Data processing method, device and electronic equipment for cross-report collaboration

Abstract

The application discloses a cross-report collaborative data processing method and device and electronic equipment. It relates to the fields of artificial intelligence and financial technology. The method comprises the following steps: obtaining account data in multiple reports; extracting multi-modal features for each field in the account data; inputting the multi-modal features into a target encoder, and outputting a hidden space vector corresponding to each field through the target encoder; generating a mixed fingerprint for each field according to the hidden space vector, wherein the mixed fingerprint is a binary code obtained by fusing the hidden space vector and context information of each field; performing target retrieval on the mixed fingerprint, and establishing an association matrix between heterogeneous fields in different reports according to the retrieval result; and determining an account checking result between the multiple reports according to the association matrix. The application solves the problem that the existing technology needs manual judgment and arrangement of heterogeneous fields with different field names but similar numerical distributions in different reports, resulting in low cross-report account checking efficiency.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and fintech, and more specifically, to a data processing method, apparatus, and electronic device for cross-report collaboration. Background Technology

[0002] In the accounting process of financial institutions, it is usually necessary to reconcile multiple reports from different business systems, such as personal loan systems, general ledger systems, and cross-border clearing systems, to ensure data consistency.

[0003] In related technologies, accounting reconciliation mainly relies on manually defined rule matching or hard-coded methods based on structured query language scripts. These methods require that the fields used for reconciliation in different reports have completely consistent field names. However, in actual business operations, different business systems often use different field naming conventions for the same type of accounting data, resulting in inconsistent field names for the same data in different reports. When field names are inconsistent, manual judgment and reconciliation rules must be written one by one. As the number of reports and fields increases, the workload of manual intervention rises sharply, leading to low efficiency in cross-report accounting reconciliation.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This application provides a cross-report collaborative data processing method, apparatus, and electronic device to at least solve the technical problem in the prior art that requires manual judgment and sorting of heterogeneous fields with different field names but similar numerical distributions in different reports, resulting in low efficiency of cross-report accounting reconciliation.

[0006] According to one aspect of the embodiments of this application, a cross-report collaborative data processing method is provided, comprising: acquiring accounting data from multiple reports; extracting multimodal features from each field in the accounting data, wherein the multimodal features include at least semantic embedding features and statistical distribution features; inputting the multimodal features into a target encoder, and outputting a latent space vector corresponding to each field through the target encoder, wherein the latent space vector is a feature representation obtained by the target encoder after dimensionality reduction encoding of the multimodal features, used to characterize the numerical distribution structure of each field; generating a hybrid fingerprint of each field based on the latent space vector, wherein the hybrid fingerprint is a binary code obtained by fusing the latent space vector with the context information of each field, used to uniquely identify the data features of each field; performing target retrieval on the hybrid fingerprint, and establishing an association matrix between heterogeneous fields in different reports based on the retrieval results, wherein the target retrieval is used to map each hybrid fingerprint to a hash bucket through multiple hash functions, and to ensure that hybrid fingerprints with similarity greater than a threshold fall into the same hash bucket, and then retrieving hybrid pointer pairs that are neighbors within the same hash bucket; and determining the accounting reconciliation results between multiple reports based on the association matrix.

[0007] According to another aspect of the embodiments of this application, a cross-report collaborative data processing apparatus is also provided, comprising: a data acquisition unit for acquiring accounting data from multiple reports; a feature extraction unit for extracting multimodal features from each field in the accounting data, wherein the multimodal features include at least semantic embedding features and statistical distribution features; an encoding processing unit for inputting the multimodal features into a target encoder and outputting a latent space vector corresponding to each field through the target encoder, wherein the latent space vector is a feature representation obtained by the target encoder after dimensionality reduction encoding of the multimodal features, used to characterize the numerical distribution structure of each field; and a vector processing unit for processing the latent space vector according to the latent space... The vector generation unit generates a hybrid fingerprint for each field, where the hybrid fingerprint is a binary code obtained by fusing the latent space vector with the context information of each field, used to uniquely identify the data features of each field; the retrieval processing unit performs target retrieval on the hybrid fingerprints and establishes an association matrix between heterogeneous fields in different reports based on the retrieval results, where the target retrieval is used to map each hybrid fingerprint to a hash bucket through multiple hash functions, and hybrid fingerprints with similarity greater than a threshold fall into the same hash bucket, and then retrieves hybrid pointer pairs that are neighbors within the same hash bucket; the result determination unit determines the accounting reconciliation results between multiple reports based on the association matrix.

[0008] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described cross-report collaborative data processing method.

[0009] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the above-described cross-report collaborative data processing method.

[0010] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described cross-report collaborative data processing method.

[0011] In this embodiment, the data processing system can simultaneously extract semantic embedding features and statistical distribution features for each field. Semantic embedding features reflect the semantic information of the field name and its context, while statistical distribution features reflect the overall distribution pattern of the numerical sequence under the field. Even if two fields have completely different names, as long as their semantics are similar or their numerical distribution patterns are similar, their multimodal features can exhibit high similarity, which is beneficial for providing quantitative basis for subsequent automatic matching and can reduce subjective judgments by humans regarding the meaning and numerical form of fields. The data processing system then inputs the multimodal features into the target encoder and outputs a dimensionality-reduced latent space vector, which represents the numerical distribution structure of the field. This latent space vector is then fused with the context information of the field to generate a unique binary hybrid fingerprint. This binary hybrid fingerprint encodes the semantics, numerical distribution, and contextual position of the field into a compact binary string, making the fingerprints of two fields from the same origin very close in binary space. This facilitates the automatic digital identification of field identities and reduces the need for manual naming of each field or the creation of mapping tables. Furthermore, the data processing system maps each hybrid fingerprint to a hash bucket using multiple hash functions. Fingerprints with similarity exceeding a threshold fall into the same bucket. Then, it searches for neighboring hybrid pointer pairs within the same bucket, quickly finding matching field pairs. This retrieval method utilizes the similarity of binary fingerprints to automatically filter out fields that should be associated in different reports and establish an association matrix accordingly. This association matrix records the matching relationships between heterogeneous fields, reducing the need for manual field mapping. Finally, based on the established association matrix, the data processing system can automatically complete subsequent numerical comparisons and verifications, reducing manual intervention. Thus, this embodiment transforms the work that originally required manual judgment and organization based on field names and numerical experience into a computer-executable binary fingerprint similarity calculation and rapid retrieval process. Identifying and associating heterogeneous fields in cross-report accounting reconciliation reduces reliance on manual labor, thereby improving reconciliation efficiency. This solves the technical problem of low efficiency in cross-report accounting reconciliation caused by the need for manual judgment and organization of heterogeneous fields with different names but similar numerical distributions in different reports. Attached Figure Description

[0012] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0013] Figure 1 This is a schematic diagram of an optional cross-report collaborative data processing method according to an embodiment of this application;

[0014] Figure 2 This is a flowchart of an optional cross-report collaborative data processing method according to an embodiment of this application;

[0015] Figure 3 This is a schematic diagram of an optional cross-report collaborative data processing apparatus according to an embodiment of this application. Detailed Implementation

[0016] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0017] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] According to an embodiment of this application, a method embodiment of a cross-report collaborative data processing method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0019] It should be noted that the information collected in this application (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for display, data used for analysis, etc.) are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data all comply with relevant laws, regulations, and standards, necessary confidentiality measures have been taken, and they do not violate public order and good morals. Corresponding access points are provided for users to choose to authorize or refuse. For example, interfaces are set up between this system and relevant users or organizations, providing users with corresponding access points to choose to agree to or refuse automated decision-making results; if the user chooses to refuse, the process proceeds to the expert decision-making stage.

[0020] According to the embodiments of this application, a data processing system can be used as the execution subject of the cross-report collaborative data processing method of this application embodiment. The system can be a software system or an embedded system combining software and hardware. Of course, the method execution subject in the embodiments of this application can also be other forms of execution subject, such as devices, equipment, etc. It should be known by those skilled in the art that this application does not particularly limit the specific form of the method execution subject.

[0021] Figure 1 This is a schematic diagram of a cross-report collaborative data processing method according to an embodiment of this application, such as... Figure 1 As shown, this cross-report collaborative data processing method includes the following steps:

[0022] Step S101: Obtain accounting data from multiple reports.

[0023] For example, a report can refer to an accounting record document or data table generated by different business systems, and accounting data can refer to the field names recorded in the report and the numerical sequence corresponding to each field name. By acquiring accounting data from multiple reports, the data processing system can facilitate the subsequent extraction of multimodal features from each field in the accounting data, the generation of hybrid fingerprints, and the establishment of correlation matrices between heterogeneous fields. This helps reduce the workload of manually collecting and organizing multi-source report data and improves the data processing efficiency of cross-report accounting reconciliation.

[0024] In some embodiments, the data processing system can obtain accounting data from multiple reports by reading structured data tables stored in a database. For example, the data processing system can connect to a personal loan system database and a general ledger system database respectively, reading accounting data from the personal loan details report from the personal loan system database and accounting data from the general ledger report from the general ledger system database. This embodiment facilitates obtaining accounting data directly from the source database, ensuring the real-time nature and accuracy of the accounting data obtained by the data processing system.

[0025] In other embodiments, the data processing system can obtain accounting data from multiple reports by parsing spreadsheet files or comma-separated value files. For example, the data processing system can receive multiple report files uploaded by users, parse the format of each report file, and extract the field names and the corresponding numerical sequences of each field name as accounting data. This embodiment facilitates compatibility with report files of different formats and improves the applicability of the data processing system to heterogeneous data sources.

[0026] Step S102: Extract multimodal features from each field in the accounting data, wherein the multimodal features include at least semantic embedding features and statistical distribution features.

[0027] For example, multimodal features can refer to a set of features formed by extracting multiple different types of features from the same field of accounting data. Semantic embedding features can refer to the vectorized representation obtained by encoding the field name and its contextual path in the report using a natural language processing model. Statistical distribution features can refer to high-order statistics obtained by statistical analysis of the numerical sequence under the field, including mean, variance, skewness, and kurtosis. By simultaneously extracting semantic embedding features and statistical distribution features from each field in the accounting data, the data processing system can effectively characterize the field's identity information from both the semantic level of the field name and the level of the field's numerical distribution, thereby improving the accuracy of subsequent generation of latent space vectors and hybrid fingerprints based on multimodal features.

[0028] In some embodiments, the data processing system can use a pre-trained language model to semantically encode the name of each field and the hierarchical path of the report to which the field belongs, obtaining semantic embedding features. Simultaneously, the data processing system can calculate the mean, variance, skewness, and kurtosis of the numerical sequences under each field, combining these statistics into statistical distribution features. The data processing system concatenates the semantic embedding features and statistical distribution features into a multimodal feature vector. This embodiment is advantageous in utilizing semantic information and statistical distribution information to jointly represent field features, improving the accuracy of heterogeneous field identification.

[0029] In other embodiments, the data processing system may further include information entropy features in the multimodal features. Information entropy features refer to the feature values ​​obtained by calculating the Shannon entropy of the numerical sequence under a field, used to characterize the complexity of the field's numerical sequence. The data processing system integrates semantic embedding features, statistical distribution features, and information entropy features into multimodal features; for example, it can combine them into a 512-dimensional multimodal feature vector space, which is beneficial for comprehensively characterizing the field's identity information. This embodiment is advantageous for distinguishing accounting data fields of different complexities through information entropy features, further improving the discriminative power of subsequent fingerprint generation.

[0030] Step S103: Input the multimodal features into the target encoder and output the latent space vector corresponding to each field through the target encoder. The latent space vector is the feature representation obtained by the target encoder after dimensionality reduction encoding of the multimodal features, which is used to characterize the numerical distribution structure of each field.

[0031] For example, the target encoder can refer to the encoder network structure in a pre-trained variational autoencoder. The latent space vector can refer to the feature representation output by the target encoder after mapping the input multimodal features to a low-dimensional latent space. The dimension of the latent space vector is lower than the dimension of the input multimodal features, and it is used to characterize the numerical distribution structure of the fields. The data processing system inputs the multimodal features to the target encoder, and the target encoder outputs the latent space vector corresponding to each field. This helps to compress the original high-dimensional multimodal features into a low-dimensional, compact representation, thereby helping to filter out redundant information and noise interference in the multimodal features and improving the noise resistance and discriminativeness of the subsequently generated hybrid fingerprints.

[0032] Optionally, the training objective of a variational autoencoder includes minimizing the sum of a reconstruction loss function and a distribution difference constraint loss function. The reconstruction loss function measures the difference between the features reconstructed by the decoder and the multimodal features of the input, while the distribution difference constraint loss function constrains the difference between the distribution of the latent space vectors and the standard normal distribution. The combined effect of the reconstruction loss function and the distribution difference constraint loss function ensures that the extracted latent space vectors retain key information from the original features while possessing good noise robustness. For example, the training objective of a variational autoencoder can be to learn a low-dimensional representation of the data by minimizing a loss function. This loss function includes a reconstruction error term and a Kolb-Leibler divergence term. The reconstruction error measures the difference between the features reconstructed by the decoder and the multimodal features of the input, while the Kolb-Leibler divergence constrains the difference between the distribution of the latent space vectors and the standard normal distribution. Through training, the latent space vectors extracted by the target encoder become the numerical fingerprints of the fields, which possess noise robustness.

[0033] In some embodiments, the target encoder may include a neural network composed of multiple fully connected layers. The data processing system passes the multimodal features as input data to the input layer of the target encoder. After nonlinear transformation by each fully connected layer in the target encoder, the latent space vector is obtained at the output layer of the target encoder. This embodiment is advantageous for extracting the most discriminative low-dimensional representation from the multimodal features through the adaptive learning capability of the neural network.

[0034] In other embodiments, the target encoder can output latent space distribution parameters, including a mean vector and a log-variance vector. The data processing system randomly samples from the Gaussian distribution described by the mean and log-variance vectors, using the sampling results as the latent space vector. This embodiment helps to introduce randomness to enhance the generalization ability of the latent space vector and reduces the risk of the target encoder overfitting to specific noise in the training data.

[0035] Step S104: Generate a hybrid fingerprint for each field based on the latent space vector. The hybrid fingerprint is a binary code obtained by fusing the latent space vector with the context information of each field, which is used to uniquely identify the data features of each field.

[0036] For example, a hybrid fingerprint can refer to a binary encoded sequence obtained by fusing latent space vectors with contextual information. Contextual information may include field names, the name of the report to which the field belongs, and the hierarchical path of the field within the report. The data processing system generates a hybrid fingerprint for each field based on the latent space vector, which helps to fuse the numerical distribution structure features and semantic positional features of the field into a compact unique identifier. This ensures that fields belonging to the same type of accounting data in different reports have a smaller distance in the binary encoding space, improving the accuracy of subsequent field matching.

[0037] In some embodiments, the data processing system can perform a secure hash algorithm on the context information to obtain a first hash value, perform a hash quantization operation on the latent space vector to obtain a quantized vector, and then perform an XOR operation on the first hash value and the quantized vector to obtain a hybrid fingerprint. This embodiment is beneficial for uniformly converting discrete text information and continuous numerical vectors into binary encoding form, which facilitates efficient similarity retrieval in the subsequent process.

[0038] In other embodiments, the data processing system can first encode the context information into a binary string, then convert the latent space vector into a binary string, and finally merge the two binary strings into a longer binary string as the hybrid fingerprint according to a preset concatenation order. This embodiment is advantageous for flexibly adjusting the fusion method according to different data types, improving the adaptability of the hybrid fingerprint generation process.

[0039] Step S105: Perform target retrieval on the hybrid fingerprint and establish an association matrix between heterogeneous fields in different reports based on the retrieval results. The target retrieval is used to map each hybrid fingerprint to a hash bucket through multiple hash functions, and to ensure that hybrid fingerprints with similarity greater than a threshold fall into the same hash bucket. Then, hybrid pointer pairs that are neighbors are retrieved within the same hash bucket.

[0040] For example, target retrieval can refer to an approximate nearest neighbor retrieval process based on locality-sensitive hashing. A hash bucket can refer to a storage unit obtained by mapping hybrid fingerprints using a hash function; hybrid fingerprints with a similarity greater than a threshold fall into the same hash bucket. A hybrid pointer pair can refer to two hybrid fingerprints that are neighbors retrieved within the same hash bucket. The data processing system performs target retrieval on hybrid fingerprints and establishes an association matrix between heterogeneous fields in different reports based on the retrieval results. This facilitates the transformation of high-dimensional hybrid fingerprint similarity comparison into fast retrieval within hash buckets, thereby efficiently identifying matching field pairs in a large-scale field set and improving the processing speed of cross-report heterogeneous field associations.

[0041] In some embodiments, the data processing system may employ multiple independent families of locality-sensitive hash functions, each mapping the mixed fingerprint to a hash bucket. The data processing system counts the number of times the mixed fingerprint of each pair of fields falls into the same hash bucket under multiple hash functions, and marks field pairs that fall into the same hash bucket more than a preset threshold as candidate matching pairs. This embodiment helps to reduce the probability of false positive matches and improve the accuracy of target retrieval through a multi-hash function voting mechanism.

[0042] In other embodiments, the data processing system can first distribute the mixed fingerprints into coarse-grained hash buckets using a first-level hash function, and then perform fine-grained bucketing retrieval using a second-level hash function within the set of mixed fingerprints falling into the same coarse-grained hash bucket. The data processing system ultimately uses the mixed pointer pairs that are neighbors within the same fine-grained hash bucket as the retrieval results. This embodiment is advantageous in balancing retrieval speed and accuracy in a multi-level bucketing structure, adapting to the retrieval needs of datasets of different sizes.

[0043] Step S106: Determine the accounting reconciliation results among multiple reports based on the correlation matrix.

[0044] For example, an association matrix can refer to a two-dimensional data structure used to record the matching relationships between heterogeneous fields in different reports. Each element in the association matrix represents whether a field pair is identified as a common source field. The accounting reconciliation result can refer to the judgment result obtained by the data processing system after comparing accounting data in different reports based on the field matching relationships recorded in the association matrix, indicating whether there are differences or no differences. The data processing system determines the accounting reconciliation results between multiple reports based on the association matrix, which facilitates the direct application of the automatic association results of heterogeneous fields to the consistency verification of accounting data. This reduces the workload of manually writing reconciliation rules based on matching relationships and improves the efficiency of generating cross-report accounting reconciliation results.

[0045] In some embodiments, the data processing system can automatically generate numerical comparison rules between each matching field pair based on the field matching relationships recorded in the association matrix, and calculate the difference between the field values ​​in the first report and the corresponding field values ​​in the second report according to the numerical comparison rules. The data processing system compares the difference value with a preset difference threshold. If the difference value is less than or equal to the difference threshold, the accounting reconciliation result of the field pair is determined to be consistent; otherwise, it is considered inconsistent. This embodiment facilitates the automatic execution of numerical comparisons between fields, reducing manual intervention.

[0046] In other embodiments, the data processing system can identify many-to-one matching relationships between multiple fields and a target field based on an association matrix. The system combines the values ​​of the multiple fields according to the operational rules recorded in the association matrix to obtain a combined calculation result. This combined calculation result is then compared with the value of the target field, and the accounting reconciliation result is determined based on the comparison result. This embodiment is advantageous for handling complex one-to-many or many-to-one accounting reconciliation scenarios and improves the coverage of reconciliation rules.

[0047] In some optional embodiments, determining the accounting reconciliation results among multiple reports based on the association matrix includes: the data processing system determining accounting reconciliation rules based on the association matrix; obtaining target state parameters corresponding to the accounting data, wherein the target state parameters include account sensitivity, audit history, and transaction peak; inputting the target state parameters into a reinforcement learning agent, which determines an error adjustment amount based on the target state parameters, wherein the error adjustment amount indicates the adjustment range of the allowable error ratio during the accounting reconciliation process; obtaining a preset initial reconciliation threshold, wherein the initial reconciliation threshold is a pre-set allowable error ratio benchmark value used to determine whether the accounting data matches; adjusting the initial reconciliation threshold according to the error adjustment amount to obtain a dynamic reconciliation threshold; and reconciling the accounting data in multiple reports according to the accounting reconciliation rules and the dynamic reconciliation threshold to obtain the accounting reconciliation results among the multiple reports.

[0048] For example, the accounting reconciliation rules can refer to the numerical comparison logic automatically generated based on the matching relationships of heterogeneous fields recorded in the association matrix. Target state parameters can include account sensitivity, audit history, and transaction peaks. Account sensitivity represents the tolerance of different accounting accounts for deviations, audit history represents the types and frequencies of errors that occurred in past reconciliations, and transaction peaks represent the volume of accounting transactions per unit time. The reinforcement learning agent can refer to a decision model trained based on a proximal policy optimization algorithm, used to output an error adjustment amount based on the input target state parameters. This error adjustment amount represents the magnitude of the increase or decrease in the allowable error ratio. The initial reconciliation threshold can refer to a pre-set allowable error ratio benchmark value, and the dynamic reconciliation threshold refers to the actual comparison threshold obtained after adjusting the initial reconciliation threshold according to the error adjustment amount. The data processing system obtains the error adjustment amount by acquiring the target state parameters and inputting them into the reinforcement learning agent. It then uses the error adjustment amount to dynamically correct the initial reconciliation threshold, enabling the dynamic reconciliation threshold to adaptively adjust according to changes in account sensitivity, historical error patterns, and current transaction load. Based on this, the data processing system combines dynamic verification thresholds with accounting verification rules to perform numerical comparisons, which helps to avoid underreporting of highly sensitive items due to fixed thresholds and a large number of false reports during low-sensitivity peak periods, thereby improving the reliability of accounting verification results.

[0049] In some embodiments, the data processing system compares the current peak transaction volume collected in real time with the historical average peak transaction volume. If the current peak transaction volume exceeds twice the historical average peak transaction volume, the reinforcement learning agent outputs a positive error adjustment, for example, increasing it by 20%. The data processing system calculates the sum of the current peak transaction volume and the positive error adjustment to obtain a first sum value. Then, it multiplies the initial verification threshold by this first sum value to obtain a dynamic verification threshold, ensuring that the dynamic verification threshold is greater than the initial verification threshold. The data processing system then calculates the percentage difference of corresponding fields in different reports according to accounting verification rules. When the percentage difference is less than or equal to the dynamic verification threshold, the verification is considered consistent. This embodiment, by automatically expanding the allowable error range during peak transaction periods, helps reduce invalid difference alarms caused by instantaneous data fluctuations.

[0050] In some embodiments, the account sensitivity in the target state parameters can include two categories: highly sensitive accounts and low-sensitive accounts. When the reinforcement learning agent identifies the currently processed account as a highly sensitive account, it outputs a negative error adjustment, such as lowering it by 10%, so that the dynamic verification threshold is less than the initial verification threshold, thereby implementing a stricter verification standard. When the reinforcement learning agent identifies the currently processed account as a low-sensitive account, it outputs a positive error adjustment, so that the dynamic verification threshold is greater than the initial verification threshold, thereby implementing a more lenient verification standard. The data processing system verifies highly sensitive and low-sensitive accounts separately according to the dynamic verification threshold and accounting verification rules, which helps to reduce the verification cost of non-core accounts while ensuring the accuracy of core account accounting. For example, during the peak period of month-end closing, the volume of accounting data transactions surges and there are many reasonable time differences. The data processing system obtains the current state parameters through the reinforcement learning agent, including high transaction peaks and low account sensitivity, outputs a positive error adjustment, and automatically increases the allowable error ratio threshold for non-sensitive accounts. This adjustment reduces the number of invalid error alarms that were originally caused by minor time differences, thus reducing the workload of manual review.

[0051] For example, in some embodiments, the data processing system can construct a logical execution plan based on a parsing framework. For instance, the data processing system transforms the fingerprint relationships in the generated association matrix into an abstract syntax tree. The data processing system automatically selects the optimal join operation strategy based on the characteristics of the underlying data source. For example, when the data source is a columnar storage database, the data processing system selects a mapping-end join strategy to reduce data transmission overhead; when the data source is a data warehouse, the data processing system selects a shuffling join strategy to utilize distributed computing resources. Through these optimizations, the data processing system can achieve minute-level reconciliation of terabyte-level accounting data.

[0052] In some optional embodiments, after verifying the accounting data in multiple reports according to the accounting verification rules and dynamic verification thresholds to obtain the accounting verification results between the multiple reports, the method further includes: the data processing system can determine a reward value based on the accounting verification results, wherein the reward value is used to characterize the false alarm rate and / or false negative rate in the accounting data verification process; and the reward value is fed back to the reinforcement learning agent to update the network parameters of the reinforcement learning agent.

[0053] For example, the reward value can refer to a numerical value calculated based on the accounting reconciliation results, used to measure the quality of the reconciliation effect caused by the error adjustment amount output by the reinforcement learning agent. The false positive rate can refer to the proportion of actually consistent accounting data that is judged as inconsistent during the accounting reconciliation process. The false negative rate can refer to the proportion of actually inconsistent accounting data that is judged as consistent. The data processing system determines the reward value based on the accounting reconciliation results and feeds the reward value back to the reinforcement learning agent to update the network parameters. This helps the reinforcement learning agent gradually learn the mapping relationship between different target state parameters and the optimal error adjustment amount during multiple reconciliation processes, thereby continuously improving the accuracy of the error adjustment amount output and reducing the false positive and false negative rates in subsequent reconciliation processes.

[0054] In some embodiments, the data processing system counts the number of false alarms and false negatives during the current reconciliation process based on the accounting reconciliation results. The false alarm rate is obtained by dividing the number of false alarms by the total number of reconciliation fields, and the false negative rate is obtained by dividing the number of false negatives by the total number of reconciliation fields. The data processing system can use the weighted sum of the false alarm and false negative rates as a reward value, where the weight corresponding to the false alarm rate is higher than that corresponding to the false negative rate, indicating that reducing false alarms is more important than reducing false negatives. The data processing system inputs the reward value as a feedback signal into the reinforcement learning agent, which calculates the policy gradient and updates the network parameters based on the proximal policy optimization algorithm. This embodiment helps the reinforcement learning agent prioritize learning policies that reduce false alarms, reducing the workload of manually reviewing false alarm warnings.

[0055] In other embodiments, the data processing system determines the reward value solely based on the false negative rate, using the negative of the false negative rate as the reward value. When the false negative rate is zero, the reward value is zero; when the false negative rate is greater than zero, the reward value is negative. The data processing system feeds the reward value back to the reinforcement learning agent. Upon receiving a negative reward value, the reinforcement learning agent updates its network parameters to reduce the probability of generating the current error adjustment. This embodiment helps the reinforcement learning agent prioritize avoiding false negatives and is suitable for reconciliation scenarios with extremely high requirements for accounting consistency.

[0056] In some optional embodiments, multimodal features are input to a target encoder, and the target encoder outputs a latent space vector corresponding to each field. This includes: the data processing system inputting multimodal features into the encoder network of the target encoder; encoding the multimodal features through the encoder network to obtain latent space distribution parameters corresponding to the multimodal features, wherein the latent space distribution parameters include the mean and log-variance of the multimodal features, the mean and log-variance being used to describe the Gaussian distribution of the multimodal features in the latent space; sampling the Gaussian distribution determined by the mean and log-variance, and using the sampling result as the latent space vector.

[0057] For example, the encoder network of the target encoder can refer to the neural network module in a variational autoencoder responsible for mapping input features to latent space distribution parameters. The latent space distribution parameters can include the mean and log-variance, where the mean of the multimodal features describes the central position of the multimodal features in the latent space, and the log-variance describes the dispersion of the multimodal features in the latent space. The mean and log-variance of the multimodal features together determine a Gaussian distribution. The data processing system inputs the multimodal features into the encoder network of the target encoder, encodes the mean and log-variance through the encoder network, and then samples from this Gaussian distribution to obtain the latent space vector. This allows the latent space vector to not only carry a compressed representation of the multimodal features but also introduce controllable randomness, thereby enhancing the generalization ability of the latent space vector and reducing the risk of overfitting the encoder network to specific noise in the training data.

[0058] In some embodiments, the data processing system inputs multimodal feature vectors into an encoder network, which consists of multiple fully connected layers. The last fully connected layer branches into two output heads, outputting a mean vector and a log-variance vector, respectively. The data processing system constructs a multidimensional Gaussian distribution based on the mean and log-variance vectors, and randomly samples once from this distribution, using the sampling result as the latent space vector. This embodiment, by simultaneously outputting distribution parameters through a branching structure, helps maintain the end-to-end training characteristics of the encoder network.

[0059] In other embodiments, the data processing system employs reparameterization techniques to generate latent space vectors when sampling from a Gaussian distribution. For example, the system samples a random noise vector from a standard normal distribution, multiplies this random noise vector by the standard deviation vector obtained by exponential operation on the logarithm and variance, and then adds it to the mean vector to obtain the latent space vector. This embodiment facilitates making the sampling operation differentiable, allowing the gradient of the encoder network to propagate back through the sampling nodes, thus facilitating gradient descent optimization training.

[0060] In some optional embodiments, generating a hybrid fingerprint for each field based on the latent space vector includes: the data processing system obtaining context information for each field, wherein the context information includes the field name, the report name to which the field belongs, and the level to which the field belongs; performing a hash operation on the context information to obtain a hash value; performing hash quantization on the latent space vector to obtain a quantization vector; and performing an XOR operation between the hash value and the quantization vector to obtain a hybrid fingerprint.

[0061] For example, context information may include the field name, the report name containing the field, and the hierarchical path within the report containing the field. Hash operation refers to an algorithmic operation that converts context information into a fixed-length binary string, such as the SHA3-256 secure hash algorithm. Hash value refers to the binary string obtained after performing a hash operation on the context information. Hash quantization refers to a mapping operation that converts continuous values ​​in the latent space vector into discrete binary strings. Quantized vector refers to the binary string obtained after performing hash quantization on the latent space vector. XOR operation refers to a binary operation that performs a bitwise logical XOR operation. Hybrid fingerprint refers to the binary string obtained by performing an XOR operation on the hash value and the quantized vector. By acquiring context information and generating hash values, simultaneously performing hash quantization on the latent space vector to obtain quantized vectors, and then performing an XOR operation on the hash value and the quantized vector to obtain a hybrid fingerprint, the data processing system facilitates the fusion of the textual semantic information and numerical distribution information of a field into a unified binary identifier. This allows fields with similar semantics and similar numerical distributions to have a shorter distance in the binary space, improving the accuracy of subsequent field matching.

[0062] In some embodiments, the data processing system concatenates the field name, the report name containing the field, and the hierarchical path of the field into a string in a preset order, and performs a SHA3-256 hash operation on this string to obtain a 256-bit hash value. The data processing system binarizes the values ​​of each dimension in the latent space vector; for example, it maps dimension values ​​greater than or equal to zero to binary 1s and dimension values ​​less than zero to binary 0s, obtaining a 256-bit quantization vector. The data processing system performs a bitwise XOR operation between the hash value and the quantization vector, and uses the result as a hybrid fingerprint. This embodiment simplifies the quantization vector generation process through binarization, reducing computational complexity.

[0063] In other embodiments, the data processing system employs a family of locality-sensitive hash functions to perform hash quantization on the latent space vector. For example, the data processing system randomly generates multiple random vectors with the same dimension as the latent space vector. For each random vector, it calculates the dot product between the random vector and the latent space vector. If the dot product is greater than or equal to zero, it maps to binary 1; otherwise, it maps to binary 0. The mapping results of all random vectors are combined into a quantized vector. The data processing system then performs an XOR operation between the hash value and the quantized vector to obtain a hybrid fingerprint. This embodiment helps maintain the distance similarity between the quantized vector and the latent space vector, ensuring that quantized vectors corresponding to fields with similar latent space vectors are also similar in Hamming distance.

[0064] Alternatively, the hybrid fingerprint can be a 256-bit binary code, which can be calculated according to the following formula:

[0065] Formula (1);

[0066] in, The final generated hybrid fingerprint is a binary code used to uniquely identify the data characteristics of each field; This indicates a secure hash algorithm used to convert input data into a fixed-length hash value; This indicates the context information of the field, including the field name, the report name in which the field is located, and the level at which the field is located; This represents a bitwise XOR operation, used to perform a bitwise XOR operation between the hash value and the quantization vector to achieve fingerprint fusion; This represents a hash quantization function that converts a latent space vector into a binary string; This represents the low-dimensional latent space vector extracted by the target encoder, used to characterize the numerical distribution structure of the field. This construction method helps to ensure that semantically similar fields with similar numerical distributions have shorter distances in the Hamming space, improving the accuracy of subsequent field matching.

[0067] In some optional embodiments, target retrieval includes the following steps: mapping each hybrid fingerprint to a hash bucket using multiple hash functions; counting the number of times different fields fall into the same hash bucket, and marking field pairs that fall into the same hash bucket more than a preset threshold as candidate matching pairs; calculating the Hamming distance for field pairs in the candidate matching pairs, and determining field pairs with a Hamming distance less than a preset distance threshold as matching field pairs, wherein the Hamming distance is the number of different characters appearing at corresponding positions between two equal-length binary strings; and using the determined matching field pairs as the retrieval results of the target retrieval.

[0068] For example, a hash function can refer to a mapping function that maps hybrid fingerprints to hash buckets, and multiple hash functions constitute a family of hash functions. A hash bucket can refer to a storage unit for the mapping results of a hash function; different hybrid fingerprints may fall into the same hash bucket. A candidate matching pair can refer to a pair of fields that fall into the same hash bucket more than a preset threshold. Hamming distance can refer to the number of different characters appearing at corresponding positions between two binary strings of equal length. A matching field pair can refer to a pair of fields whose Hamming distance is less than a preset distance threshold. The data processing system maps each hybrid fingerprint to a hash bucket using multiple hash functions, counts the number of times different fields fall into the same hash bucket and filters candidate matching pairs, and then calculates the Hamming distance for the candidate matching pairs for precise verification. This is beneficial for quickly eliminating obviously dissimilar field pairs from a large set of fields, and performing precise distance calculations only on a small number of candidate matching pairs. This reduces the amount of retrieval computation while maintaining matching accuracy and improves the retrieval efficiency of heterogeneous field associations.

[0069] In some embodiments, the data processing system may employ a separate set of hash functions, each mapping the mixed fingerprint to a hash bucket identifier. For any two fields of a mixed fingerprint, the data processing system counts the number of times the two mixed fingerprints fall into the same hash bucket under that set of hash functions. If the number of times they fall into the same hash bucket exceeds a preset threshold, for example, more than half the total number of hash functions, then the field pair corresponding to these two fields is marked as a candidate matching pair. This embodiment uses a multi-hash function voting mechanism to filter candidate matching pairs, which helps reduce the number of false positive candidate matching pairs and lowers the overhead of subsequent Hamming distance calculation.

[0070] In other embodiments, the data processing system divides all hash functions into multiple hash function groups, each hash function group independently mapping the mixed fingerprint to a corresponding hash bucket. The data processing system marks field pairs that fall into the same hash bucket in any hash function group as candidate matching pairs. For each candidate matching pair, the data processing system calculates the Hamming distance between the two mixed fingerprints. If the Hamming distance is less than a preset distance threshold, the candidate matching pair is determined as a matching field pair. This embodiment improves the recall rate through grouped hash retrieval, which helps to avoid missing truly matching field pairs due to accidental collisions in a single group of hash functions.

[0071] When processing massive amounts of data, brute-force traversal to calculate the similarity between all field pairs results in a time complexity of O(N²), which is insufficient for practical applications. Data processing systems can employ locality-sensitive hashing (LSH) algorithms based on minimum hashing for efficient retrieval. For example, the system maps each hybrid fingerprint to hash buckets using multiple hash functions. Adjusting the number of hash function combinations and the number of buckets helps balance recall and precision. By appropriately configuring the number of hash function combinations and the number of buckets, the system can achieve second-level matching field pair retrieval in a database of tens of millions of fields.

[0072] In some optional embodiments, after acquiring accounting data from multiple reports, the following steps are further performed sequentially: the data processing system can parse the institution codes of multiple institution nodes; based on the institution codes, construct a cross matrix of administrative levels divided according to the administrative management relationship of the institutions and business levels divided according to the business accounting relationship; construct a directed acyclic graph (DAG) based on the hierarchical position of each institution node in the cross matrix, wherein the DAG is used to represent the asset and liability summary relationship between different institution nodes, and the edges in the DAG represent the flow of funds or business affiliation relationship; generate a dynamic summary path based on the DAG, wherein the dynamic summary path is the node traversal order determined according to the DAG for hierarchically summarizing and verifying accounting data, and the dynamic summary path is used to automatically update the accounting verification path when institution nodes change.

[0073] For example, an institutional node can refer to an organizational unit within a financial institution system, such as a bank, that has independent accounting or management functions; an institutional code can refer to a unique identifier for an institutional node. Administrative hierarchy can refer to a hierarchical structure divided according to the administrative affiliation of institutions; business hierarchy can refer to a hierarchical structure divided according to the business accounting affiliation. A cross matrix can refer to a two-dimensional relationship matrix obtained by cross-combining administrative and business hierarchies. A directed acyclic graph (DAG) can refer to a graph structure with institutional nodes as vertices and directed edges representing fund flows or business affiliations, where no cycles exist. A dynamic aggregation path can refer to a node traversal sequence determined according to the topological order of the DAG, used for hierarchically aggregating and verifying accounting data. The data processing system constructs a cross matrix by parsing the organization code, then builds a directed acyclic graph (DAG) based on the hierarchical position in the cross matrix, and generates a dynamic summary path based on the DAG. This facilitates the integration of administrative management relationships and business accounting relationships into the path planning of accounting summary. When an organization node is merged or transferred, the data processing system can automatically regenerate the dynamic summary path based on the updated DAG, reducing the maintenance cost of manually adjusting and verifying the path. Furthermore, through fingerprint mapping, it helps to improve the reconciliation coverage.

[0074] In some embodiments, the data processing system can read the full organizational information table from the database, parse the organizational code of each organizational node, and the fields representing administrative affiliation and business affiliation within the organizational code. The data processing system constructs an administrative hierarchy tree based on the administrative affiliation field and a business hierarchy tree based on the business affiliation field. Then, it combines the two hierarchy trees using a Cartesian product to obtain a cross matrix of administrative and business levels. The data processing system treats each organizational node as a vertex in a directed acyclic graph (DAG), adding directed edges based on the hierarchical relationships between organizational nodes in the cross matrix to obtain the DAG. The data processing system performs a topological sort on the DAG and uses the node order obtained from the topological sort as a dynamic aggregation path. This embodiment facilitates the automatic parsing of hierarchical relationships through the field positions of the organizational code, improving the automation level of DAG construction.

[0075] In other embodiments, the data processing system can monitor organizational information change events. When it receives a notification of organizational merger or transfer, the system updates the administrative or business level position of the adjusted organizational node in the cross-matrix according to the notification content. The system then reconstructs the directed acyclic graph (DAG) based on the updated cross-matrix and performs a topological sort on the reconstructed DAG to generate an updated dynamic summary path. The system replaces the original dynamic summary path with the updated one for subsequent accounting reconciliation. This embodiment facilitates real-time response of the dynamic summary path to organizational changes, reducing the time window during which the reconciliation path becomes invalid due to organizational adjustments.

[0076] In some optional embodiments, after establishing the association matrix between heterogeneous fields in different reports, the method further includes: when the accounting reconciliation results between multiple reports indicate inconsistencies in the accounting data in different reports, the data processing system inputs the accounting discrepancy data into a gradient boosting decision tree model, and ranks the attributes of the accounting discrepancy data by importance using the gradient boosting decision tree model to obtain the root cause analysis results, wherein the importance ranking is used to arrange the accounting discrepancy data from high to low according to the contribution of each attribute to the discrepancy; based on the root cause analysis results, the source node of the discrepancy is located in the directed acyclic graph, wherein the source node of the discrepancy is the institutional node associated with the accounting discrepancy in the directed acyclic graph; drill-down is performed on the accounting data corresponding to the source node of the discrepancy, wherein the drill-down includes: expanding the lower-level detailed data corresponding to the source node of the discrepancy level by level to locate the specific location where the discrepancy occurs; and a repair suggestion script is generated based on the drill-down results, wherein the repair suggestion script is a code snippet including structured repair instructions, used to automatically execute the correction operation of the accounting discrepancy.

[0077] For example, a gradient boosting decision tree model can refer to a machine learning model that integrates multiple decision trees based on a gradient boosting framework, used to rank the importance of various attributes in accounting discrepancy data. Importance ranking can refer to the result of arranging each attribute from highest to lowest contribution to the discrepancy, with contribution measured by the number of times an attribute is selected as a splitting feature in the gradient boosting decision tree model or the reduction in impurity. The discrepancy root cause analysis result can refer to the importance ranking result and the attributes with the highest contribution. The discrepancy source node can refer to the institutional node associated with the accounting discrepancy in a directed acyclic graph. Breakpoint drill-down can refer to the process of progressively expanding the lower-level detailed data corresponding to the discrepancy source node to locate the specific location where the discrepancy occurred. The repair suggestion script can refer to a code snippet containing structured repair instructions, used to automatically execute the correction operation for the accounting discrepancy. When discrepancies are indicated by accounting reconciliation results, the data processing system inputs the discrepancy data into a gradient boosting decision tree model to rank the attributes by importance to obtain the root cause analysis results. Based on these results, the system locates the source node of the discrepancy from the directed acyclic graph. Drilling down is then performed on the accounting data corresponding to the source node, and finally, a repair suggestion script is generated based on the drill-down results. This process transforms the discrepancy investigation process from manually reviewing reports at each level to an automated process of root cause location and correction suggestion generation, reducing error location time and improving discrepancy processing efficiency.

[0078] In some embodiments, the data processing system uses the organization code, channel identifier, business date, and account code associated with each discrepancy record in the accounting discrepancy data as input attributes, and inputs them into a pre-trained gradient boosting decision tree model. The gradient boosting decision tree model outputs the importance score of each attribute and sorts the importance scores from high to low to obtain the importance ranking result. The data processing system uses the organization code corresponding to the attribute ranked first in the importance ranking result as the target organization code, searches for an organization node matching the target organization code in the directed acyclic graph, and identifies this organization node as the source node of the discrepancy. The data processing system obtains a list of subordinate organization nodes of the source node of the discrepancy from the directed acyclic graph, and reads the detailed accounting data of each subordinate organization node level by level until the specific accounting record that caused the discrepancy is located. Based on the difference between the located accounting record and the correct value, the data processing system generates a structured query language script containing update or reversal statements as a repair suggestion script. This embodiment is beneficial for quickly locating the organizational level and specific accounting record that caused the discrepancy, and automatically generating executable repair instructions, reducing the workload of manually writing repair scripts.

[0079] In other embodiments, the data processing system inputs accounting discrepancy data into a gradient boosting decision tree model to obtain importance ranking results for multiple attributes. The system selects the top three attribute combinations from the importance ranking results and, based on the organization code, channel identifier, and business date within these attribute combinations, simultaneously locates the source node and associated channel of the discrepancy in a directed acyclic graph. For the accounting data corresponding to the source node, the system filters out the lower-level detailed data for the discrepancy period according to the business date and displays the summaries of discrepancy amounts for each channel, grouped by channel identifier. Based on the summaries of discrepancy amounts for each channel, the system generates a script containing multiple repair instructions for reversal or supplementation by channel. This embodiment is beneficial for distinguishing the source of discrepancies in different channels in multi-channel mixed business scenarios and generating repair instructions for each channel, improving the targeting and accuracy of the repair suggestion script.

[0080] In some embodiments, the data processing system can fully record each step of fingerprint matching, rule execution, and tolerance adjustment to generate a data lineage graph. This data lineage graph is used to trace the complete processing chain of accounting data from the original reports to the final verification results, which helps meet the regulatory requirements for data traceability and audit compliance.

[0081] In some optional embodiments, the target encoder is pre-trained through the following training steps: acquiring a training dataset, which includes training multimodal features of multiple sample fields; inputting the training multimodal features into the encoder network, and outputting training latent space distribution parameters corresponding to the training multimodal features through the encoder network, the training latent space distribution parameters including the training mean and training log-variance; sampling from a Gaussian distribution determined by the training mean and training log-variance to obtain a training latent space vector; inputting the training latent space vector into the decoder network, and reconstructing the training multimodal features through the decoder network to obtain reconstructed features; constructing a reconstruction loss function and a distribution difference constraint loss function, wherein the reconstruction loss function is used to measure the difference between the reconstructed features and the input training multimodal features, and the distribution difference constraint loss function is used to constrain the difference between the distribution of the training latent space vector and the preset standard distribution; and iteratively updating the encoder network and decoder network multiple times with the goal of minimizing the sum of the reconstruction loss function and the distribution difference constraint loss function until the convergence condition is met, and then using the encoder network obtained from the last update as the target encoder.

[0082] For example, the training dataset can refer to a set of samples used to train the target encoder, including training multimodal features across multiple sample fields. The encoder network can refer to the neural network module in a variational autoencoder responsible for mapping input features to latent space distribution parameters. The training latent space distribution parameters can include the training mean and training log-variance, where the training mean describes the central location of the training multimodal features in the latent space, and the training log-variance describes the dispersion of the training multimodal features in the latent space. The decoder network can refer to the neural network module in a variational autoencoder responsible for reconstructing the latent space vectors into the original feature space. The reconstructed feature can refer to the feature vector obtained by the decoder network after decoding the training latent space vectors. The reconstruction loss function measures the difference between the reconstructed feature and the input training multimodal features, while the distribution difference constraint loss function constrains the difference between the distribution of the training latent space vectors and a preset standard distribution. The data processing system first acquires a training dataset containing training multimodal features across multiple sample fields. Then, it inputs these training multimodal features into the encoder network to obtain the training mean and training log-variance, which together determine a Gaussian distribution. Next, it samples the training latent space vector from this Gaussian distribution and inputs it into the decoder network to reconstruct the original features. The data processing system can construct a reconstruction loss function and a distribution difference constraint loss function, and jointly iteratively update the encoder and decoder networks with the goal of minimizing their sum. This approach helps to simultaneously ensure reconstruction accuracy and latent space regularity. It also helps the trained target encoder map the input multimodal features into a noise-resistant and regularly distributed latent space vector, which is beneficial for obtaining stable feature representations in subsequent fingerprint generation and improving the error detection rate.

[0083] Optionally, the training objective of a variational autoencoder includes minimizing the sum of a reconstruction loss function and a distribution difference constraint loss function. The reconstruction loss function measures the difference between the features reconstructed by the decoder and the multimodal features of the input, while the distribution difference constraint loss function constrains the difference between the distribution of the latent space vectors and the standard normal distribution. The combined effect of the reconstruction loss function and the distribution difference constraint loss function ensures that the extracted latent space vectors retain key information from the original features while possessing good noise robustness. For example, the training objective of a variational autoencoder can be to learn a low-dimensional representation of the data by minimizing a loss function. This loss function includes a reconstruction error term and a Kolb-Leibler divergence term. The reconstruction error measures the difference between the features reconstructed by the decoder and the multimodal features of the input, while the Kolb-Leibler divergence constrains the difference between the distribution of the latent space vectors and the standard normal distribution. Through training, the latent space vectors extracted by the target encoder become the numerical fingerprints of the fields, which possess noise robustness.

[0084] For example, the loss function of a variational autoencoder can be expressed as:

[0085] Formula (2);

[0086] in, This represents the total loss function value of the variational autoencoder, which is the sum of the reconstruction loss function and the distribution difference constraint loss function. It is used to measure the performance of the variational autoencoder given an input. The training error at that time. The training objective is to minimize this loss function. This represents the model parameters of the decoder network, which is responsible for reconstructing the latent space vectors into the original feature space. This represents the model parameters of the encoder network, which is responsible for mapping the multimodal features of the input to the latent space distribution parameters. This indicates training multimodal features, which are the input data, specifically the multimodal feature vectors of the fields. This represents the training latent space vector, which is the latent space vector sampled from the encoder distribution. This represents the latent space distribution parameters, i.e., the conditional probability distribution of the encoder network output. The reconstruction loss function, also known as the negative reconstruction expectation, is used to measure the difference between the features reconstructed by the decoder and the input. The differences between them. The distribution difference constraint loss function, also known as the KL divergence term, is used to measure the difference between the encoder output distribution and the prior distribution. This indicates a pre-defined standard distribution.

[0087] In some embodiments, the data processing system uses mean squared error as the reconstruction loss function to calculate the average of the squared Euclidean distances between the reconstructed features and the input training multimodal features. The data processing system uses Kolb-Leibler divergence as the distribution difference constraint loss function to calculate the difference between the distribution of the training latent space vectors and the standard normal distribution. The data processing system adds the reconstruction loss function and the distribution difference constraint loss function according to preset weight coefficients to obtain the total loss value. The data processing system aims to minimize the total loss value and uses a stochastic gradient descent optimization algorithm to iteratively update the parameters of the encoder and decoder networks multiple times. When the total loss value no longer decreases after several consecutive iterations, the convergence condition is met, training is stopped, and the encoder network obtained from the last update is used as the target encoder. This embodiment, by weighting and combining the two loss functions, helps to ensure the reconstruction accuracy while making the latent space vector distribution regular, facilitating subsequent hash quantization.

[0088] In other embodiments, the data processing system employs a binary cross-entropy loss function as the reconstruction loss function, suitable for training multimodal features whose values ​​have been normalized and fall within the range of zero to one. The data processing system introduces an adjustable temperature coefficient into the distribution difference constraint loss function to control the strength of the distribution difference constraint. During training, the system dynamically adjusts the temperature coefficient, setting a smaller coefficient in the early stages to reduce the impact of the distribution difference constraint, allowing the encoder network to prioritize learning reconstruction capabilities. In the later stages of training, the temperature coefficient is gradually increased to enhance the distribution regularity constraint. After completing iterative updates, the data processing system uses the encoder network obtained from the last update as the target encoder. This embodiment balances reconstruction accuracy and distribution regularity by dynamically adjusting the temperature coefficient, which is beneficial for adapting to training datasets with different feature distributions.

[0089] For example, in a cross-table alignment application scenario for personal loan interest, the field name in the first report is INT_AMT_CALC, with the unit being cents; the field name in the second report is Interest Income_Personal Loan Category, with the unit being yuan. The data processing system can extract the multimodal features of the two fields and then generate a latent space vector through a target encoder, thereby obtaining a hybrid fingerprint. The variational autoencoder captures that the fluctuation curves of the two sets of numerical sequences almost overlap. The data processing system determines that the Hamming distance between the two hybrid fingerprints is extremely short through target retrieval, thus automatically associating the two fields and establishing a numerical conversion relationship: the interest value in the second report is equal to the sum of the interest values ​​in the first report divided by one hundred.

[0090] For example, in a cross-border trade finance reconciliation application scenario, the trade finance amount to be settled in the head office's financial statements is denominated in RMB, while the statements reported by overseas branches are denominated in the original currency such as USD or EUR. The fingerprint mapping engine of the data processing system identifies the covariance relationship in the numerical distribution of data in different currencies, automatically detects hidden conversion coefficients through latent variable analysis, and determines heterogeneous fields as logically similar without relying on external exchange rate tables, thus achieving cross-currency account reconciliation.

[0091] For example, in the application scenario of interbank netting reconciliation, a single interbank statement corresponds to hundreds of interbank lending records and multiple interest and fee items across multiple branches. The data processing system utilizes the generated feature fingerprints, combined with a multi-indicator association rule mining algorithm, to automatically identify the combination matching relationship between a single external statement and multiple internal debits and credits. For instance, the system identifies a combination feature where one external statement equals 24 internal debits plus 2 credits; the combination vector in the fingerprint space highly matches the target field, thus completing the netting reconciliation in complex netting scenarios.

[0092] Figure 2A flowchart illustrating a cross-report collaborative data processing method provided in an embodiment of this application is shown. Figure 2 As shown, the method includes the following steps:

[0093] Step 1, Adaptive Hierarchical Construction: The data processing system first acquires accounting data from multiple reports, then parses the organizational codes of multiple organizational nodes, constructing a cross matrix between administrative levels (divided according to organizational administrative management relationships) and business levels (divided according to business accounting relationships), and generates a directed acyclic graph (DAG) based on this cross matrix. Subsequently, a dynamic summary path is generated based on the DAG for subsequent hierarchical summary and verification of accounting data.

[0094] Step 2, heterogeneous data fingerprint mapping, involves the data processing system extracting multimodal features for each field. These multimodal features include at least semantic embedding features and statistical distribution features. Next, the multimodal features are input into the target encoder, which outputs the latent space vector for each field. The latent space vector is then fused with the context information of each field to generate a hybrid fingerprint for each field. Afterward, locality-sensitive hashing (LSH) target retrieval is performed on the hybrid fingerprint, and an association matrix between heterogeneous fields in different reports is established based on the retrieval results.

[0095] Step 3 involves the dynamic execution of collaborative verification rules. The data processing system determines the accounting verification rules based on the association matrix established in Step 2, and simultaneously acquires the target state parameters corresponding to the accounting data, including account sensitivity, audit history, and transaction peaks. These target state parameters are input into the reinforcement learning agent, which then outputs a dynamic verification threshold. Finally, based on the accounting verification rules and the dynamic verification threshold, distributed verification is performed on the accounting data across multiple reports to obtain the accounting verification results across the various reports.

[0096] Then, branch judgment and consistency verification can be performed. If the verification passes, an audit lineage report is generated, recording each step of fingerprint matching, rule execution, and tolerance adjustment, and the process ends. If the verification fails, i.e., there is a discrepancy, then proceed to step 4 for intelligent root cause diagnosis of the error.

[0097] Step 4, Intelligent Root Cause Diagnosis: The data processing system inputs the accounting discrepancy data into a gradient boosting decision tree model. This model ranks the attributes of the accounting discrepancy data by importance, obtaining the root cause analysis results. Based on these results, the source node of the discrepancy is located in the directed acyclic graph constructed in Step 1. Drill-down is then performed on the accounting data corresponding to this source node, expanding the lower-level detailed data level by level to pinpoint the specific location of the discrepancy. A repair suggestion script, such as a structured query language script, is generated based on the drill-down results to automatically execute the correction of the accounting discrepancy. After correction, the updated information is returned to Step 1 through feedback or manual confirmation mechanisms to rebuild or update the directed acyclic graph and dynamic summary path, thereby achieving closed-loop optimization.

[0098] See Figure 3 According to another aspect of the embodiments of this application, a cross-report collaborative data processing apparatus is also provided, including: a data acquisition unit, a feature extraction unit, an encoding processing unit, a vector processing unit, a retrieval processing unit, and a result determination unit.

[0099] The system comprises the following components: a data acquisition unit for acquiring accounting data from multiple reports; a feature extraction unit for extracting multimodal features from each field of the accounting data, wherein the multimodal features include at least semantic embedding features and statistical distribution features; an encoding processing unit for inputting the multimodal features into a target encoder, which outputs a latent space vector corresponding to each field, wherein the latent space vector is a feature representation obtained by the target encoder after dimensionality reduction encoding of the multimodal features, used to characterize the numerical distribution structure of each field; a vector processing unit for generating a hybrid fingerprint for each field based on the latent space vector, wherein the hybrid fingerprint is a binary code obtained by fusing the latent space vector with the context information of each field, used to uniquely identify the data features of each field; a retrieval processing unit for performing target retrieval on the hybrid fingerprints and establishing an association matrix between heterogeneous fields in different reports based on the retrieval results, wherein the target retrieval is used to map each hybrid fingerprint to a hash bucket using multiple hash functions, and hybrid fingerprints with similarity greater than a threshold fall into the same hash bucket, and then retrieve hybrid pointer pairs that are neighbors within the same hash bucket; and a result determination unit for determining the accounting reconciliation results between multiple reports based on the association matrix.

[0100] Optionally, the result determination unit includes: a rule determination subunit, used to determine accounting reconciliation rules based on the association matrix; a state acquisition subunit, used to acquire target state parameters corresponding to the accounting data, wherein the target state parameters include account sensitivity, audit history, and transaction peak; an error adjustment subunit, used to input the target state parameters into the reinforcement learning agent, and the reinforcement learning agent determines the error adjustment amount based on the target state parameters, wherein the error adjustment amount is used to indicate the adjustment range of the allowable error ratio during the accounting reconciliation process; a threshold acquisition subunit, used to acquire a preset initial reconciliation threshold, wherein the initial reconciliation threshold is a pre-set allowable error ratio benchmark value used to determine whether the accounting data matches; a threshold adjustment subunit, used to adjust the initial reconciliation threshold according to the error adjustment amount to obtain a dynamic reconciliation threshold; and a reconciliation execution subunit, used to reconcile the accounting data in multiple reports according to the accounting reconciliation rules and the dynamic reconciliation threshold to obtain the accounting reconciliation results between the multiple reports.

[0101] Optionally, the cross-report collaborative data processing device further includes: a reward determination unit, used to determine a reward value based on the accounting verification results after the verification execution subunit obtains the accounting verification results between multiple reports, wherein the reward value is used to characterize the false alarm rate and / or false alarm rate in the accounting data verification process; and a parameter update unit, used to feed the reward value back to the reinforcement learning agent to update the network parameters of the reinforcement learning agent.

[0102] Optionally, the encoding processing unit includes: an input subunit for inputting multimodal features into the encoder network of the target encoder; an encoding subunit for encoding the multimodal features through the encoder network to obtain the latent space distribution parameters corresponding to the multimodal features, wherein the latent space distribution parameters include the mean and log-variance of the multimodal features, the mean and log-variance being used to describe the Gaussian distribution of the multimodal features in the latent space; and a sampling subunit for sampling the Gaussian distribution determined by the mean and log-variance, and using the sampling result as the latent space vector.

[0103] Optionally, the vector processing unit includes: a context acquisition subunit for acquiring context information for each field, wherein the context information includes the field name, the report name to which the field belongs, and the level to which the field belongs; a hash operation subunit for performing a hash operation on the context information to obtain a hash value; a quantization subunit for performing hash quantization on the latent space vector to obtain a quantized vector; and a fusion subunit for performing an XOR operation on the hash value and the quantized vector to obtain a hybrid fingerprint.

[0104] Optionally, the retrieval processing unit includes: a mapping subunit, used to map each hybrid fingerprint to a hash bucket using multiple hash functions; a statistics subunit, used to count the number of times different fields fall into the same hash bucket, and mark field pairs that fall into the same hash bucket more than a preset threshold as candidate matching pairs; a distance calculation subunit, used to calculate the Hamming distance for field pairs in the candidate matching pairs, and determine field pairs with a Hamming distance less than a preset distance threshold as matching field pairs, wherein the Hamming distance is the number of different characters appearing at corresponding positions between two equal-length binary strings; and a result output subunit, used to use the determined matching field pairs as the retrieval results of the target retrieval.

[0105] Optionally, the cross-report collaborative data processing device further includes: a code parsing unit, used to parse the organization codes of multiple organization nodes after the data acquisition unit acquires accounting data from multiple reports; a cross matrix construction unit, used to construct a cross matrix based on the organization codes, dividing administrative levels according to organizational administrative management relationships and business levels according to business accounting relationships; a directed acyclic graph (DAG) construction unit, used to construct a DAG based on the hierarchical position of each organization node in the cross matrix, wherein the DAG is used to represent the asset and liability summary relationship between different organization nodes, and the edges in the DAG represent the flow of funds or business affiliation relationship; and a path generation unit, used to generate a dynamic summary path based on the DAG, wherein the dynamic summary path is the node traversal order determined by the DAG for hierarchically summarizing and verifying accounting data, and the dynamic summary path is used to automatically update the accounting verification path when an organization node changes.

[0106] Optionally, the cross-report collaborative data processing device further includes: a discrepancy analysis unit, used to input accounting discrepancy data into a gradient boosting decision tree model when the accounting reconciliation results between multiple reports indicate inconsistencies in the accounting data in different reports, after the retrieval processing unit establishes the association matrix between heterogeneous fields in different reports; the gradient boosting decision tree model then ranks the attributes of the accounting discrepancy data by importance to obtain the root cause analysis results, wherein the importance ranking is used to arrange the data according to the contribution of each attribute of the accounting discrepancy data to the discrepancy from high to low; and a source location unit. The system is used to locate the source node of the discrepancy in the directed acyclic graph (DAG) based on the root cause analysis results. The source node is the institutional node in the DAG associated with the accounting discrepancy. The drill-down processing unit is used to perform breakpoint drill-down on the accounting data corresponding to the source node of the discrepancy. The breakpoint drill-down includes expanding the lower-level detailed data corresponding to the source node of the discrepancy level by level to locate the specific location where the discrepancy occurred. The script generation unit is used to generate a repair suggestion script based on the drill-down results. The repair suggestion script is a code snippet containing structured repair instructions, which is used to automatically execute the correction operation of the accounting discrepancy.

[0107] Optionally, the cross-report collaborative data processing device further includes a training unit for pre-training the target encoder. The training unit includes: a dataset acquisition subunit for acquiring a training dataset, which includes training multimodal features of multiple sample fields; an encoder network subunit for inputting the training multimodal features into the encoder network and outputting training latent space distribution parameters corresponding to the training multimodal features, the training latent space distribution parameters including the training mean and training log-variance; a latent space sampling subunit for sampling from a Gaussian distribution determined by the training mean and training log-variance to obtain a training latent space vector; and a decoder network subunit for processing the training latent space... The vector input is fed into the decoder network, which reconstructs the training multimodal features to obtain the reconstructed features. The loss construction subunit is used to construct the reconstruction loss function and the distribution difference constraint loss function. The reconstruction loss function measures the difference between the reconstructed features and the input training multimodal features, and the distribution difference constraint loss function constrains the difference between the distribution of the training latent space vectors and the preset standard distribution. The iterative update subunit is used to iteratively update the encoder network and decoder network multiple times with the goal of minimizing the sum of the reconstruction loss function and the distribution difference constraint loss function until the convergence condition is met. The encoder network obtained from the last update is then used as the target encoder.

[0108] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, which stores a computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located performs the above-described cross-report collaborative data processing method.

[0109] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the above-described cross-report collaborative data processing method.

[0110] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program or instructions, which, when executed by a processor, implement the above-described cross-report collaborative data processing method.

[0111] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0112] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0113] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0114] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0115] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0116] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0117] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A cross-report collaborative data processing method, characterized in that, include: Retrieve accounting data from multiple reports; Multimodal features are extracted from each field in the accounting data, wherein the multimodal features include at least semantic embedding features and statistical distribution features; The multimodal features are input into the target encoder, and the target encoder outputs the latent space vector corresponding to each field. The latent space vector is the feature representation obtained by the target encoder after dimensionality reduction encoding of the multimodal features, which is used to characterize the numerical distribution structure of each field. A hybrid fingerprint for each field is generated based on the latent space vector, wherein the hybrid fingerprint is a binary code obtained by fusing the latent space vector with the context information of each field, used to uniquely identify the data features of each field; Target retrieval is performed on the hybrid fingerprint, and an association matrix between heterogeneous fields in different reports is established based on the retrieval results. The target retrieval is used to map each hybrid fingerprint to a hash bucket through multiple hash functions, and to make hybrid fingerprints with similarity greater than a threshold fall into the same hash bucket. Then, hybrid pointer pairs that are neighbors are retrieved in the same hash bucket. The accounting reconciliation results among the multiple reports are determined based on the correlation matrix.

2. The method according to claim 1, characterized in that, Determining the accounting reconciliation results among the multiple reports based on the correlation matrix includes: The accounting reconciliation rules are determined based on the aforementioned correlation matrix; Obtain the target status parameters corresponding to the accounting data, wherein the target status parameters include account sensitivity, audit history, and transaction peak. The target state parameters are input into the reinforcement learning agent, which then determines the error adjustment amount based on the target state parameters. The error adjustment amount is used to indicate the adjustment range of the allowable error ratio during the accounting reconciliation process. Obtain a preset initial verification threshold, wherein the initial verification threshold is a preset allowable error ratio benchmark value used to determine whether accounting data matches; The initial verification threshold is adjusted according to the error adjustment amount to obtain the dynamic verification threshold; Based on the accounting reconciliation rules and the dynamic reconciliation threshold, the accounting data in the multiple reports are reconciled to obtain the accounting reconciliation results between the multiple reports.

3. The method according to claim 2, characterized in that, After verifying the accounting data in the multiple reports according to the accounting reconciliation rules and the dynamic reconciliation threshold to obtain the accounting reconciliation results between the multiple reports, the method further includes: The reward value is determined based on the accounting reconciliation results, wherein the reward value is used to characterize the false alarm rate and / or omission rate in the accounting data reconciliation process; The reward value is fed back to the reinforcement learning agent to update the network parameters of the reinforcement learning agent.

4. The method according to claim 1, characterized in that, The multimodal features are input into the target encoder, which outputs the latent space vector corresponding to each field, including: The multimodal features are input into the encoder network of the target encoder; The multimodal features are encoded by the encoder network to obtain the latent space distribution parameters corresponding to the multimodal features. The latent space distribution parameters include the mean and logarithmic variance of the multimodal features, which are used to describe the Gaussian distribution of the multimodal features in the latent space. The Gaussian distribution determined by the mean and the logarithmic variance is sampled, and the sampling result is used as the latent space vector.

5. The method according to claim 1, characterized in that, Generate a hybrid fingerprint for each field based on the latent space vector, including: Obtain the context information for each field, wherein the context information includes the field name, the report name in which the field is located, and the level in which the field is located; Perform a hash operation on the context information to obtain a hash value; Perform hash quantization on the latent space vector to obtain the quantized vector; The hash value and the quantization vector are XORed to obtain the hybrid fingerprint.

6. The method according to claim 1, characterized in that, The target retrieval includes the following steps: Each of the hybrid fingerprints is mapped to a hash bucket using multiple hash functions; Count the number of times different fields fall into the same hash bucket, and mark field pairs that fall into the same hash bucket more than a preset threshold as candidate matching pairs; For the field pairs in the candidate matching pairs, calculate the Hamming distance, and determine the field pairs whose Hamming distance is less than a preset distance threshold as matching field pairs, wherein the Hamming distance is the number of different characters appearing at corresponding positions between two binary strings of equal length; The determined matching field pairs are used as the search results for the target search.

7. The method according to claim 1, characterized in that, After retrieving accounting data from multiple reports, the following steps are performed sequentially: Parse the organization codes of multiple organization nodes; Based on the institution code, construct a cross matrix of administrative levels divided according to the institution's administrative management relationship and business levels divided according to the business accounting relationship; A directed acyclic graph is constructed based on the hierarchical position of each institutional node in the cross matrix. The directed acyclic graph is used to represent the asset and liability summary relationship between different institutional nodes, and the edges in the directed acyclic graph represent the flow of funds or business affiliation. A dynamic summary path is generated based on the directed acyclic graph, wherein the dynamic summary path is the node traversal order determined by the directed acyclic graph for summarizing and verifying accounting data level by level, and the dynamic summary path is used to automatically update the accounting verification path when the institutional node changes.

8. The method according to claim 7, characterized in that, After establishing the association matrix between heterogeneous fields in different reports, the method further includes: When the accounting reconciliation results among the multiple reports indicate inconsistencies in the accounting data in different reports, the accounting discrepancy data is input into a gradient boosting decision tree model. The gradient boosting decision tree model is used to rank the attributes of the accounting discrepancy data by importance to obtain the root cause analysis results. The importance ranking is used to arrange the accounting discrepancy data from high to low according to the degree of contribution of each attribute of the accounting discrepancy data to the discrepancy. Based on the root cause analysis results, the source node of the difference is located from the directed acyclic graph, wherein the source node of the difference is the institutional node in the directed acyclic graph associated with the accounting difference; Drill down the breakpoints in the accounting data corresponding to the source node of the difference, wherein the drill-down includes: expanding the lower-level detailed data corresponding to the source node of the difference level by level to locate the specific location where the difference occurred. A repair suggestion script is generated based on the drilling results. The repair suggestion script is a code snippet that includes structured repair instructions and is used to automatically perform the correction operation for accounting discrepancies.

9. The method according to claim 1, characterized in that, The target encoder is pre-trained through the following training steps: Obtain a training dataset, which includes training multimodal features for multiple sample fields; The training multimodal features are input into the encoder network, and the encoder network outputs the training latent space distribution parameters corresponding to the training multimodal features. The training latent space distribution parameters include the training mean and the training log variance. The training latent space vector is obtained by sampling from the Gaussian distribution determined by the training mean and the training log-variance; The trained latent space vectors are input into the decoder network, and the trained multimodal features are reconstructed through the decoder network to obtain the reconstructed features; Construct a reconstruction loss function and a distribution difference constraint loss function, wherein the reconstruction loss function is used to measure the difference between the reconstructed feature and the input training multimodal feature, and the distribution difference constraint loss function is used to constrain the difference between the distribution of the training latent space vector and the preset standard distribution; With the goal of minimizing the sum of the reconstruction loss function and the distribution difference constraint loss function, the encoder network and the decoder network are iteratively updated multiple times until the convergence condition is met. The encoder network obtained from the last update is then used as the target encoder.

10. A data processing device for cross-report collaboration, characterized in that, include: The data acquisition unit is used to acquire accounting data from multiple reports; The feature extraction unit is used to extract multimodal features from each field in the accounting data, wherein the multimodal features include at least semantic embedding features and statistical distribution features; The encoding processing unit is used to input the multimodal features into the target encoder and output the latent space vector corresponding to each field through the target encoder. The latent space vector is the feature representation obtained by the target encoder after dimensionality reduction encoding of the multimodal features, which is used to characterize the numerical distribution structure of each field. A vector processing unit is configured to generate a hybrid fingerprint for each field based on the latent space vector, wherein the hybrid fingerprint is a binary code obtained by fusing the latent space vector with the context information of each field, and is used to uniquely identify the data features of each field; The retrieval processing unit is used to perform target retrieval on the hybrid fingerprint and establish an association matrix between heterogeneous fields in different reports based on the retrieval results. The target retrieval is used to map each hybrid fingerprint to a hash bucket through multiple hash functions, and to make hybrid fingerprints with similarity greater than a threshold fall into the same hash bucket. Then, hybrid pointer pairs that are neighbors are retrieved in the same hash bucket. The result determination unit is used to determine the accounting reconciliation results among the multiple reports based on the correlation matrix.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device on which the computer-readable storage medium is located performs the cross-report collaborative data processing method according to any one of claims 1 to 9.

12. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the cross-report collaborative data processing method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the cross-report collaborative data processing method of any one of claims 1 to 9.

Images