A method and system for automatically generating a finance and tax digital resource by fusing machine learning artificial intelligence

Abstract

This invention discloses a method and system for automatically generating financial and tax digital resources by integrating machine learning and artificial intelligence, belonging to the field of digital resource generation technology. It includes: setting up a multi-source heterogeneous data input layer to receive multi-source heterogeneous data and obtain input data items, whereby the multi-source heterogeneous data includes text data, image data, time-series data, and structured data; and a feature extraction module based on the input data items to extract the analysis results of text data, image data, time-series data, and structured data respectively, thereby obtaining an analysis dataset. This invention, through the use of an OCR-RCNN hybrid model, can not only accurately recognize handwritten invoice information such as amount and date, but also automatically perform image correction, such as tilting and blurring, significantly improving the recognition accuracy and robustness of the original image data, and overcoming the limitations of applying OCR technology alone.

Description

Technical Field

[0001] This invention relates to the field of digital resource generation technology, specifically to a method and system for automatically generating financial and tax digital resources that integrates machine learning and artificial intelligence. Background Technology

[0002] Tax digital resources refer to the collection, processing, analysis, and application of relevant data, information, and services in the financial and tax fields using digital technologies such as the Internet, big data, artificial intelligence, and cloud computing. These resources aim to improve the efficiency, accuracy, and transparency of financial and tax management and are widely used in corporate financial management, government tax collection and administration, and personal tax processing. The core of tax digital resources is the integration of digital tools and platforms, including financial data such as income and expenditure records and reports, tax information such as declaration data, policies and regulations, and related technical services such as algorithm models and cloud storage. It emphasizes the automation and intelligentization of financial and tax processes through technological means, such as using big data analysis to predict tax trends or using artificial intelligence to optimize financial decisions.

[0003] The patent publication number CN119166732A discloses a method, apparatus, device, and medium for generating and using digital resource maps. This method acquires digital resource information from multiple management departments, including various resources with corresponding relationships between them. A digital resource map is constructed based on these relationships, displaying the supply and demand relationships and usage status of various resources. The digital resource map is then analyzed to obtain the usage efficiency of each resource within the digital resource information. This application-centric method draws resource maps, displays the relationships between resources, statistically analyzes the usage of various resources, avoids redundant resource construction, achieves intensive management of digital resources, and improves resource supply and demand efficiency and service experience.

[0004] Traditional methods rely on structured data, such as ERP spreadsheets, but cannot efficiently process unstructured data, such as handwritten invoices and contract texts. This leads to the omission of key information, which in turn affects the integrity and accuracy of financial and tax digital resources. For example, the amount on a handwritten invoice is difficult to extract accurately through optical character recognition, and special clauses in a contract text may not be automatically recognized and parsed. At the same time, the phenomenon of data silos is serious, and heterogeneous source data such as bank statements, tax systems, and supply chain logs are difficult to integrate, resulting in incomplete resource coverage. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for automatically generating financial and tax digital resources that integrates machine learning and artificial intelligence, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for automatically generating financial and tax digital resources by integrating machine learning and artificial intelligence, comprising:

[0007] A multi-source heterogeneous data input layer is set up to receive multi-source heterogeneous data and obtain input data items. The multi-source heterogeneous data includes text data, image data, time series data and structured data.

[0008] The feature extraction module extracts text data, image data, time series data, and structured data analysis results based on the input data items, thereby obtaining the analysis dataset.

[0009] Based on the analysis dataset, modality alignment is performed, a multimodal fusion engine is set up, and feature vectors of different modalities are weighted and aggregated through an attention mechanism to generate a unified embedding representation, which is then input into a multimodal generative adversarial network to obtain integrated data items;

[0010] Based on the creation of a dynamic generation and optimization module based on integrated data items, the GAN generator outputs preliminary financial and tax resources, including draft tax reports, and obtains preliminary output items;

[0011] By introducing a data credibility scoring mechanism through a discriminator, and correcting the generation bias based on reinforcement learning feedback, the initial output items are optimized and corrected to obtain the optimized output items.

[0012] Based on the output optimization items, a blockchain audit and output layer is set up. The generated results are stored in the blockchain after being hashed and encrypted. An interface report with interpretable annotations is output to obtain the output results items. Then, based on artificial intelligence, end-to-end trusted financial and tax resources are generated, covering defects in the entire process such as data silos, dynamic compliance, and audit traceability.

[0013] Furthermore, the method for obtaining the input data items includes:

[0014] Text data includes scanned copies of contracts and PDF terms; image data includes VAT invoices and handwritten vouchers; time-series data includes bank transaction records and supply chain event logs; and structured data includes CSV and XML tables exported from the ERP system, resulting in a multi-source dataset.

[0015] Data caching and queue management are implemented based on multi-source datasets. Data priority sorting is set, and asynchronous transmission is achieved using Kafka message queues based on the priority sorting results. A memory database is set to temporarily store the original data, thereby obtaining the input data items.

[0016] Furthermore, the multi-source heterogeneous data input layer also includes an adaptive parsing submodule and a data normalization pipeline:

[0017] The adaptive parsing submodule can include an image module and a text module. The image module uses an OCR-RCNN hybrid model to recognize handwritten invoices and automatically correct tilted and blurred images. The text module integrates a knowledge graph of fiscal and tax policies and matches contract terms and legal entries through entity links.

[0018] The data normalization pipeline defines a unified data specification across modalities, converts heterogeneous data such as bank transaction records and tax codes into Tensor format, deploys a data cleaning agent, sets a missing threshold (a fixed percentage threshold), and filters invalid inputs that reach the missing threshold based on a rule engine.

[0019] Furthermore, the method for obtaining the analysis dataset includes:

[0020] Image data is parsed using a convolutional neural network to extract key fields from invoices, including amount and tax number, to obtain image analysis items. A Transformer-based NLP model is used to process text modalities and identify clause semantics to obtain text analysis items. A time series model based on LSTM is used to analyze financial data flow and detect abnormal fluctuations to obtain time series analysis items and structural analysis items.

[0021] A lightweight feature extractor is designed to extract features based on image analysis, text analysis, time series analysis, and structural analysis, respectively, thereby obtaining the analysis dataset.

[0022] Furthermore, the method for obtaining the integrated data items includes:

[0023] Based on image analysis, text analysis, temporal analysis, and structural analysis, a cross-modal attention mechanism is designed to create an intermodal attention weight matrix. The semantic correlation between the text description and the invoice image is calculated to obtain the associated information items.

[0024] Based on the generative adversarial network architecture of associated information items, a hierarchical attention aggregator is set up, including intramodal self-attention and cross-modal collaborative attention. Intramodal self-attention is used to extract deep features of a single modality, and cross-modal collaborative attention is used to generate joint embedding vectors to obtain integrated data items.

[0025] Furthermore, the method for obtaining the initial output items includes:

[0026] A multimodal Transformer with a U-Net structure is designed to encode and decode the data items by inputting a fusion vector and outputting the hidden state. At the same time, based on the U-Net skip connection structure, the hidden state and text data are connected to generate a report field, thus obtaining the initial output items.

[0027] Furthermore, the method for obtaining the output optimization term includes:

[0028] The credibility scoring mechanism includes a reinforcement learning agent design, which includes a state space and a reward function. The state space is used to generate reports on field completeness, logical consistency, and policy compliance. The reward function is the weighted calculation result of the compliance score and the manual review pass rate, respectively.

[0029] A correction threshold is set, which is the score of the reward function. When the score is lower than the correction threshold, iterative regeneration is triggered. When the score is higher than the correction threshold, the report is directly output to the blockchain audit layer, thereby obtaining the output optimization items.

[0030] Furthermore, the method for obtaining the output result item includes:

[0031] Based on the output optimization terms, the decision points in the generation process are written into the smart contract, and the original data hash and the report hash are stored on a dual-chain basis to obtain the stored information items;

[0032] Based on the stored information items, a PDF annotation report is generated through SHAP value analysis to mark key decision factors, thereby obtaining the output results.

[0033] An automatic generation system for financial and tax digital resources integrating machine learning and artificial intelligence, using the aforementioned method for automatic generation of financial and tax digital resources integrating machine learning and artificial intelligence, includes:

[0034] Data acquisition module: Set up a multi-source heterogeneous data input layer to receive multi-source heterogeneous data and obtain input data items;

[0035] Feature extraction module: Based on the input data items, the feature extraction module extracts the analysis results of text data, image data, time series data and structured data respectively, and then obtains the analysis dataset;

[0036] Data alignment module: Based on the analysis dataset, modality alignment is performed. A multimodal fusion engine is set up, and feature vectors of different modalities are weighted and aggregated through an attention mechanism to generate a unified embedding representation. This representation is then input into a multimodal generative adversarial network to obtain integrated data items.

[0037] Optimized output module: Based on the integrated data items, a dynamic generation and optimization module is created. The generator of GAN outputs preliminary financial and tax resources, including draft tax reports, to obtain preliminary output items. A data credibility scoring mechanism is introduced through a discriminator. Based on reinforcement learning feedback, the generation bias is corrected, and the preliminary output items are optimized and corrected to obtain optimized output items.

[0038] Output annotation module: Based on the output optimization items, the blockchain audit and output layer is set, the generated results are stored in the blockchain after being hashed and encrypted, and the interface report with interpretable annotations is output to obtain the output result items.

[0039] Compared with the prior art, the beneficial effects of the present invention are:

[0040] This method and system for automatically generating financial and tax digital resources, which integrates machine learning and artificial intelligence, utilizes an OCR-RCNN hybrid model. It not only accurately recognizes handwritten invoice information such as amounts and dates, but also automatically corrects images, such as tilting and blurring, significantly improving the accuracy and robustness of original image data recognition. This overcomes the limitations of applying OCR technology alone. Furthermore, the text modality integrates a financial and tax policy knowledge graph. Through entity linking technologies, such as semantic similarity calculation based on BERT embedding, it intelligently matches contract terms with legal entries, deeply understanding the semantics of financial and tax clauses expressed in natural language, thus solving the problem of automatically parsing contract and other textual information.

[0041] Meanwhile, data normalization defines a unified data standard across modalities, such as converting to Tensor format, and deploys a data cleaning agent to filter invalid input based on a rule engine and missing threshold, standardizing heterogeneous data such as bank statements and tax codes, laying the foundation for subsequent processing. Asynchronous transmission via Kafka queues and temporary storage in Redis in-memory databases ensure system response speed and reliability under high-concurrency data input, avoiding data congestion and loss, fundamentally solving the problem that traditional methods cannot effectively utilize unstructured data sources, and significantly improving the integrity of financial and tax data resources and the accuracy of source data.

[0042] Furthermore, by calculating the semantic relevance between different data sources through a cross-modal attention mechanism and deeply fusing heterogeneous features using a hierarchical attention aggregator of a multimodal generative adversarial network, a unified embedded representation is generated. The U-Net-structured multimodal Transformer generator, combined with a financial and tax knowledge graph and preset templates, dynamically generates preliminary financial and tax resources, solving the problems of traditional methods generating content that is singular and unable to integrate multi-source information. A data credibility scoring mechanism is introduced, the core of which is a reinforcement learning agent. The field completeness, logical consistency, and policy compliance of the generated report are used as the state space. A reward function is designed by combining the compliance score and the manual review pass rate, realizing automated quality assessment and dynamic optimization of the report, significantly improving the logical consistency and policy compliance of the generated results. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the multi-source heterogeneous data input layer process of the present invention;

[0044] Figure 2 This is a schematic diagram of the feature extraction and modal analysis process of the present invention;

[0045] Figure 3 This is a schematic diagram of the modal alignment and fusion process of the present invention;

[0046] Figure 4 This is a schematic diagram of the dynamic generation and optimization process of the present invention;

[0047] Figure 5 This is a schematic diagram of the blockchain auditing and output process of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Traditional methods for automatically generating financial and tax digital resources often rely on structured data, such as tabular data from Enterprise Resource Planning (ERP) systems. This data has clear field definitions and standardized data formats, making it easy to extract, transform, and load (ETL). However, in real-world scenarios, financial and tax information comes from diverse sources, including a large amount of unstructured data such as handwritten invoices, contract texts, and email correspondence. This data has high information density but lacks a clear structure, making it difficult to process effectively using traditional methods. For example, even with Optical Character Recognition (OCR) technology, the accuracy of extracting information such as amounts and dates from handwritten invoices is difficult to guarantee, easily leading to recognition errors or omissions. Furthermore, special clauses in contract texts often exist in natural language, requiring complex Natural Language Processing (NLP) technologies for automatic recognition and parsing. Relying solely on structured data will result in the omission of a large amount of key information, seriously affecting the integrity and accuracy of financial and tax digital resources. This application provides a method for automatically generating financial and tax digital resources that integrates machine learning and artificial intelligence, employing an OCR-RCNN hybrid model... This system not only accurately recognizes handwritten invoice information such as amounts and dates, but also automatically corrects images, such as tilting and blurring, significantly improving the recognition accuracy and robustness of raw image data. It overcomes the limitations of standalone OCR technology. Furthermore, the text modality integrates a knowledge graph of fiscal and tax policies. Through entity linking technologies, such as semantic similarity calculation based on BERT embedding, it intelligently matches contract terms with legal entries, deeply understanding the semantics of fiscal and tax clauses expressed in natural language. This solves the problem of automatic parsing of textual information such as contracts. Data normalization defines a unified data standard across modalities, such as converting to Tensor format, and deploys a data cleaning agent. Based on a rule engine and missing thresholds, it filters invalid input, standardizing heterogeneous data such as bank statements and tax codes, laying the foundation for subsequent processing. Asynchronous transmission via Kafka queues and temporary storage in a Redis in-memory database ensure system response speed and reliability under high-concurrency data input, avoiding data congestion and loss. This fundamentally solves the problem of traditional methods failing to effectively utilize unstructured data sources, significantly improving the integrity of fiscal and tax data resources and the accuracy of source data. Figures 1-5 As shown, it includes steps S100-S600.

[0050] Step S100: Set up a multi-source heterogeneous data input layer to receive multi-source heterogeneous data and obtain input data items.

[0051] It is important to note that, such as Figure 1As shown, the multi-source heterogeneous data includes text data, image data, time-series data, and structured data. The methods for obtaining input data items include: text data including scanned copies of contracts and PDF terms; image data including VAT invoices and handwritten vouchers; time-series data including bank transaction records and supply chain event logs; and structured data including CSV and XML tables exported from the ERP system, resulting in a multi-source dataset. Based on this multi-source dataset, data caching and queue management are implemented, data priority sorting is set, and based on the priority sorting results, asynchronous transmission is achieved using a Kafka message queue. An in-memory database is set to temporarily store the original data, thus obtaining the input data items.

[0052] Specifically, data priority is set, such as prioritizing urgent tax reports. Asynchronous transmission is achieved using Kafka message queues to ensure high throughput and reliability. In-memory databases, such as Redis, are deployed to temporarily store raw data, solving congestion and latency issues during data input and improving system response speed. Through queue management, the system can handle high-concurrency data and avoid data loss. Priority sorting ensures that critical data, such as expired invoices, is processed first, adapting to dynamic business needs.

[0053] In the specific implementation process, the monthly financial and tax processing of a certain e-commerce platform involves the following data collection phase: Text data: Supplier contract PDFs, such as payment terms documents; Image data: Sales VAT invoice images, including handwritten notes; Time-series data: Daily bank transaction records (CSV), recording income and expenses; Structured data: Order data XML exported from the ERP system, including customer information and product sales, resulting in a multi-source dataset. Data caching and queue management are then implemented. System priorities are set: Bank transaction data has the highest priority due to its involvement in real-time auditing, followed by contract data. Asynchronous transmission using Kafka queues is employed: for example, when the ERP system exports data, Kafka processes it in batches to avoid system overload. In-memory database temporary storage is used: Redis stores raw data, such as invoice images, with a temporary storage time not exceeding 5 minutes to ensure fast access. The priority mechanism ensures timely generation of tax reports and reduces human error.

[0054] It is important to note that the multi-source heterogeneous data input layer also includes an adaptive parsing submodule and a data normalization pipeline. The adaptive parsing submodule can include image and text modules. The image module uses an OCR-RCNN hybrid model for handwritten invoice recognition, automatically correcting tilted and blurred images. The text module integrates a knowledge graph of fiscal and tax policies, matching contract terms and legal entries through entity links. The data normalization pipeline defines a unified data specification across modalities, converting heterogeneous data such as bank transaction records and tax codes into Tensor format, deploying a data cleaning agent, setting a missing threshold (a fixed percentage threshold), and filtering invalid inputs that reach the missing threshold based on a rule engine.

[0055] Specifically, the image module adopts an OCR-RCNN hybrid model with a deepening mechanism. The RCNN part handles handwritten character recognition, such as using geometric transformations for tilt correction. The OCR part extracts text fields and combines them with correction. The text module integrates knowledge graphs, such as a tax policy database, and matches entity links to contract terms and regulations. It uses BERT embedding to calculate similarity to ensure semantic consistency. It converts heterogeneous data such as bank transaction records and tax codes into Tensor format, deploys a data cleaning agent, sets a missing threshold of 10%, and filters invalid inputs that reach the missing threshold based on a rule engine.

[0056] Step S200: Based on the input data items, the feature extraction module extracts the text data, image data, time series data, and structured data analysis results respectively, and then obtains the analysis dataset.

[0057] It is important to note that, such as Figure 2 As shown, the method for obtaining the analysis dataset includes: using a convolutional neural network to parse image data, extracting key fields of invoices, including amount and tax number, to obtain image analysis items; using a Transformer-based NLP model to process text modalities, recognizing clause semantics, to obtain text analysis items; using an LSTM-based time series model to analyze financial data flow, detecting abnormal fluctuations, to obtain time series analysis items and structural analysis items; setting a lightweight feature extractor to extract features based on image analysis items, text analysis items, time series analysis items, and structural analysis items, respectively, to obtain the analysis dataset.

[0058] Specifically, image data is processed using convolutional neural networks, such as a lightweight version of ResNet-18, to extract fields such as invoice amount. Key regions are located through convolutional layers. Based on the set lightweight feature extractor, image analysis items, text analysis items, temporal analysis items, and structural analysis items are integrated to generate a unified analysis dataset. Image features are output as 512-dimensional convolutional feature vectors, text features are generated as 768-dimensional BERT embedding vectors, and temporal and structural features are encoded using LSTM to generate 128-dimensional state vectors.

[0059] Step S300: Based on the analysis dataset, perform modality alignment, set up a multimodal fusion engine, and aggregate different modal feature vectors through an attention mechanism to generate a unified embedding representation. Input the unified embedding representation into a multimodal generative adversarial network to obtain integrated data items.

[0060] It is important to note that, such as Figure 3 As shown, the method for obtaining integrated data items includes: based on image analysis items, text analysis items, temporal analysis items, and structural analysis items, a cross-modal attention mechanism is set to design an inter-modal attention weight matrix, and the semantic correlation between the text description and the invoice image is calculated to obtain the associated information items; based on the associated information items, a generative adversarial network architecture is generated, and a hierarchical attention aggregator is set, including intra-modal self-attention and cross-modal collaborative attention. Intra-modal self-attention is used to extract deep features of a single modality, and cross-modal collaborative attention is used to generate a joint embedding vector to obtain integrated data items.

[0061] Specifically, a cross-modal attention matrix is ​​constructed, and the association weights between the text description and the invoice image are calculated, such as the cosine similarity between the contract amount and the invoice amount. The weights are dynamically adjusted, such as by using the Softmax function, to output associated information items. The hierarchical attention aggregator of the Generative Adversarial Network (GAN) architecture includes intra-modal self-attention and cross-modal collaborative attention. The fusion logic generates unified integrated data items through concatenation and linear transformation. GAN training uses Wasserstein loss to improve stability.

[0062] Step S400: Based on the integrated data items, a dynamic generation and optimization module is created. The GAN generator outputs preliminary financial and tax resources, including draft tax reports, to obtain preliminary output items.

[0063] It is important to note that, such as Figure 4 As shown, the method for obtaining the initial output items includes: setting up a multimodal Transformer with a U-Net structure, designing an encoder-decoder, integrating data items with input fusion vectors, outputting hidden states, and connecting the hidden states and text data based on the U-Net skip connection structure to generate report fields, thereby obtaining the initial output items.

[0064] Specifically, in the encoder-decoder design, the encoder is a 12-layer Transformer, and the input fused vector... Output hidden state The decoder uses a U-Net skip connection structure, which will... Combined with shallow features, such as invoice amount concatenation, to generate report fields, field constraints are generated through pre-set financial and tax templates, such as profit and loss statement structures, and the output format is restricted through pointer networks; dynamic variable filling: such as "income tax = taxable income × tax rate", where the tax rate is obtained in real time from the knowledge graph.

[0065] Step S500: Introduce a data credibility scoring mechanism through a discriminator, correct the generation bias based on reinforcement learning feedback, optimize and correct the initial output items, and obtain the output optimization items.

[0066] It is important to note that the methods for obtaining output optimization items include: the credibility scoring mechanism includes a reinforcement learning agent design, which includes a state space and a reward function. The state space is used to generate the report's field completeness, logical consistency, and policy compliance. The reward function is the weighted calculation result corresponding to the compliance score and the manual review pass rate, respectively. A correction threshold is set, which is the reward function score. When the score is lower than the correction threshold, iterative regeneration is triggered. When the score is higher than the correction threshold, the report is directly output to the blockchain audit layer, thereby obtaining the output optimization items.

[0067] Specifically, the reward function is a weighted calculation of the compliance score and the manual review pass rate, with the compliance score having a weight of 0.7 and the manual review pass rate having a weight of 0.3. A higher reward value indicates a higher report quality. The set correction threshold is 0.7. When the score is lower than the correction threshold, iterative regeneration is triggered. At the same time, abnormal fields are highlighted in the output layer, such as input tax deduction conflicts. When the score is higher than the correction threshold, the report is directly output to the blockchain audit layer, thereby obtaining output optimization items.

[0068] Step S600: Set up the blockchain audit and output layer according to the output optimization items, generate the results and store them in the blockchain after hash encryption, output an interface report with interpretable annotations, and obtain the output result items.

[0069] It is important to note that, based on artificial intelligence, end-to-end trusted financial and tax resources are generated, covering defects throughout the entire process such as data silos, dynamic compliance, and audit traceability. The methods for obtaining output items include: based on the output optimization items, decision points in the generation process are written into smart contracts, and the original data hash and report hash are stored on a dual-chain basis to obtain stored information items; based on the stored information items, a PDF annotation report is generated through SHAP value analysis to mark key decision factors, thereby obtaining output items.

[0070] Specifically, decision points in the generation process, such as modal weights and reward function scores, are written into smart contracts. The original data hash and report hash are stored on a dual-chain via Hyperledger Fabric channels to obtain stored information items. Based on the stored information items, a PDF annotation report is generated through SHAP value analysis to mark key decision factors, thereby obtaining output results. The SHAP value analysis engine includes field attribution calculation. For numerical fields in the generated report, such as "tax payable", the Shapley value of each input feature is calculated. For example, if the SHAP value of the invoice amount is 0.7 and the SHAP value of the policy item is 0.3, it is marked as "mainly driven by invoice A". Then, visualization rendering is performed. The PDF report embeds a heatmap to highlight high-impact fields and provides an interactive web interface to support users to manually adjust the SHAP weights.

[0071] An automatic generation system for financial and tax digital resources integrating machine learning and artificial intelligence is disclosed. This system utilizes the aforementioned method for automatic generation of financial and tax digital resources integrating machine learning and artificial intelligence, and includes: a data acquisition module: setting up a multi-source heterogeneous data input layer to receive multi-source heterogeneous data and obtain input data items; a feature extraction module: performing feature extraction based on the input data items to extract text data, image data, time-series data, and structured data analysis results, thereby obtaining an analysis dataset; and a data alignment module: performing modal alignment based on the analysis dataset, setting up a multi-modal fusion engine, and weightedly aggregating different modal features through an attention mechanism. The system generates a unified embedded representation, which is then input into a multimodal generative adversarial network (GAN) to obtain integrated data items. The output optimization module is created based on these integrated data items. The GAN's generator outputs preliminary financial and tax resources, including draft tax reports, resulting in preliminary output items. A data credibility scoring mechanism is introduced through a discriminator, and reinforcement learning feedback corrects generation biases, optimizing the preliminary output items to obtain optimized output items. The output annotation module sets up a blockchain audit and output layer based on the optimized output items. The generated results are hashed, encrypted, and stored in the blockchain. An interface report with interpretable annotations is output, resulting in the final output item.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended embodiments and their equivalents.

Claims

1. A method for automatically generating financial and tax digital resources by integrating machine learning and artificial intelligence, comprising: A multi-source heterogeneous data input layer is set up to receive multi-source heterogeneous data and obtain input data items. The multi-source heterogeneous data includes text data, image data, time series data and structured data. Its characteristic is that it further includes: The feature extraction module extracts text data, image data, time series data, and structured data analysis results based on the input data items, thereby obtaining the analysis dataset. Based on the analysis dataset, modality alignment is performed, a multimodal fusion engine is set up, and feature vectors of different modalities are weighted and aggregated through an attention mechanism to generate a unified embedding representation, which is then input into a multimodal generative adversarial network to obtain integrated data items; Based on the creation of a dynamic generation and optimization module based on integrated data items, the GAN generator outputs preliminary financial and tax resources, including draft tax reports, and obtains preliminary output items; By introducing a data credibility scoring mechanism through a discriminator, and correcting the generation bias based on reinforcement learning feedback, the initial output items are optimized and corrected to obtain the optimized output items. Based on the output optimization items, a blockchain audit and output layer is set up. The generated results are stored in the blockchain after being hashed and encrypted. An interface report with interpretable annotations is output to obtain the output results items. Then, based on artificial intelligence, end-to-end trusted financial and tax resources are generated, covering the defects of data silos, dynamic compliance, and audit traceability throughout the entire process.

2. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 1, characterized in that: The method for obtaining the input data items includes: Text data includes scanned copies of contracts and PDF terms; image data includes VAT invoices and handwritten vouchers; time-series data includes bank transaction records and supply chain event logs; and structured data includes CSV and XML tables exported from the ERP system, resulting in a multi-source dataset. Data caching and queue management are implemented based on multi-source datasets. Data priority sorting is set, and asynchronous transmission is achieved using Kafka message queues based on the priority sorting results. A memory database is set to temporarily store the original data, thereby obtaining the input data items.

3. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 2, characterized in that: The multi-source heterogeneous data input layer also includes an adaptive parsing submodule and a data normalization pipeline: The adaptive parsing submodule can include an image module and a text module. The image module uses an OCR-RCNN hybrid model to recognize handwritten invoices and automatically correct tilted and blurred images. The text module integrates a knowledge graph of fiscal and tax policies and matches contract terms and legal entries through entity links. The data normalization pipeline defines a unified data specification across modalities, converts heterogeneous data such as bank transaction records and tax codes into Tensor format, deploys a data cleaning agent, sets a missing threshold (a fixed percentage threshold), and filters invalid inputs that reach the missing threshold based on a rule engine.

4. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 1, characterized in that: The methods for obtaining the analysis dataset include: Image data is parsed using a convolutional neural network to extract key fields from invoices, including amount and tax number, to obtain image analysis items. A Transformer-based NLP model is used to process text modalities and identify clause semantics to obtain text analysis items. A time series model based on LSTM is used to analyze financial data flow and detect abnormal fluctuations to obtain time series analysis items and structural analysis items. A lightweight feature extractor is designed to extract features based on image analysis, text analysis, time series analysis, and structural analysis, respectively, thereby obtaining the analysis dataset.

5. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 4, characterized in that: The method for obtaining the integrated data items includes: Based on image analysis, text analysis, temporal analysis, and structural analysis, a cross-modal attention mechanism is designed to create an intermodal attention weight matrix. The semantic correlation between the text description and the invoice image is calculated to obtain the associated information items. Based on the generative adversarial network architecture of associated information items, a hierarchical attention aggregator is set up, including intramodal self-attention and cross-modal collaborative attention. Intramodal self-attention is used to extract deep features of a single modality, and cross-modal collaborative attention is used to generate joint embedding vectors to obtain integrated data items.

6. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 1, characterized in that: The method for obtaining the initial output items includes: A multimodal Transformer with a U-Net structure is designed to encode and decode the data items by inputting a fusion vector and outputting the hidden state. At the same time, based on the U-Net skip connection structure, the hidden state and text data are connected to generate a report field, thus obtaining the initial output items.

7. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 1, characterized in that: The method for obtaining the output optimization term includes: The credibility scoring mechanism includes a reinforcement learning agent design, which includes a state space and a reward function. The state space is used to generate reports on field completeness, logical consistency, and policy compliance. The reward function is the weighted calculation result of the compliance score and the manual review pass rate, respectively. A correction threshold is set, which is the score of the reward function. When the score is lower than the correction threshold, iterative regeneration is triggered. When the score is higher than the correction threshold, the report is directly output to the blockchain audit layer, thereby obtaining the output optimization items.

8. The method for automatically generating financial and tax digital resources integrating machine learning and artificial intelligence according to claim 1, characterized in that: The methods for obtaining the output result items include: Based on the output optimization terms, the decision points in the generation process are written into the smart contract, and the original data hash and the report hash are stored on a dual-chain basis to obtain the stored information items; Based on the stored information items, a PDF annotation report is generated through SHAP value analysis to mark key decision factors, thereby obtaining the output results.

9. A system for automatically generating financial and tax digital resources that integrates machine learning and artificial intelligence, characterized in that: The method for automatically generating financial and tax digital resources by integrating machine learning and artificial intelligence, as described in any one of claims 1-8, includes: Data acquisition module: Set up a multi-source heterogeneous data input layer to receive multi-source heterogeneous data and obtain input data items; Feature extraction module: Based on the input data items, the feature extraction module extracts the analysis results of text data, image data, time series data and structured data respectively, and then obtains the analysis dataset; Data alignment module: Based on the analysis dataset, modality alignment is performed. A multimodal fusion engine is set up, and feature vectors of different modalities are weighted and aggregated through an attention mechanism to generate a unified embedding representation. This representation is then input into a multimodal generative adversarial network to obtain integrated data items. Optimized output module: Based on the integrated data items, a dynamic generation and optimization module is created. The generator of GAN outputs preliminary financial and tax resources, including draft tax reports, to obtain preliminary output items. A data credibility scoring mechanism is introduced through a discriminator. Based on reinforcement learning feedback, the generation bias is corrected, and the preliminary output items are optimized and corrected to obtain optimized output items. Output annotation module: Based on the output optimization items, the blockchain audit and output layer is set, the generated results are stored in the blockchain after being hashed and encrypted, and the interface report with interpretable annotations is output to obtain the output result items.

Images