configured to feed back a result of the artificial review to the system to fine-tune the pre-trained large language model and the schema correction module;

[0054] The embodiment provides a document recognition management system based on artificial intelligence; the system constructs a complete technical closed loop from preliminary analysis of unknown documents, cross-document business logic verification, adaptive correction based on feedback to hierarchical artificial review;

[0055] The candidate paradigm generation unit aims to extract structured information from any format of document to be processed without any preset templates or zero samples, and generate a set of the most likely structured data paradigms, i.e., candidate paradigms. In this embodiment, the unit acquires image data of the document to be processed through a high-speed scanner or digital interface. Subsequently, the integrated multimodal feature extraction module starts working. This module uses an optical character recognition engine to extract basic text content, precise position coordinates of each text block, and logical hierarchy of the text blocks from the image data. At the same time, a deep learning-based visual analysis model parses the image data to extract layout features, including but not limited to table borders. The text describes the division of areas such as titles, headers, and footers. The multimodal features here are a collection of information integrating text, location, and layout dimensions, providing rich contextual information for subsequent semantic understanding. After feature extraction, the candidate paradigm generation unit uses a pre-trained large language model to perform deep semantic analysis on the multimodal features. This large language model does not rely on fixed templates but instead uses its natural language understanding capabilities to directly identify potential entities and values ​​from the text content, combining location coordinates and page layout features to generate one or more candidate paradigms. Finally, the unit calculates the paradigm inference confidence score for each candidate paradigm, a quantitative score used to evaluate the reliability of the generated paradigms.

[0056] The consistency verification unit aims to verify whether the data extracted by the candidate paradigm generation unit is reasonable and consistent with the actual business logic. This unit generates instantiated data based on the candidate paradigm, that is, fills in the entities in the candidate paradigm with the identified specific values ​​to form structured data records.

[0057] For example, for a candidate paradigm containing the key "product name" and the value "roller", the unit will construct a data object with the field name "product name" and the field value "roller". This process involves mapping unstructured or semi-structured information identified by the large language model to key-value pairs and converting data types according to a predefined paradigm structure, such as converting numeric text to floating-point format.

[0058] This unit compares these instantiated data with a pre-built historical business knowledge graph. This historical business knowledge graph is a graph database that stores all manually verified correct document data in the company's history. The nodes in the graph represent business entities, and the edges represent business relationships. By querying and reasoning in this graph, this unit can calculate the cross-document verification rate, which quantifies the degree of consistency between the new document data and the historical business process.

[0059] The adaptive correction unit aims to endow the system with the ability to learn and evolve, enabling it to dynamically adjust its parsing strategy for document formats based on verification results. This unit compares the normalization inference confidence level generated by the preceding steps with the cross-document verification rate to determine the type of deviation. Deviation types are divided into business logic errors in data instances or inference errors in the normalization itself. When the deviation type is determined to be a normalization inference error, it means that the system's understanding of the document format is flawed. In this case, the unit will calculate the normalization correction update amount, which is a guiding vector used to update the internal model parameters used for normalization inference.

[0060] The tiered review unit aims to establish an efficient and intelligent human-machine collaborative process to balance automation efficiency and data accuracy. This unit integrates paradigm inference confidence and cross-document verification rate to generate a comprehensive credibility score. This comprehensive credibility score serves as an indicator for comprehensively evaluating the reliability of data processing results and acts as a decision-making basis for triggering human intervention. Subsequently, based on this comprehensive credibility score and preset thresholds, the unit triggers different levels of tiered human review. A core function of this unit is that it also feeds back the results of human review as high-quality labeled data to the system for periodic fine-tuning of the pre-trained large language model and paradigm correction module.

[0061] This embodiment, through the collaborative work of the aforementioned units, realizes an end-to-end automated identification and management system capable of handling documents with unknown formats. It not only solves the problem that traditional template matching methods cannot cope with format changes, but also ensures the business correctness of data through cross-document consistency verification. More importantly, through adaptive correction and closed-loop feedback mechanisms, the system has the ability to continuously learn and evolve, becoming increasingly intelligent and accurate as the volume of business processed increases, significantly improving the automation level of document processing and data quality.

[0062] The process by which the candidate paradigm generation unit calculates the confidence score of paradigm inference includes:

[0063] Get the total number of entities in the candidate paradigm;

[0064] For each entity in the candidate paradigm, calculate the semantic consistency score between the key and the value and the layout proximity score respectively;

[0065] The semantic consistency score and layout proximity score of all entities are combined and weighted and summed, and then averaged according to the total number of entities to generate the paradigm inference confidence score.

[0066] This embodiment is a concretization of the process of calculating the confidence of paradigm inference in the candidate paradigm generation unit described above; the process aims to quantitatively evaluate the quality of candidate paradigms inferred from zero samples through a rigorous mathematical model.

[0067] Calculate the confidence level of paradigm inference The process is as follows: Obtain the total number of entities extracted from the candidate paradigm. And for each entity in the paradigm, calculate its key separately. AND value Semantic consistency score between Proximity score Finally, the system combines the semantic consistency score and layout proximity score of all entities into a weighted sum, and then averages the sum based on the total number of entities to generate the paradigm inference confidence score. The calculation follows the formula:

[0068] ;

[0069] in, The confidence score for paradigm inference is a floating-point number in the range [0, 1]. Its physical meaning is the probability that the candidate paradigm is the correct data paradigm for the document, which is calculated by this formula.

[0070] : Total number of entities, an integer, obtained by counting the number of key-value pairs in the candidate normal form;

[0071] : Loop variable, representing the loop from the first entity to the second. One entity;

[0072] The key of the i-th entity is a string, such as a product name, which is identified from the text content by a pre-trained large language model.

[0073] The value of the i-th entity is a string, such as "roller needle," which is identified from the text content by a pre-trained large language model.

[0074] Semantic consistency score, a floating-point number in the range [0, 1], is calculated by the subsequent natural language inference model;

[0075] Layout proximity score, a floating-point number ranging from [0, 1], is calculated by... and The geometric positional relationship of the corresponding text blocks is obtained;

[0076] : Semantic weight factor, a floating-point number in the range [0, 1], whose value is a hyperparameter, determined by grid search optimization in the calibration dataset; its role is to balance the importance of semantic information and layout information in the final score;

[0077] This embodiment combines semantic matching degree with physical proximity degree to provide a more comprehensive and robust paradigm quality assessment method. Compared with single-dimensional assessment, this method can more accurately determine the correctness of key-value pair matching, significantly improve the accuracy of confidence scoring, and enable the system to reliably distinguish between high-quality candidate paradigms and potentially erroneous paradigms, providing a solid decision-making basis for subsequent automated processing or graded review.

[0078] The semantic consistency score is determined by inputting the key and value into a pre-trained natural language inference model to obtain the probability of implication relations; the layout proximity score is determined by calculating a normalized layout proximity value based on the geometric position of the text blocks of the key and value.

[0079] This embodiment further defines the calculation methods for semantic consistency score and layout proximity score mentioned above, aiming to clarify the specific technical implementation path of these two core scores;

[0080] Semantic consistency score The method for determining this is: to use the key Sum As a pair of sentences, the input is fed into a pre-trained natural language reasoning model; this natural language reasoning model is a deep learning model pre-trained on massive amounts of text data, capable of judging the logical relationship between two sentences; in the specific technical context of this invention, it is used to determine whether a specific value can serve as a reasonable instance of a certain abstract entity; the model outputs a probability value, representing... yes The probability of an implied instance; in this embodiment, the probability of the implied relation is directly obtained as the semantic consistency score;

[0081] To adapt to key-value relationships in document recognition, the model can be pre-trained on a large-scale natural language inference dataset and then fine-tuned on domain-specific data containing document key-value pairs such as product model-model and invoice number-123456, thereby better understanding and judging entity matching relationships in documents.

[0082] Layout proximity score The method of determination is: through key-based AND value The geometric position of each text block is determined by calculating a normalized layout proximity value;

[0083] The formula for calculating the layout proximity score can be specified, for example:

[0084] ;

[0085] in, and These represent the horizontal and vertical distances of the key-value text blocks, respectively. It maps distance to Functions over an interval; It is a function that measures the degree of alignment; This is a weighting factor used to balance the importance of different layout features; and : No. The keys and values ​​of an entity;

[0086] In terms of specific implementation, the system obtains... and The bounding box coordinates of the text blocks are used to calculate the geometric distance between them or to consider their horizontal and vertical alignment. The calculation result is normalized and mapped to an interval of [0, 1] to form the final layout proximity score.

[0087] This embodiment quantifies semantic relevance by employing a natural language inference model and combines it with precise geometric position calculation to quantify layout relevance, which greatly improves the objectivity and accuracy of the paradigm inference confidence calculation. This method makes the scoring no longer dependent on heuristic rules, but based on the deep learning model's understanding of language and geometric analysis of the layout, thus enabling more accurate handling of various complex and fuzzy key-value pair relationships and further enhancing the robustness of the entire system.

[0088] Example 2

[0089] The process by which the consistency check unit calculates the cross-document check rate includes:

[0090] Get the total number of core entities extracted from the new document that require cross-document validation;

[0091] For each entity instance to be verified, calculate the conditional probability under the given historical business knowledge graph.

[0092] The average of the conditional probabilities of all entity instances to be verified is used to generate the cross-document verification rate.

[0093] This embodiment is a concretization of the process of calculating the cross-document verification rate by the consistency verification unit mentioned above. The core is to quantify the logical consistency between the newly extracted document data and the historical business panorama.

[0094] Calculate cross-document validation rate The process is as follows: From the new document after normalization parsing, identify and obtain the total number of core entities that need to be validated across documents. The core entities here refer to those fields that are crucial to the business process and can establish relationships between different documents; subsequently, for each entity instance to be validated... The system calculates its knowledge in a given historical business knowledge graph. Conditional probability under given conditions This probability represents the likelihood that the entity instance conforms to historical business patterns. Finally, the system calculates the average of the conditional probabilities of all entity instances to be verified to generate the cross-document verification rate. The calculation follows the formula: ;

[0095] in, Cross-document verification rate is a floating-point number in the range [0, 1]. Its physical meaning is the overall consistency between new document data and historical business maps, which is calculated by this formula.

[0096] The total number of core entities is an integer, obtained by analyzing the instantiation data of new documents.

[0097] The j-th entity instance to be verified is a data structure extracted from the instantiation data of the new document.

[0098] Historical business knowledge graph is a graph database object that is pre-built from manually verified document data that has been processed in the enterprise's history.

[0099] : Loop variable, representing the sequence from the first core entity to the second core entity. One core entity;

[0100] Conditional probability is a floating-point number with a value in the range [0, 1], obtained by querying and verifying in the graph G;

[0101] This embodiment establishes a quantitative model for measuring the correctness of new data business logic; it transcends the limitations of traditional document information verification and achieves in-depth verification of implicit correlations in business processes by comparing with a dynamically evolving historical knowledge graph; this enables the system to not only detect format errors but also identify deep-seated business logic errors, thereby greatly improving the effectiveness of the final data entering the database.

[0102] The process of determining conditional probability includes:

[0103] Search the historical business knowledge graph for paths or values ​​related to the entity instance to be verified;

[0104] If the path is found and the value matches the preset business rules, the conditional probability is set to 1.

[0105] If the path is not found or the value conflicts with the preset business rules, the conditional probability will be set to 0.

[0106] If entity instances cannot be directly verified, the conditional probability is set to the estimated value inferred from similar entities.

[0107] Similar entities can be defined based on text content, for example, through word vector similarity, entity type, such as belonging to the same product category or their connection relationship in the business graph; the inference process can be completed by calculating the average consistency score between the entity instance to be verified and all similar entities in the historical knowledge graph;

[0108] This embodiment addresses the conditional probability mentioned above. Further refinement of the determination process provides clear and executable rules for calculating this probability;

[0109] The process of determining conditional probability includes the following steps:

[0110] The system uses historical business knowledge graphs Querying and verifying entity instances Related paths or values;

[0111] Make a judgment based on the query results:

[0112] If the path is found and the value matches the preset business rules, the conditional probability is set to 1. The preset business rules here refer to the business constraints encoded into the verification logic, such as the quantity of goods entering the warehouse must be less than or equal to the quantity of goods purchased in the corresponding purchase order.

[0113] If the path is not found or the value conflicts with the preset business rules, the conditional probability will be set to 0.

[0114] Handling special cases that cannot be directly verified:

[0115] If the entity instance cannot be directly verified, the conditional probability is set to a predicted value inferred from similar entities; the predicted value can be a preset neutral value, such as 0.5, or in a preferred embodiment, a dynamic probability value calculated by an auxiliary regression model based on other features associated with the entity.

[0116] The input to the auxiliary regression model can be an instance of the entity to be validated. Other relevant features include document type, supplier name, and transaction amount range; the model outputs a dynamic probability value between 0 and 1, representing the likelihood that the entity instance conforms to business rules given this contextual information; the model can perform supervised learning on historical data to predict the compliance of new data instances;

[0117] This embodiment transforms complex business logic verification into a calculable and repeatable standardized process by defining a clear set of probability assignment rules. This design makes the calculation process of cross-document verification rate transparent and interpretable, which not only ensures the stability of verification results, but also makes the maintenance and expansion of business rules more convenient, providing the system with a strong guarantee of business correctness.

[0118] Example 3

[0119] The process by which the adaptive correction unit determines the type of deviation includes:

[0120] When the confidence level of the paradigm inference is high and the cross-document verification rate is low, the deviation type is determined to be a business logic error of the data instance.

[0121] When both the confidence level of the paradigm reasoning and the cross-document verification rate are low, the deviation type is determined to be a paradigm reasoning error, and the paradigm correction process is triggered.

[0122] This embodiment is a specific implementation of the above-mentioned adaptive correction unit's process for determining the type of deviation. The core is to establish an intelligent diagnostic mechanism to distinguish between different types of identification errors.

[0123] The process of determining the type of bias is based on the confidence level of the paradigm inference. and cross-document validation rate The joint analysis of two key indicators; the threshold used for the judgment was determined based on the receiver operating characteristic curve analysis on the validation set, selecting the working point that maximizes diagnostic accuracy while ensuring high specificity;

[0124] When the confidence level of paradigm inference High cross-document validation rate When the error is low, the system determines that the deviation type is a business logic error of the data instance. The underlying logic of this judgment is that the system is highly confident in the structured parsing of the document format, but the data that is correctly extracted fails to pass the historical business logic verification, so the problem is determined to be in the data itself.

[0125] When the confidence level of paradigm inference Cross-document validation rate When both are low, the system determines the deviation type as a paradigm reasoning error and triggers the paradigm correction process. The underlying logic of this judgment is that the system lacks confidence in the candidate paradigm it generates, and the data extracted by the paradigm cannot pass the business verification. This strongly indicates that the error stems from a deviation in the understanding of the document format.

[0126] The dual-index diagnostic logic designed in this embodiment can accurately attribute errors to their causes. It avoids lumping all verification failures into identification errors, but intelligently distinguishes between two fundamentally different situations: paradigm errors and data content errors. This accurate diagnostic capability is a prerequisite for achieving efficient adaptive learning, ensuring that the system can focus its corrective actions on the paradigm inference stage that truly needs optimization, thereby improving the efficiency and accuracy of the entire adaptive correction process.

[0127] The process of calculating the paradigm correction update amount by the adaptive correction unit includes:

[0128] The verification error term is obtained by subtracting the cross-document verification rate from 1;

[0129] Construct a loss function to measure the degree of inconsistency between the current paradigm elements and the historical business knowledge graph, and calculate the gradient of the loss function with respect to the paradigm elements;

[0130] The preset learning rate, validation error term, and gradient are multiplied together to generate the paradigm correction update.

[0131] This embodiment is an implementation of the above-mentioned adaptive correction unit's process for calculating the paradigm correction update amount; this process employs a mechanism inspired by the gradient descent algorithm to mathematically guide the paradigm inference model on how to perform self-optimization;

[0132] Computational paradigm correction update amount The process is as follows: by cross-document validation rate Subtracting from 1 yields the verification error term. Simultaneously, construct elements for measuring the current paradigm. With historical business knowledge graph Loss function for the degree of inconsistency between them And calculate the loss function for the normal form elements. gradient The loss function here is quantized by the normal form elements. Extracted instance data With knowledge graph A mathematical expression for the degree of conflict;

[0133] This loss function can be designed to penalize entity relationships that do not conform to established business rules in the knowledge graph. For example, it can be defined as a sharp increase in loss when the total amount of a document does not match the sum of the amounts of its constituent items.

[0134] In this embodiment, the loss function It can be defined as a weighted sum of multiple sub-loss terms. For example, for amount verification, the mean squared error loss can be used. For entity existence verification, binary cross-entropy loss can be used. ,in These are the actual labels of entities existing in the knowledge graph, and It predicts probabilities; the final loss function can be expressed as: ;

[0135] Alternatively, when the part number cannot be found in the knowledge graph, the loss value will also be set to a high value; the specific design needs to mathematically quantify the business rules, for example, by using the binary cross-entropy loss function or the mean squared error function to measure inconsistency.

[0136] The gradient specifies the norm element for fastest reduction of this loss. The direction in which the parameters should be adjusted;

[0137] The gradient calculation process can be understood as follows: through the backpropagation algorithm, the loss value caused by inconsistencies in business logic is passed back layer by layer to the elements used to generate the normal form, following a chain rule. The model parameters; for example, if the model incorrectly identifies a product model number as a product quantity, the resulting loss will inform the model through gradient information, prompting it to adjust its internal weights to better identify these two types of entities in the future. This allows the model to learn from business logic errors, thereby correcting its underlying inference strategy;

[0138] Set the preset learning rate The error term and the gradient are multiplied together to generate the normalization correction update. The calculation follows the formula:

[0139] ;

[0140] in, : The paradigm correction update is a vector representing the update of the th candidate paradigm. The correction amount and direction of each element are calculated by this formula;

[0141] The learning rate is a positive scalar hyperparameter. Its initial value is set based on experience and is dynamically adjusted during system operation using an adaptive learning rate algorithm.

[0142] : Verification error term, a floating-point number, is calculated from the cross-document verification rate in the previous step; Paradigm element;

[0143] Instance data, consisting of normalized elements. Extracted; Historical business knowledge graph;

[0144] The loss function is a scalar function that is pre-designed based on specific business requirements.

[0145] The gradient of the loss function with respect to the norm elements is a vector, calculated using the backpropagation algorithm.

[0146] This embodiment provides a precise and automated paradigm correction mechanism. By quantifying the feedback of business verification into error signals and using gradient information to guide the correction direction, the system can iteratively optimize its own paradigm reasoning ability. This mechanism is the core of the system's adaptive learning of unknown paradigms, ensuring that the system can learn from errors and gradually converge to the correct parsing paradigm.

[0147] Example 4

[0148] The process of triggering tiered manual review by a tiered review unit includes:

[0149] When the overall credibility score exceeds the first preset threshold, the data will be automatically processed and archived.

[0150] When the overall confidence score is between the first preset threshold and the second preset threshold, the low confidence field is marked and sent to a human reviewer.

[0151] When the overall credibility score is less than the second preset threshold, the entire document will be sent to a human for full review.

[0152] This embodiment describes a strategy for triggering a tiered manual review process using the aforementioned tiered review unit. This strategy aims to achieve differentiated processing of data with different levels of credibility by setting multiple thresholds.

[0153] The tiered review process is triggered based on a comprehensive credibility score, and a first and a second preset threshold are set. The value of the preset threshold is determined by statistical analysis of a large amount of historical processing data and in combination with the company's risk control requirements.

[0154] For example, in one implementation plan, the first preset threshold can be set to 0.95 and the second preset threshold can be set to 0.8. This means that when the overall confidence score is greater than 0.95, the system will process it automatically; when the score is between 0.8 and 0.95, the system will mark the low confidence field and push it to manual review; when the score is less than 0.8, the entire document will be pushed to full review.

[0155] The triggering logic is as follows:

[0156] When the overall credibility score is greater than the first preset threshold, the system will perform fully automatic processing and directly archive the data, realizing a highly efficient processing mode without human intervention;

[0157] When the overall credibility score is between the first preset threshold and the second preset threshold, the system will mark the specific fields with lower confidence and push only these fields to human reviewers for quick review, thus achieving targeted review and reducing the workload of human reviewers.

[0158] When the overall credibility score is less than the second preset threshold, the system will determine the entire document as high-risk and push it to a human for full review. At this time, the system will present the original document, the recognition result, and highlight the verification failure steps to assist human decision-making.

[0159] This embodiment achieves refined management of review resources by establishing an intelligent hierarchical review strategy. It frees human experts from the attention of massive amounts of simple documents, allowing them to focus on handling complex, ambiguous, or high-risk cases. This human-machine collaboration not only maximizes the overall automation rate of document processing, but also ensures the final data quality through strict control of high-risk data.

[0160] The hierarchical review unit uses the results of manual review as labeled data to fine-tune the pre-trained large language model and paradigm correction module, so as to form a closed-loop iterative optimization.

[0161] This embodiment is a further functional description of the aforementioned hierarchical review unit, focusing on its closed-loop iterative optimization mechanism; this mechanism is the core to ensure that the system can continuously learn and evolve.

[0162] In this embodiment, the hierarchical review unit is the end point of the processing flow and the starting point of a new learning cycle. The implementation method is as follows: the hierarchical review unit records the result of each manual review, whether it is confirmation, correction or rejection, as a high-quality, human expert-verified labeled data.

[0163] This accumulated labeled data was used for:

[0164] Fine-tuning of pre-trained large language models to continuously improve their extraction accuracy and generalization ability in zero-shot scenarios;

[0165] Optimize the paradigm correction module, for example, by adjusting the learning rate. or loss function The design makes its correction process more efficient and stable;

[0166] Through the above methods, the system has constructed a complete closed-loop iterative optimization process of processing, review, feedback, and learning;

[0167] This embodiment establishes a closed-loop feedback mechanism, transforming the tacit knowledge obtained through manual review into explicit data that the model can learn from. This transforms the system from a static tool into a dynamic system capable of self-improvement. Each manual intervention contributes to improving the accuracy of the next automated processing, thus forming a positive cycle. This allows the overall performance of the system to continuously improve over time and with the increase in the amount of data processed, and enables it to cope with future challenges such as highly heterogeneous formats and ambiguous contexts.

[0168] It should be noted that the above 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 preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A document recognition and management system based on artificial intelligence, characterized in that, include: The candidate paradigm generation unit is used to collect image data of the documents to be processed and extract multimodal features from the image data; Multimodal features include text content, location coordinates, and page layout features; The candidate paradigm generation unit is also used to perform semantic analysis on multimodal features using a pre-trained large language model to generate candidate paradigms and to calculate the paradigm inference confidence of the candidate paradigms. The consistency verification unit is used to generate instantiated data based on candidate paradigms and compare the instantiated data with a pre-built historical business knowledge graph to calculate the cross-document verification rate. An adaptive correction unit is used to compare the paradigm inference confidence with the cross-document validation rate to determine the type of bias. When a paradigm inference error is identified, the paradigm correction update amount is calculated to update the model parameters used for paradigm inference. A tiered review unit is used to integrate paradigm reasoning confidence and cross-document verification rate to generate a comprehensive credibility score. The system triggers tiered manual review based on the comprehensive credibility score and preset thresholds. The tiered review unit is also used to feed the results of the manual review back to the system for fine-tuning the pre-trained large language model and paradigm correction module. The process by which the consistency check unit calculates the cross-document check rate includes: Get the total number of core entities extracted from the new document that require cross-document validation; For each entity instance to be verified, calculate the conditional probability under the given historical business knowledge graph. The average of the conditional probabilities of all entity instances to be verified is used to generate the cross-document verification rate.

2. The document recognition and management system based on artificial intelligence according to claim 1, characterized in that, The process by which the candidate paradigm generation unit calculates the confidence score of paradigm inference includes: Get the total number of entities in the candidate paradigm; For each entity in the candidate paradigm, calculate the semantic consistency score between the key and the value and the layout proximity score respectively; The semantic consistency score and layout proximity score of all entities are combined and weighted and summed, and then averaged according to the total number of entities to generate the paradigm inference confidence score.

3. The document recognition and management system based on artificial intelligence according to claim 2, characterized in that, The semantic consistency score is determined by inputting the key and value into a pre-trained natural language inference model to obtain the probability of implication relations; the layout proximity score is determined by calculating a normalized layout proximity value based on the geometric position of the text blocks of the key and value.

4. The document recognition and management system based on artificial intelligence according to claim 1, characterized in that, The process of determining conditional probability includes: Search the historical business knowledge graph for paths or values ​​related to the entity instance to be verified; If the path is found and the value matches the preset business rules, the conditional probability is set to 1. If the path is not found or the value conflicts with the preset business rules, the conditional probability will be set to 0. If entity instances cannot be directly verified, the conditional probability is set to an estimated value inferred from similar entities.

5. The document recognition and management system based on artificial intelligence according to claim 1, characterized in that, The process by which the adaptive correction unit determines the type of deviation includes: When the confidence level of the paradigm inference is high and the cross-document verification rate is low, the deviation type is determined to be a business logic error of the data instance. When both the confidence level of the paradigm inference and the cross-document verification rate are low, the deviation type is determined to be a paradigm inference error, and the paradigm correction process is triggered.

6. The document recognition and management system based on artificial intelligence according to claim 1, characterized in that, The process of calculating the paradigm correction update amount by the adaptive correction unit includes: The verification error term is obtained by subtracting the cross-document verification rate from 1; Construct a loss function to measure the degree of inconsistency between the current paradigm elements and the historical business knowledge graph, and calculate the gradient of the loss function with respect to the paradigm elements; The preset learning rate, validation error term, and gradient are multiplied together to generate the paradigm correction update.

7. The document recognition and management system based on artificial intelligence according to claim 1, characterized in that, The process of triggering tiered manual review by a tiered review unit includes: When the overall credibility score exceeds the first preset threshold, the data will be automatically processed and archived. When the overall confidence score is between the first preset threshold and the second preset threshold, the low confidence field is marked and sent to a human reviewer. When the overall credibility score is less than the second preset threshold, the entire document will be sent to a human for full review.

8. The document recognition and management system based on artificial intelligence according to claim 1, characterized in that, The hierarchical review unit uses the results of manual review as labeled data to fine-tune the pre-trained large language model and paradigm correction module, forming a closed-loop iterative optimization.

Images