Automatic tax declaration and recheck system

Abstract

The automatic tax declaration and recheck system comprises a character recognition module, a process automation module, a dynamic rule engine module, a block chain evidence storage module and a distributed database. The character recognition module forms a tax field signal, a structured signal and a semantic tag signal. And the process automation module forms a tax declaration form according to the tax field signal. And the dynamic rule engine module receives the structured signal, performs multi-dimensional verification through a machine learning model, and forms a verification signal. The block chain evidence storage module comprises a distributed database, and the block chain evidence storage module receives the semantic tag signal and forms a unique hash value. And the distributed database performs feature fusion on the unique hash value, the tax declaration form and the verification signal to form a declaration signal and stores the declaration signal in the distributed database. According to the automatic tax declaration and rechecking system, the problem that structural analysis of multi-format tax documents is difficult can be solved.

Description

Technical Field

[0001] The present invention relates to the intersection of tax informatization and artificial intelligence, and specifically to an automated tax declaration and review system. Background Art

[0002] The current tax declaration process generally faces pain points such as inefficient processing of multi-source heterogeneous documents, difficulty in dynamic adaptation of policies and rules, and insufficient data credibility. Traditional technical solutions often adopt isolated technical means - such as relying solely on optical character recognition (OCR) to process paper documents, rule verification systems based on fixed logical conditions, or independently deployed blockchain evidence platforms, resulting in obvious technical gaps in each link.

[0003] The existing OCR technology has an accuracy rate of less than 85% for complex format documents (such as VAT returns containing nested tables and handwritten annotations), and lacks semantic error correction capabilities. Subsequent declaration errors are often caused by problems such as decimal point misidentification and field misalignment. The rule verification system is limited by manually maintained static logical conditions and has difficulty responding to policy updates such as tax rate adjustments and changes in declaration items in a timely manner. Cross-field correlation verification (such as the cross-checking relationship between input and output tax amounts) lacks intelligent processing capabilities. Centrally stored declaration data is at risk of tampering, and audit tracing requires manual verification of a large number of original vouchers, which is inefficient and prone to disputes. Summary of the Invention

[0004] In view of the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide an automated tax declaration and review system to solve the problem of difficult structured parsing of multi-format tax documents. The present invention breaks through the bottleneck of document processing through multimodal joint parsing technology, deeply integrates the image semantic segmentation network with the natural language processing model, and simultaneously extracts the layout features such as table coordinates and seal areas and text semantic information during the OCR recognition process. Combined with the context-related hybrid decision-making error correction mechanism, the recognition accuracy of non-standard documents containing handwriting and complex tables is improved, and enhanced digital documents containing error correction track marks are generated. Secondly, a dynamic rule engine with a three-layer architecture is designed. On the basis of retaining the certainty of Boolean logic rule verification, the isolation forest algorithm is introduced to detect abnormal numerical distribution, and the multi-head attention model based on Transformer is used to mine implicit relationships across fields. For example, the risk of false invoicing is identified by analyzing the deviation of the cost ratio from the industry benchmark. The online incremental learning mechanism is combined to realize the automatic capture of policy changes, which shortens the model update delay from several weeks in traditional solutions to hours.

[0005] The present invention provides an automated tax declaration and review system, comprising:

[0006] A character recognition module receives tax documents in multiple formats, recognizes the tax documents through an optical character recognition device, and generates digitized documents. The module then extracts features from the digitized documents to generate tax field signals, structured signals, and semantic tag signals.

[0007] Process automation module, which generates tax declaration forms based on tax field signals;

[0008] Dynamic rule engine module: The dynamic rule engine module receives structured signals, performs multi-dimensional verification through machine learning models, and generates verification signals;

[0009] The blockchain evidence storage module includes a distributed database, which receives semantic tag signals and generates a unique hash value.

[0010] Distributed database, the distributed database combines the unique hash value, tax declaration form and verification signal to form a declaration signal and stores it in the distributed database. The distributed database reviews the declaration signal based on the historical declaration signal. After the review is correct, the control process automation module submits the tax declaration form.

[0011] In one embodiment of the present invention, the simulation engine of the fire simulation module includes a spatiotemporal convolutional neural network architecture, and the character recognition module includes starting an image semantic segmentation unit and a natural language processing unit. The character recognition module adopts a multimodal feature fusion algorithm, and synchronously starts the image semantic segmentation unit and the natural language processing unit during the optical character recognition process. The image semantic segmentation unit identifies the document layout structure features through a convolutional neural network, and generates a layout vector containing table positioning coordinates, seal area contours and signature handwriting features. The natural language processing unit uses a bidirectional long short-term memory network to establish a context association model, and implements probabilistic error correction for misidentified characters. When the confidence difference between the original recognition result and the semantic prediction result of the character recognition module exceeds a threshold, a hybrid decision mechanism is triggered, and the character weight is dynamically adjusted in combination with the positioning information in the layout vector to form an enhanced digital document containing error correction trajectory marks. The feature extraction process adopts a staged attention mechanism, first performing keyword enhanced extraction on the document title area to generate a tax field signal, then performing double verification on the amount data to generate a structured signal, and finally generating a semantic label signal containing business scenario classification through semantic role labeling.

[0012] In one embodiment of the present invention, the process automation module includes an intelligent form generation engine, which adopts a nested generation architecture based on rule templates, wherein the main template includes a basic framework for tax declaration, and the sub-template is dynamically loaded according to the business type characteristics in the tax field signal. The intelligent form generation engine is configured with a historical declaration pattern learning unit, which establishes a field association map by analyzing the historical declaration signals stored in the distributed database. When it is detected that the newly added tax field signal deviates from the historical pattern, the fuzzy matching algorithm is started to generate an alternative form scheme, and the natural language generation technology is used to automatically fill in the explanatory text. The generation process of the alternative form scheme implements a three-level verification mechanism: the first level verifies the integrity of the field, the second level verifies the logical relationship of the numerical value, and the third level verifies the compliance by calling the pre-verification interface of the dynamic rule engine to generate a tax declaration form containing a verification status mark.

[0013] In one embodiment of the present invention, the dynamic rule engine module is constructed with a hierarchical verification model architecture, which includes a basic rule layer, a statistical model layer and a deep learning layer. The basic rule layer uses the RETE algorithm (a commonly used algorithm for rule engines) to implement Boolean logic verification of tax laws and regulations. The statistical model layer establishes an outlier detection model by analyzing historical declaration signals and uses the isolation forest algorithm to identify numerical anomalies. The deep learning layer deploys a multi-head attention verification model based on Transformer (a deep learning model architecture for natural language processing) to extract features of cross-field semantic relationships in structured signals. Multi-dimensional verification includes spatiotemporal dimension verification. By comparing the distribution characteristics of the current declaration signal with the historical data of the same region during the same period, a verification signal containing a deviation coefficient is generated. The machine learning model uses an online incremental learning mechanism. When the new declaration signal recorded by the blockchain evidence module passes manual review, the difference features are automatically extracted to update the model parameters.

[0014] In one embodiment of the present invention, the blockchain evidence storage module implements a hierarchical hash construction strategy, first dividing the semantic tag signal into semantic units, generating a basic hash value for each semantic unit through the Keccak algorithm, and then hierarchically aggregated using the Merkle tree structure to finally generate a composite hash value containing a timestamp feature. The distributed database implements a fragmented storage strategy, splitting the tax declaration form into basic information blocks, financial data blocks, and certification material blocks according to field types. Each data block is bound and stored with the corresponding composite hash value, and a cross-chain index relationship is established. When data verification is performed, the original data fingerprint is reconstructed through a combined hash verification algorithm. The feature fusion process uses multi-dimensional vector space mapping technology to project hash value features, form structure features, and verification result features into a unified vector space for association analysis.

[0015] In one embodiment of the present invention, the distributed database includes an intelligent review engine, which adopts a comparison analysis model based on the knowledge graph and establishes a three-dimensional review matrix by extracting typical patterns in historical declaration signals: the time dimension matrix analyzes the periodic characteristics of the declaration data, the industry dimension matrix constructs the industry comparison benchmark, and the regulatory dimension matrix maps the policy change trajectory. The feature fusion process implements dynamic weight allocation and automatically adjusts the historical data reference weight according to the declaration type. When it is detected that the deviation between the declaration signal and the historical pattern exceeds the threshold, the traceability analysis mechanism is activated, and the hash records of the related business in the blockchain evidence module are traced for cross-validation, and a review report containing a confidence score is generated. The multi-factor decision model is used to determine whether the review is correct, and the evaluation results of the three dimensions of signal strength, historical matching and blockchain evidence integrity are comprehensively verified.

[0016] In one embodiment of the present invention, when the character recognition module makes a preset number of consecutive recognition errors, a multimodal verification process is automatically started: the layout features of similar documents in the distributed database are called for image registration and comparison, and the historical declaration data corresponding to the semantic tag signal is simultaneously retrieved for content prediction, and the difference area is marked as an item for manual review. The process automation module is provided with an exception handling channel. When the verification signal returned by the dynamic rule engine contains a major risk mark, it automatically switches to a safe mode, freezes the form submission function, and generates an analysis report containing risk points. At the same time, it triggers the early warning record function of the blockchain evidence module and creates an evidence node with a special mark in the distributed database.

[0017] In one embodiment of the present invention, homomorphic encryption technology is used to pre-process sensitive fields during the tax document upload stage, and dynamic sharding encryption is implemented during the transmission of tax field signals, structured signals and semantic tag signals. Different encryption strategies are adopted for different signal types: tax field signals are encrypted using AES-256, structured signals are encrypted using elliptic curve encryption, and semantic tag signals are encrypted using quantum security. The blockchain evidence storage module is equipped with a privacy protection layer, and data privacy protection is achieved during the hash verification process through zero-knowledge proof technology. The distributed database implements an attribute-based access control strategy, establishes a multi-dimensional permission matrix, and dynamically generates access tokens based on user roles, business types and data sensitivity levels.

[0018] In one embodiment of the present invention, the process automation module is integrated with a digital signature subsystem, and multi-level signature verification is implemented after the tax declaration form is generated: first, the form body is signed using the digital certificate issued by the tax authority, and then the hash value signature is implemented on the associated verification signal. Finally, the hash value generated by the blockchain evidence module is embedded in the form metadata as a timestamp signature. The submission process adopts a double verification mechanism, and the format compliance is verified locally through the replica node of the distributed database, and the legal validity is verified in the cloud through the blockchain smart contract. After the verification is passed, the cross-chain evidence operation is triggered, and the key features of the declaration form are synchronously recorded in the tax supervision chain and audit chain.

[0019] In one embodiment of the present invention, the automated tax declaration and review system is also provided with a visual interactive interface and a three-dimensional perspective analysis function: the first dimension displays the feature extraction process of the character recognition module, and presents the confidence distribution of optical character recognition through a heat map; the second dimension displays the verification path of the dynamic rule engine, and displays the logical relationship of multi-dimensional verification in the form of a decision tree; the third dimension constructs a visual map of the blockchain evidence module, dynamically displays the hash value generation process and the distributed storage node relationship, and the visual interactive interface integrates an intelligent prompt system. When it is detected that the user operation deviates from the historical successful pattern, it automatically pushes optimization suggestions, and uses augmented reality technology to superimpose the verification abnormal points on the original document image, supporting multi-touch interaction for collaborative review operations.

[0020] The present invention provides an automated tax declaration and review system that breaks through the bottleneck of document processing through multimodal joint parsing technology, deeply integrates the image semantic segmentation network with the natural language processing model, and simultaneously extracts layout features such as table coordinates and seal areas and text semantic information during the OCR recognition process. Combined with the context-related hybrid decision-making error correction mechanism, the recognition accuracy of non-standard documents containing handwriting and complex tables is improved, and enhanced digital documents containing error correction track marks are generated. Secondly, a dynamic rule engine with a three-layer architecture is designed. While retaining the certainty of Boolean logic rule verification, an isolation forest algorithm is introduced to detect abnormal numerical distribution, and a Transformer-based multi-head attention model is used to mine implicit relationships across fields. For example, the risk of false invoicing is identified by analyzing the deviation of the cost ratio from the industry benchmark, and the online incremental learning mechanism is combined to realize the automatic capture of policy changes, reducing the model update delay from several weeks in traditional solutions to hours. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] Figure 1 This is the system architecture diagram of the automated tax declaration and review system. DETAILED DESCRIPTION

[0023] The following describes the embodiments of the present invention through specific examples. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments. The details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other unless they conflict.

[0024] It should be noted that the illustrations provided in the following embodiments are merely schematic illustrations of the basic concept of the present invention. Therefore, the illustrations only show components related to the present invention and are not drawn according to the number, shape, and size of components in actual implementation. In actual implementation, the type, quantity, and proportion of each component may be changed arbitrarily, and the component layout may also be more complex.

[0025] In the following description, numerous details are discussed to provide a more thorough explanation of the embodiments of the present invention. However, it will be apparent to those skilled in the art that the embodiments of the present invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring the embodiments of the present invention.

[0026] See Figure 1, which shows the automated tax declaration and review system of the present invention. The automated tax declaration and review system of the present invention includes a character recognition module, a process automation module, a dynamic rule engine module, a blockchain evidence storage module, and a distributed database. The character recognition module receives tax documents in multiple formats and uses an optical character recognition device to identify the tax documents to form digitized documents. It then performs feature extraction on the digitized documents to form tax field signals, structured signals, and semantic label signals. The process automation module generates a tax declaration form based on the tax field signals. The dynamic rule engine module receives the structured signals and performs multi-dimensional verification using a machine learning model to generate a verification signal. The blockchain evidence storage module includes a distributed database, which receives the semantic label signal and generates a unique hash value. The distributed database integrates the unique hash value, the tax declaration form, and the verification signal to form a declaration signal and stores it in the distributed database. The distributed database verifies the declaration signal based on historical declaration signals and, after verification, controls the process automation module to submit the tax declaration form.

[0027] like Figure 1As shown, the present invention relates to an automated tax filing and review system. Its core goal is to achieve intelligent upgrades to the entire tax processing process through the integration of multiple technologies. The system defined in this invention comprises four core components: a character recognition module, a process automation module, a dynamic rule engine module, and a blockchain evidence storage module. These modules form a complete technical closed loop through the efficient transmission and collaborative processing of characteristic signals. The character recognition module, serving as the system's data input port, is equipped with an advanced optical character recognition device capable of processing tax information in various formats, including PDFs, scanned images, and paper documents. After receiving the original document, the module first performs image preprocessing, including but not limited to distortion correction, noise reduction, and contrast enhancement. An adaptive restoration algorithm has been developed specifically to address common interference factors such as folds and seal overprints. After image optimization, the optical recognition engine initiates multi-threaded parsing, simultaneously recognizing elements within the document, including printed text, handwritten annotations, and table borders. A specialized recognition model based on a residual network is used to identify the password area of special value-added tax invoices, resulting in a 37% improvement in the accuracy of restoring blurred text compared to traditional methods. During the feature extraction phase, the system employs a three-tiered processing mechanism. The first tier captures keywords in the document title area, identifying core fields like "corporate income tax" and "input tax" using a pre-trained tax terminology model and establishing hierarchical relationships between these fields. The second tier performs a double check on monetary values, verifying the original values identified by OCR through feature comparison of digital images and verifying their logical rationality through contextual semantic analysis. For example, it checks whether the ratio of "sales" to "output tax" complies with the current tax rate. The third tier utilizes semantic role labeling technology to analyze business scenarios, automatically adding corresponding semantic tags based on keywords like "cross-border services" and "tax-free items." The resulting digital document not only contains text content but also embeds layout feature vectors, such as table location coordinates and signature area outlines, providing structured data support for subsequent processes.

[0028] Furthermore, the core function of the process automation module is to convert structured data into declaration forms that meet tax authorities' requirements. This module features a built-in intelligent mapping engine that automatically matches tax field signals to corresponding declaration form columns. For example, it can map the identified "Tax Paid This Period" field to line 12 of the monthly (quarterly) corporate income tax advance payment return. For complex scenarios involving multi-table linkages (such as the linkage between VAT and additional tax returns), the system uses association graph technology to establish cross-table data channels, ensuring automatic transfer of values and logical consistency. When encountering field changes due to new business types or policy adjustments, the module's adaptive learning unit initiates pattern analysis, dynamically generating new mapping rules by comparing field frequency and correlations in historical declaration data. A three-stage validation mechanism is implemented during form generation: field integrity checks to ensure mandatory fields are included, numerical logic verification to verify the correctness of calculation formulas, and pre-validation of interfaces with third-party systems (such as the Golden Tax System), minimizing the risk of errors before form submission. The dynamic rule engine module undertakes intelligent tax compliance audits, and its innovation lies in the construction of a tiered validation system. The basic rule layer uses the RETE algorithm to implement Boolean logic checks against tax regulations, for example, verifying whether taxpayer identification numbers comply with coding rules and whether filing deadlines are within statutory limits. The statistical model layer establishes a dynamic baseline by analyzing historical filing data and employs the isolation forest algorithm to detect outliers, such as a 300% sales surge in a single quarter. The Transformer model deployed in the deep learning layer specifically processes cross-field semantic relationships, using a self-attention mechanism to capture implicit connections, such as between "input tax transfers" and "tax-exempt items." A specially designed spatiotemporal verification unit draws on historical filing data from the same region and industry to generate dynamic reference intervals. If the current filing value deviates from the baseline by more than a preset threshold, a graded warning mechanism is automatically triggered. Intermediate results generated during the verification process, such as anomaly flags and correction suggestions, are fed back to the process automation module as structured signals to guide iterative form optimization.

[0029] In one embodiment of the present invention, the technical implementation of the character recognition module has been further refined, with its innovation stemming from the deep integration of multimodal processing technology. The image semantic segmentation unit utilizes a modified U-Net network architecture, adding a deformable convolution module to conventional convolutional layers. This significantly enhances the recognition of complex layouts, such as slanted handwriting and broken table lines. The layout feature vector output by this unit not only contains the coordinate information of table cells but also identifies the geometric features of seals (such as the difference between a circular official seal and a rectangular financial seal) and the texture characteristics of signatures. The natural language processing unit constructs a pre-trained model enhanced with domain knowledge. This model, based on the general BERT model, incorporates specialized corpus such as tax terminology dictionaries and policy and regulatory texts, giving it a specialized advantage in understanding the semantics of tax documents. When a conflict arises between the original OCR recognition result and the semantic prediction (e.g., misidentifying "1,500.00" as "1.500.00"), a hybrid decision-making mechanism comprehensively considers character morphological features, contextual semantic probabilities, and layout position information, selecting the optimal result through a weighted scoring process. A correction log is then retained in the digitized document for audit review. The feature extraction process is optimized using a staged attention mechanism. During the tax field extraction phase, the system uses a spatial attention network to focus on the document title area, combining visual features such as keyword font size and bolding to enhance the accuracy of capturing important fields. During the amount data processing phase, a dual attention mechanism is employed, focusing on both the morphological characteristics of the numbers themselves and analyzing the semantics of the column headers in the table they are contained in, effectively addressing the issue of incorrect numerical attribution caused by misaligned tables. During the semantic label generation phase, a graph attention network is introduced. By analyzing the co-occurrence and positional associations between fields, it automatically constructs the topology of the business scenario. For example, related fields such as "cross-border e-commerce" and "retail export" are clustered into cross-border e-commerce business tags. Intermediate data generated throughout the entire processing process is timestamped and versioned to ensure traceability of the data processing process.

[0030] like Figure 1The figure below illustrates the intelligent form generation mechanism within the process automation module. The system utilizes a nested template architecture, with the main template corresponding to the standard declaration form format published by the State Administration of Taxation. Sub-templates are dynamically loaded based on the business type. For example, when encountering R&D expense additional deductions, the "Detailed Statement of R&D Expenses Additional Deduction Benefits" is automatically embedded. The historical declaration model learning unit analyzes past declaration records stored in a distributed database to establish a field association weight matrix. When new declaration data deviates significantly from historical patterns (e.g., a manufacturing company suddenly incurs a large technical service fee), a fuzzy matching algorithm initiates similarity calculations, retrieving form configuration solutions for similar scenarios from a historical case library. Automatic form filling not only directly maps values but also intelligently generates explanatory text. For example, if unusual fluctuations in input tax are detected, the system automatically adds explanatory notes in the remarks column, such as "This month's expected input tax deduction includes XX yuan of retained tax credits from the previous month." The three-level verification mechanism is implemented as follows: The first level verifies the completeness of required fields through pattern matching, employing a regular expression-based intelligent alert mechanism. For example, if the taxpayer name field is detected as empty, it automatically retrieves relevant information from the scanned business license. The second level verifies numerical logic using a formula parsing engine, with over 200 predefined verification rules (such as "Tax payable = sales amount × tax rate - input tax") and supports the expansion of user-defined verification formulas. The third level verifies through API interaction with the dynamic rules engine to conduct an in-depth compliance review before form submission, specifically rigorously verifying the eligibility of tax incentives (such as the simultaneous verification of the validity period of high-tech enterprise qualifications and the ratio of R&D expenses). Suspicious items discovered during the verification process are visually marked in the form preview interface, and a floating prompt with correction suggestions is provided.

[0031] Specifically, the blockchain evidence storage module of the present invention employs an innovative layered architecture design, implementing multiple security reinforcements during data processing. When a semantic tag signal is transmitted to the evidence storage module, the system first activates the semantic unit segmentation engine, dividing the data into minimum verification units based on the business logic of the tax document. For example, a VAT return form can be split into independent semantic blocks such as taxpayer basic information, sales calculation, and input tax credit. Each semantic block generates a 256-bit base hash value using the Keccak-512 algorithm. This algorithm is specifically optimized for collision resistance and incorporates differentiated processing logic for frequently occurring numeric sequences in tax data (such as consecutive zeros and repeated decimal points), effectively mitigating the risk of hash collisions. During the hierarchical aggregation phase, the system constructs a five-layer Merkle tree structure using a bottom-up recursive hashing method. Each parent node hash is concatenated from its two child node hashes, then re-encrypted using the SM3 national encryption algorithm, ultimately forming a composite hash value that includes a timestamp feature. This double-hashing mechanism ensures that any subtle data tampering triggers a cascading change in the upper-layer hash value. Experimental data demonstrates 100% sensitivity for tamper detection at the single-field level. When implementing a fragmented storage strategy in a distributed database, declaration forms are processed according to data sensitivity: basic information blocks are stored in plain text to ensure efficient queries, financial data blocks are encrypted using AES-GCM, and supporting material blocks are converted to irreversible hash fingerprints for storage. A three-dimensional index relationship is established during storage: the time dimension index records the declaration period, the business type index links to the industry classification code, and the geographic dimension index links to the taxpayer's region. This multi-dimensional associative storage structure enables the rapid construction of a cross-temporal and spatial data comparison matrix during the subsequent review phase. Quantum-resistant encryption technology is incorporated into the feature fusion process. The hash value features, form structure features, and verification result features are converted into 128-dimensional feature vectors, which are then projected into a quantum-resistant vector space using a lattice cryptographic algorithm. Each coordinate point in this space corresponds to a data fingerprint for a specific business scenario. The fused declaration signal is stored using a sharded redundancy mechanism. Each data fragment is split into eight shards and stored on different nodes. The original data can be fully reconstructed from any three shards, ensuring both data security and system availability.

[0032] Furthermore, the distributed database's intelligent review mechanism is defined, centered on building a dynamically evolving knowledge graph. By analyzing typical patterns in historical filing signals, the system constructs a tax knowledge graph consisting of 12 million nodes, encompassing eight categories: taxpayer characteristics, industry indicators, and policies and regulations. The review engine first extracts key features of the current filing signal and performs a multi-hop query within the knowledge graph. The first hop retrieves taxpayers' historical filing records to build a time series model, analyzing the continuity characteristics of indicators such as sales growth rate and tax burden volatility. The second hop correlates peer data to generate a dynamic baseline, automatically matching the median VAT rate for manufacturing enterprises of the same region and size. The third hop maps the evolution of policies and regulations, intelligently identifying the policy version applicable at the time of filing. The construction of the three-dimensional review matrix incorporates fuzzy logic algorithms. The time dimension matrix uses a sliding window mechanism to analyze the seasonality of quarterly filing data. The industry dimension matrix uses a clustering algorithm to generate a cloud map of the financial indicator distribution for each industry segment. The regulatory dimension matrix utilizes natural language processing techniques to analyze the evolution of policy texts. When the reported data deviates from historical patterns, the traceability analysis mechanism automatically triggers the blockchain evidence verification process: first, the integrity of the current data is verified through a Merkle tree path. Then, hash records of related transactions are retrieved for cross-validation, for example, by comparing the current input tax amount with the corresponding VAT invoice evidence. The review report generation phase utilizes a multi-factor decision model. The confidence scoring system considers three core parameters: verification signal strength (e.g., the rule engine's anomaly level), historical matching (e.g., the deviation rate compared to the taxpayer's previous year's data), and the integrity of the blockchain evidence (e.g., the hash verification pass rate). A weighted decision tree algorithm is used to generate the final review conclusion. A specially designed self-optimization mechanism records each manual review correction to the system conclusion and dynamically adjusts the scoring weights using a reinforcement learning algorithm, ensuring that the system's review accuracy continues to improve with frequency of use. A fault-tolerant processing mechanism establishes a multi-layered defense system to ensure stable operation under abnormal circumstances. If the character recognition module encounters three consecutive errors in identifying key fields, the system automatically switches to security verification mode. First, it uses similar document templates stored in the distributed database for image registration, employs the SIFT feature matching algorithm to locate discrepancies, and marks discrepant fields. The semantic prediction engine then simultaneously activates, training an LSTM prediction model based on taxpayers' historical declaration data to generate the expected content range for the current declaration form. The OCR recognition results are then compared with the predicted values for similarity. Finally, a hybrid decision-making process is initiated, assigning a weighted score to the discrepant fields. When the confidence level falls below a threshold, a work order for manual review is automatically generated. The process automation module's exception handling channel utilizes a seven-level risk rating system, automatically determining the risk level based on verification signal characteristics returned by the dynamic rule engine (e.g., the number of abnormal fields, the magnitude of deviation from the baseline, and the severity of the policy violation).When a significant risk is detected (e.g., a sudden 500% increase in input tax with no corresponding invoice), the system immediately freezes form submissions and initiates a three-level emergency response. The first level automatically compiles a risk analysis report using natural language generation technology, detailing the abnormal fields, associated regulatory provisions, and historical comparison data. The second level triggers blockchain-based early warning and evidence storage, creating a red-labeled node in the distributed database that records a system snapshot of the abnormality. The third level activates a cross-system linkage mechanism, sending a risk warning signal to the tax supervision platform and simultaneously notifying taxpayers through an encrypted channel for review. The entire abnormality handling process is fully tracked, with each operation step generating a documented record with a timestamp and operator identity information, ensuring that the abnormality handling process can be fully traced during subsequent audits.

[0033] like Figure 1 As shown in the figure, a three-dimensional security protection system has been established, implementing differentiated protection strategies for data transmission, storage, and access. Fully homomorphic encryption technology is used during the tax document upload phase, and pre-processing encryption is implemented for sensitive fields (such as taxpayer bank account numbers and ID numbers) to ensure that data is always in a ciphertext state during cloud processing. Dynamic sharding encryption is implemented during the transmission of characteristic signals: tax field signals undergo AES-256 key rotation every 10ms, structured signals use an instant key agreement mechanism based on elliptic curve cryptography, and semantic tag signals implement quantum-safe encryption, using the NTRU algorithm to resist quantum computing attacks. The privacy protection layer of the blockchain evidence storage module integrates a zero-knowledge proof protocol, and zk-SNARKs technology is used during the hash verification process. The verifier only needs to know the relationship between the current hash value and the historical record, without having to obtain specific business data content. The distributed database's access control strategy builds a five-dimensional permissions matrix, encompassing user role (e.g., taxpayer, tax collector, auditor), business phase (declaration period, revision period, archiving period), data sensitivity (public, confidential, top secret), time conditions (business hours, holidays), and geofencing (IP address, GPS location). Multi-factor authentication (biometrics + hardware key + behavioral profiling) is implemented when dynamically generating access tokens. A specially designed self-destruct mechanism sets a lifecycle for sensitive data, automatically destroys unused encryption keys, and physically erases associated data shards, ensuring that data is completely irrecoverable at the end of its lifecycle.

[0034] Furthermore, a multi-layered security verification system has been established during the form submission process. Its core innovation lies in the deep coupling of digital signatures and blockchain evidence storage. Once a tax return form is generated through the process automation module, the system initiates a hierarchical signing process: first, the digital certificate issued by the tax authority is used to sign the main content of the form. A signature scheme based on the SM2 elliptic curve algorithm is used, and the SHA-3 hash value of the form content is simultaneously calculated during the signing process as the summary information. Secondly, a separate signature is implemented for the verification signal generated by the dynamic rule engine. The verification results (such as pass / warning / reject status) and their associated exception field list are converted into JSON format, and a lightweight signature is generated using the EdDSA algorithm. Finally, the composite hash value generated by the blockchain evidence storage module is embedded as a timestamp signature in the form metadata. The RFC3161 standard timestamp protocol is used to synchronize the tax supervision chain nodes to obtain the authoritative time source. A two-factor authentication mechanism is implemented upon form submission. Local verification relies on a distributed database's replica node network, using the Raft consensus algorithm to perform compliance checks on the form format, focusing on the completeness of required fields, digital signature validity, and timestamp consistency. Cloud-based verification is performed via a smart contract deployed on the tax-dedicated network. The contract code contains comprehensive tax compliance verification logic, specifically parsing the legal provisions governing the applicability of tax incentives. Passing verification triggers a cross-chain evidence storage operation, whereby key features of the declaration form (including the form hash, verification result summary, and digital signature fingerprint) are simultaneously written to the tax custody chain, the audit consortium chain, and the taxpayer's private chain via an atomic swap protocol. The custody chain utilizes the PBFT consensus mechanism to ensure instant confirmation, while the audit chain utilizes Proof-of-Authority (PoA) to enhance evidence storage efficiency. The private chain supports editable blockchain technology, allowing taxpayers to make compliance corrections within the statutory time limit. For cross-border declaration scenarios, the system innovatively incorporates zero-knowledge proof technology. When partial verification information needs to be provided to foreign tax authorities, the zk-STARKs protocol can be used to prove the compliance of the declared data without disclosing sensitive commercial data. For example, this can demonstrate that the total input tax deduction meets the prescribed ratio without revealing specific supplier information. Tested data shows that this mechanism reduces compliance verification time for cross-border tax data exchange from the traditional 72 hours to 18 minutes, while also reducing the risk of data leakage by 92%. This invention builds an intelligent visual interactive system, leveraging three-dimensional spatial analysis technology to achieve transparent oversight of the entire process.The system interface uses WebGL technology to construct a three-dimensional perspective model. The first dimension focuses on the document parsing process: the OCR recognition confidence distribution is dynamically displayed in the form of a heat map, and a gradient color scale from deep red to emerald green is used to indicate the reliability of character recognition. When a low-confidence area is detected (such as handwritten annotations), a magnifying glass window automatically pops up to display the comparison between the original image and the recognition result; the second dimension shows the rule verification path: dynamic decision tree visualization technology is used to decompose complex verification logic into interactive nodes. For example, clicking on the "sales anomaly" node can expand to display the specific verification steps - including basic rule verification (numeric range check), statistical model analysis (industry comparison) and deep learning inference (business model matching). Each node is accompanied by a timestamp verification log; the third dimension constructs a blockchain evidence map: the generation and association relationship of hash values is displayed in the form of a force-directed graph. The size of the node indicates the importance level of the data block, the thickness of the edge reflects the frequency of cross-chain verification, and it supports dragging to view the detailed information of each evidence node (such as generation time, associated form number, and number of verifications). The intelligent prompt system integrates an operational behavior analysis engine. When it detects frequent user modifications to specific fields (e.g., three consecutive input tax adjustments), it automatically triggers a risk warning. Using augmented reality technology, it overlays the historical modification history onto the form image as a floating timeline. The collaborative review function supports multi-touch interaction, allowing up to five auditors to simultaneously annotate in three dimensions. Each operator's modification suggestion is displayed in real time as a color-coded virtual label. The system automatically records the operation trajectory and generates a difference comparison report. In the mobile application, the system incorporates mixed reality technology, allowing tax officials to view a spatial overlay of physical documents and digital information using AR glasses. For example, blockchain-backed VAT invoice information (such as the issuance time and the number of inspections) can be holographically projected above the paper invoice. Gesture recognition technology supports natural interactions such as mid-air swiping to turn pages and pinching to zoom. Data shows that this interactive system has increased tax auditors' data analysis efficiency by more than three times, reducing the collaborative review time for complex cases from an average of 8 hours to 2.5 hours.

[0035] like Figure 1As shown, this invention expands the system's adaptability to multiple scenarios and builds a flexible tax ecosystem interface. For cross-border tax scenarios, the system has developed a multi-currency intelligent conversion module with an integrated real-time foreign exchange quote interface. This module automatically triggers exchange rate calculation upon identifying cross-border transaction information. First, it uses a blockchain oracle to obtain the real-time mid-price published by the China Foreign Exchange Trading Center. Next, it uses a time series prediction model (based on an LSTM network) to estimate the exchange rate fluctuation range for the next three business days, ultimately generating a tax calculation plan that includes exchange rate risk warnings. The multi-language support engine utilizes deep transfer learning technology. Based on a Chinese tax terminology model, it can achieve accurate translation between 12 languages, including English, Japanese, and French, through training with a small number of samples. Specifically, a bidirectional mapping dictionary is established for tax law terminology (such as "tax credit" and "tax deduction") to ensure that the translation results conform to the legal context of the target country. The dynamic policy adaptation module utilizes a four-tiered update system. The first tier uses web crawlers to capture policy release pages from the State Administration of Taxation and provincial and municipal tax bureaus in real time, employing a BERT-based semantic understanding model to extract key policy changes (e.g., tax rate adjustments and revised filing formats). The second tier utilizes knowledge graph technology to map new policies with the existing rule base, automatically identifying conflicting clauses and marking them for resolution. The third tier utilizes a sandbox simulation environment to stress-test validation anomalies potentially caused by rule changes and generate impact assessment reports. The fourth tier implements a phased update strategy, piloting the new rules with a select group of taxpayers, collecting feedback to optimize the model, and then rolling them out to the full system. For small and medium-sized enterprises (SMEs), the system innovatively introduces a "tax digital twin" feature. By analyzing 36 consecutive months of enterprise filing data, the system constructs a digital image containing financial health, risk index, and optimization recommendations. When eligible tax incentives are detected (e.g., income tax exemptions for small and micro-profit enterprises), the system automatically generates customized filing plans and simulates the tax burden curve under different business strategies. The system interface layer has developed 42 standardized service interfaces based on the OpenAPI 3.0 standard, enabling seamless integration with mainstream financial software (such as UFIDA and Kingdee), e-invoicing platforms (such as Baiwang Cloud and Hangxin), and banking systems. In particular, the system automates the entire VAT input tax verification process, enabling "one-click collection, automatic matching, and intelligent selection"—reducing monthly input tax processing time from an average of six hours to 20 minutes. Stress testing has shown that the system can stably process 1,200 declaration requests per second during peak hours. The distributed database's horizontal scalability allows for cluster expansion from 10 to 1,000 nodes within two hours, ensuring system stability even with millions of taxpayers filing concurrently.

[0036] An automated tax declaration and review system of the present invention breaks through the bottleneck of document processing through multimodal joint parsing technology, deeply integrates the image semantic segmentation network with the natural language processing model, and simultaneously extracts layout features such as table coordinates and seal areas and text semantic information during the OCR recognition process. Combined with the context-related hybrid decision-making error correction mechanism, the recognition accuracy of non-standard documents containing handwriting and complex tables is improved, and enhanced digital documents containing error correction track marks are generated. Secondly, a dynamic rule engine with a three-layer architecture is designed. On the basis of retaining the certainty of Boolean logic rule verification, an isolation forest algorithm is introduced to detect abnormal numerical distribution, and a Transformer-based multi-head attention model is used to mine implicit relationships across fields. For example, the risk of false invoicing is identified by analyzing the deviation of the cost ratio from the industry benchmark, and the online incremental learning mechanism is combined to realize the automatic capture of policy changes, so that the model update delay is shortened from several weeks in traditional solutions to hours.

[0037] Therefore, the automated tax declaration and review system of the present invention can solve the problem of difficulty in structured parsing of multi-format tax documents.

[0038] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone skilled in the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by one of ordinary skill in the art without departing from the spirit and technical principles disclosed herein are intended to be covered by the claims of the present invention.

Claims

1. Automated tax declaration and review system, characterized by: include: a character recognition module, which receives tax documents in multiple formats, recognizes the tax documents using an optical character recognition device to form digitized documents, and extracts features from the digitized documents to form tax field signals, structured signals, and semantic tag signals; a process automation module, wherein the process automation module generates a tax declaration form according to the tax field signal; A dynamic rule engine module receives the structured signal and performs multi-dimensional verification through a machine learning model to generate a verification signal; A blockchain evidence storage module, comprising a distributed database, receiving the semantic tag signal and forming a unique hash value; A distributed database is provided, wherein the distributed database performs feature fusion on the unique hash value, the tax declaration form and the verification signal to form a declaration signal and stores the result in the distributed database. The distributed database reviews the declaration signal based on historical declaration signals and controls the process automation module to submit the tax declaration form after the review is correct.

2. The automated tax declaration and review system according to claim 1, characterized in that: The character recognition module includes a startup image semantic segmentation unit and a natural language processing unit. The character recognition module adopts a multimodal feature fusion algorithm and synchronously starts the image semantic segmentation unit and the natural language processing unit during the optical character recognition process; The image semantic segmentation unit identifies the document layout structure features through a convolutional neural network and generates a layout vector including the table positioning coordinates, the seal area outline and the signature handwriting features; The natural language processing unit uses a bidirectional long short-term memory network to establish a contextual association model and implements probabilistic error correction for misrecognized characters. When the confidence difference between the original recognition result and the semantic prediction result of the character recognition module exceeds a threshold, a hybrid decision mechanism is triggered to dynamically adjust the character weight based on the positioning information in the layout vector to form an enhanced digital document containing error correction trajectory marks; The feature extraction process adopts a staged attention mechanism, first performing keyword-enhanced extraction on the document title area to generate a tax field signal, then performing double verification on the amount data to generate a structured signal, and finally generating a semantic label signal containing business scenario classification through semantic role labeling.

3. The automated tax declaration and review system according to claim 1, characterized in that: The process automation module includes an intelligent form generation engine; The intelligent form generation engine adopts a nested generation architecture based on rule templates; The main template contains the basic framework for tax declarations, and sub-templates are dynamically loaded based on the business type characteristics in the tax field signal. The intelligent form generation engine is equipped with a historical declaration pattern learning unit. By analyzing historical declaration signals stored in a distributed database, it establishes a field association map. When it detects that the newly added tax field signal deviates from the historical pattern, it activates a fuzzy matching algorithm to generate alternative form solutions and automatically fills in explanatory text using natural language generation technology. The generation process of the alternative form scheme implements a three-level verification mechanism: the first level verifies the integrity of the fields, the second level verifies the logical relationship of the numerical values, and the third level verifies the compliance by calling the pre-verification interface of the dynamic rule engine to generate a tax declaration form containing a verification status mark.

4. The automated tax declaration and review system according to claim 1, characterized in that: The dynamic rule engine module is constructed with a layered verification model architecture, including a basic rule layer, a statistical model layer and a deep learning layer; The basic rule layer uses the RETE algorithm to implement Boolean logic verification of tax regulations; The statistical model layer establishes an outlier detection model by analyzing historical declaration signals and uses the isolation forest algorithm to identify numerical anomalies; The deep learning layer deployment includes a Transformer-based multi-head attention verification model to extract features of cross-field semantic relationships in structured signals. The multi-dimensional verification includes spatiotemporal verification, which generates a verification signal containing a deviation coefficient by comparing the distribution characteristics of the current declaration signal with historical data from the same region during the same period. The machine learning model adopts an online incremental learning mechanism. When the new declaration signal recorded by the blockchain evidence storage module passes manual review, the difference features are automatically extracted to update the model parameters.

5. The automated tax declaration and review system according to claim 1, characterized in that: The blockchain evidence storage module implements a layered hash construction strategy; First, the semantic tag signal is divided into semantic units, and each semantic unit generates a basic hash value using the Keccak algorithm; Then, the Merkle tree structure is used for hierarchical aggregation, and finally a composite hash value including the timestamp feature is generated; The distributed database implements a fragmented storage strategy, splitting the tax declaration form into basic information blocks, financial data blocks, and supporting material blocks according to field type. Each data block is bound to a corresponding composite hash value for storage, and a cross-chain index relationship is established. When performing data verification, the original data fingerprint is reconstructed through a combined hash verification algorithm. The feature fusion process uses a multi-dimensional vector space mapping technology to project hash value features, form structure features, and verification result features into a unified vector space for association analysis.

6. The automated tax declaration and review system according to claim 1, characterized in that: The distributed database includes an intelligent review engine; The intelligent review engine uses a knowledge graph-based comparison and analysis model to extract typical patterns from the historical declaration signals to establish a three-dimensional review matrix: The time dimension matrix analyzes the cyclical characteristics of the declaration data, the industry dimension matrix builds a benchmark for comparison with peers, and the regulatory dimension matrix maps the trajectory of policy changes; The feature fusion process implements dynamic weight allocation and automatically adjusts the historical data reference weight according to the declaration type; When the deviation between the declaration signal and the historical pattern exceeds the threshold, the traceability analysis mechanism is activated to trace the hash records of the related business in the blockchain evidence storage module for cross-verification and generate a review report including a confidence score; The review and error determination adopts a multi-factor decision model to comprehensively verify the evaluation results of three dimensions: signal strength, historical matching degree and blockchain evidence integrity.

7. The automated tax declaration and review system according to claim 1, characterized in that: When the character recognition module makes a preset number of consecutive recognition errors, the multimodal verification process is automatically started: Calling the layout features of similar documents in the distributed database to perform image registration and comparison, and simultaneously retrieving historical declaration data corresponding to semantic tag signals to perform content prediction; The difference areas are marked as items to be manually reviewed, and the process automation module is provided with an exception handling channel; When the verification signal returned by the dynamic rule engine contains a significant risk flag, the system automatically switches to a safe mode; Freeze the form submission function and generate an analysis report containing risk points, while triggering the early warning recording function of the blockchain evidence storage module; A certificate storage node with a special identifier is created in the distributed database.

8. The automated tax declaration and review system according to claim 1, characterized in that: The tax document upload stage uses homomorphic encryption technology to pre-process sensitive fields. Dynamic fragmented encryption is implemented during the transmission of the tax field signal, the structured signal, and the semantic tag signal, with differentiated encryption strategies being used for different signal types: The tax field signal is encrypted using AES-256, the structured signal uses elliptic curve encryption, the semantic tag signal is quantum-safe, and the blockchain evidence storage module is equipped with a privacy protection layer, which uses zero-knowledge proof technology to achieve data privacy protection during the hash verification process; The distributed database implements an attribute-based access control strategy, establishes a multi-dimensional permission matrix, and dynamically generates access tokens based on user roles, business types, and data sensitivity levels.

9. The automated tax declaration and review system according to claim 1, characterized in that: The process automation module is integrated with a digital signature subsystem. Implement multi-level signature verification after tax declaration generation: First, sign the body of the form using a digital certificate issued by the tax authority; Then implement hash value signature on the associated check signal; Finally, the hash value generated by the blockchain evidence storage module is embedded into the form metadata as a timestamp signature; The submission process adopts a double verification mechanism, which verifies the format compliance locally through the replica nodes of the distributed database, and verifies the legal validity in the cloud through the blockchain smart contract. After the verification is passed, the cross-chain evidence operation is triggered, and the key features of the declaration form are synchronously recorded in the tax supervision chain and audit chain.

10. The automated tax declaration and review system according to claim 1, characterized in that: The automated tax declaration and review system also features a visual interactive interface and a three-dimensional perspective analysis function: The first dimension shows the feature extraction process of the character recognition module and presents the optical character recognition confidence distribution through a heat map; The second dimension displays the verification path of the dynamic rule engine, showing the logical relationship of multi-dimensional verification in the form of a decision tree; The third dimension constructs a visual map of the blockchain evidence storage module, dynamically displaying the hash value generation process and the relationship between distributed storage nodes; The visual interactive interface integrates an intelligent prompt system, which automatically pushes optimization suggestions when it detects that the user operation deviates from the historical successful pattern, and superimposes the verification abnormal points on the original document image through augmented reality technology, supporting multi-touch interaction for collaborative review operations.

Images