A Deep Learning-Based Multi-Source Data Automatic Accounting Method and System
By standardizing multi-source data access, text semantic encoding, and intelligent ticket flow matching, combined with the Sentence-BERT deep learning model and bipartite graph maximum weight matching, the problem of low automation in bookkeeping for micro and small enterprises has been solved. Global optimal matching and accounting subject recommendation have been achieved, thereby improving the level and accuracy of bookkeeping automation.
Patent Information
- Application Number
- CN202611221226.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-17
AI Technical Summary
Existing accounting solutions for micro and small enterprises and the bookkeeping industry suffer from problems such as difficulty in matching transaction summaries, insufficient automated recommendation of accounting subjects, lack of closed-loop learning mechanisms, and weak multi-source data governance capabilities. These issues result in low levels of accounting automation, a large workload for manual review, and insufficient efficiency and accuracy.
By employing multi-source data access standardization, text semantic encoding, intelligent ticket flow matching, accounting subject recommendation and feedback iteration modules, combined with the Sentence-BERT deep learning model and bipartite graph maximum weight matching, the system achieves globally optimal ticket-flow pairing and accounting subject recommendation, forming a closed-loop optimization mechanism.
It has achieved unified governance of multi-source data, improved the level of bookkeeping automation, increased the accuracy of matching and the precision of accounting subject recommendation, reduced the workload of manual finance, and adapted to changing financial and tax business scenarios.
Smart Images

Figure CN123118280A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of intelligent financial accounting, natural language processing and deep learning, and more specifically to a method and system for automatic accounting based on deep learning and multi-source data. Background Technology
[0002] Small and micro-sized enterprises and the bookkeeping industry need to process multiple types of financial and tax data simultaneously, including electronic invoices, bank statements, and electronic receipts, in their daily bookkeeping. Most existing bookkeeping solutions rely on fixed keyword rules to match invoices with transaction records and classify accounting subjects, which has significant shortcomings.
[0003] First, the transaction summary and product name are expressed in various forms, and simple rules are difficult to understand the semantics of the text, which easily leads to matching omissions and mismatches. Moreover, it can only achieve local pairwise matching and cannot obtain the global optimal matching result.
[0004] Secondly, accounting entries rely on manual judgment by financial personnel, and automated recommendations lack accuracy, requiring a large number of vouchers to be entered manually.
[0005] Third, the system lacks a closed-loop learning mechanism, and the rules cannot adapt autonomously after changes in business scenarios;
[0006] Fourth, the original data from multiple sources is often disorganized in format, contains a lot of duplicate data, and has weak front-end data governance capabilities. These problems result in low levels of automation in bookkeeping, a large workload for manual review, and bookkeeping efficiency and data matching accuracy that are difficult to meet the needs of large-scale bookkeeping and intelligent corporate finance. Summary of the Invention
[0007] The purpose of this invention is to provide a method and system for automatic accounting based on multi-source data, in order to solve the above-mentioned technical problems.
[0008] To achieve the above objectives, the present invention specifically adopts the following technical solution: a multi-source data automatic accounting system based on deep learning, characterized in that it includes: a multi-source data access standardization module, a text semantic encoding module, a ticket flow intelligent matching module, an accounting subject recommendation and voucher generation module, and a feedback iteration module;
[0009] The multi-source data access standardization module is used to collect multi-source structured data such as electronic invoices, bank statements, and electronic receipts, and to complete text cleaning, format unification, invalid data removal and deduplication processing. It outputs standardized invoice datasets and standardized transaction datasets, and pushes them to the text semantic encoding module and the invoice flow intelligent matching module.
[0010] The text semantic encoding module uses a Sentence-BERT deep learning model fine-tuned with a financial and tax corpus to convert invoice product descriptions and bank transaction summaries into fixed-dimensional semantic vectors; a pre-built accounting subject vector index library is used for accounting subject semantic retrieval.
[0011] The intelligent ticket flow matching module receives standardized tickets, transaction data, and semantic vectors, calculates the comprehensive confidence score of candidate pairings of tickets and transactions, constructs a bipartite graph with tickets as the first node set and transactions as the second node set, uses the confidence score as the edge weight, solves the maximum weight matching of the bipartite graph, outputs the globally optimal ticket-fund pairing relationship, and marks low-confidence abnormal pairings.
[0012] The accounting subject recommendation and voucher generation module obtains candidate accounting subjects based on invoice semantic vector retrieval, and calculates the subject confidence level by weighting the paired fund information; based on the subject results and invoice flow pairing relationship, it automatically generates accounting voucher drafts, and distributes them to the automatic accounting queue or manual review queue according to the confidence level.
[0013] The feedback iteration module collects operation records of financial personnel manually adjusting pairing relationships and modifying accounting subjects to form labeled samples, and periodically updates the semantic vector model and pairing confidence calculation parameters to form a closed-loop optimization.
[0014] Furthermore, a multi-source data automatic accounting method based on deep learning includes the following steps:
[0015] S1 Multi-Source Data Acquisition Standardization: Acquire multi-source structured data such as electronic invoices, bank statements, and electronic receipts, perform cleaning, deduplication, and format normalization to obtain standardized invoice datasets and standardized transaction datasets;
[0016] S2 Text Semantic Vectorization: Using a fine-tuned Sentence-BERT deep learning model to encode invoice product names and bank transaction summaries, a corporate accounting subject vector index library is constructed.
[0017] S3 Global Optimal Pairing of Invoices and Transactions: Calculate the multidimensional features and comprehensive confidence scores of candidate pairings of invoices and transactions, construct a bipartite graph and solve for the maximum weighted matching, output the optimal pairing results, and identify and mark low-confidence abnormal pairings;
[0018] S4 intelligently recommends accounting subjects and generates vouchers: it uses vector retrieval to obtain candidate accounting subjects, integrates fund matching information to calculate the final subject confidence level, automatically generates draft accounting vouchers, and distributes them according to confidence level.
[0019] S5 closed-loop iteration: Collect manually reviewed and modified samples, update semantic vector model parameters and pairing weight parameters, and continuously optimize the system's matching accuracy and subject recognition effect.
[0020] The beneficial effects of this invention are as follows:
[0021] 1. Structure optimization: Build an integrated architecture for multi-source data standardization, semantic encoding, invoice flow matching, voucher generation, and feedback iteration, and connect the data links of invoices and bank statements to achieve unified governance of financial and tax data.
[0022] 2. Performance Improvement: By leveraging fine-tuned Sentence-BERT to mine deep semantics of text, and employing bipartite graph maximum weight matching to achieve globally optimal ticket flow pairing, the matching accuracy is improved compared to simple rule matching, and abnormal business is accurately identified; accounting subjects are efficiently recommended by using subject vector retrieval.
[0023] 3. Work Condition Adaptation: Differentiating between automatic bookkeeping and manual review scenarios reduces the workload of finance personnel; through continuous iteration of models and matching parameters via manual operation, the system has self-optimization capabilities, adapting to diverse financial and tax business scenarios and improving the level of bookkeeping automation. Attached Figure Description
[0024] Figure 1 This is a block diagram of the overall architecture of the multi-source data automatic accounting system based on deep learning, according to an embodiment of the present invention.
[0025] Figure 2 This is a schematic diagram illustrating the internal operating principle of the ticket flow intelligent matching module in an embodiment of the present invention.
[0026] Figure 3 This is a schematic diagram of the overall process of the multi-source data automatic accounting method based on deep learning in an embodiment of the present invention. Detailed Implementation
[0027] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0028] Example 1: Basic Implementation (Bipartite Graph Maximum Weight Matching + Sentence-BERT Deep Learning Scheme)
[0029] like Figure 1 As shown, this embodiment provides a multi-source data automatic accounting system based on deep learning, including a multi-source data access standardization module, a text semantic encoding module, a ticket flow intelligent matching module, an accounting subject recommendation and voucher generation module, and a feedback iteration module. Each module works in sequence to collaboratively complete multi-source data processing, semantic encoding, ticket flow matching, subject recommendation, voucher generation, and iterative optimization.
[0030] The multi-source data access standardization module connects to the electronic invoice platform and the bank-enterprise direct connection interface to collect multi-source structured data such as electronic invoices, bank statements, and electronic receipts, including core information such as invoice code, invoice date, amount, product name, counterparty, transaction date, and transaction summary; it removes special symbols, standardizes date and amount formats, eliminates duplicate and invalid data, and outputs standardized invoice datasets and standardized transaction datasets, which are simultaneously pushed to the text semantic encoding module and the invoice flow intelligent matching module.
[0031] The text semantic encoding module loads a Sentence-BERT deep learning model that has been fine-tuned with financial and tax business corpus, transforming invoice product descriptions and bank transaction summaries into fixed-dimensional dense semantic vectors; it also uniformly encodes all accounting subject names and business definitions of the enterprise, establishing a fast-searchable subject vector index library to support subsequent subject recommendations.
[0032] The intelligent ticket flow matching module calculates a comprehensive confidence score for each candidate pairing of tickets and transaction records based on factors such as transaction entity similarity, time difference, amount difference, and text semantic similarity. It then constructs a bipartite graph model, using the confidence score as edge weights, and solves for the maximum weighted matching in the bipartite graph to obtain the globally optimal ticket-transaction pairing. Pairings with confidence scores below a preset threshold are marked as abnormal and pushed for manual verification. The internal processing logic of the intelligent ticket flow matching module is as follows: Figure 2 As shown.
[0033] The accounting subject recommendation and voucher generation module uses invoice semantic vectors to retrieve multiple candidate subjects from the subject index library. It then combines these with risk information from fund matching to weight the optimal accounting subject and its confidence score. The module automatically generates draft accounting vouchers. High-confidence vouchers are automatically posted, medium-confidence vouchers are subject to manual review, and low-confidence vouchers are subject to mandatory warnings, balancing automation efficiency with financial and tax compliance.
[0034] The feedback iteration module captures operation records of financial personnel modifying pairing relationships and adjusting accounting subjects to form labeled samples. It also updates and optimizes the Sentence-BERT deep learning model and pairing feature weights on a regular basis to continuously improve the system's matching accuracy and subject recognition effect.
[0035] Example 2: First Optional Enhancement Implementation (Zero-Shot Cold Start Scheme for Large Language Models)
[0036] This embodiment reuses Figure 1 System architecture, Figure 3 The overall process shown does not require additional diagrams.
[0037] To address the scenario where newly established enterprises lack historical accounting samples, hindering effective training of deep learning models, the vector retrieval and bipartite graph matching links are disabled. Multi-source data is assembled into standardized prompt words and input into a large language model, which directly outputs invoice-transaction log pairing suggestions and recommended accounting subjects. As business accumulates and manually labeled samples reach a certain threshold, the system can smoothly switch to the basic solution of Example 1, achieving a seamless transition from cold start to regular operation.
[0038] Example 3: A Multi-Source Data Automated Accounting Method Based on Deep Learning
[0039] This embodiment provides a multi-source data automatic accounting method based on deep learning, which can run in the system described in Embodiment 1, such as... Figure 3 As shown, it includes the following steps:
[0040] S1 Multi-Source Data Acquisition Standardization: Collects multi-source structured data such as electronic invoices, bank statements, and electronic receipts, performs text cleaning, format unification, and deduplication, and outputs standardized invoice datasets and standardized transaction datasets;
[0041] S2 Text Semantic Vectorization: The financial and tax fine-tuning Sentence-BERT deep learning model is used to encode the commodity name on the invoice and the bank transaction summary to build a vector index library of enterprise accounting subjects;
[0042] S3 Global Pairing of Invoices and Transactions: Calculate the comprehensive confidence score of candidate pairings of invoices and transactions, construct a bipartite graph to solve for the maximum weighted matching, obtain the optimal pairing relationship, and mark low-confidence abnormal pairings;
[0043] S4 Subject Recommendation and Voucher Generation: Vector retrieval of candidate accounting subjects, integration of fund matching information to calculate the final subject confidence level, automatic generation of accounting voucher drafts, and hierarchical distribution according to confidence level;
[0044] S5 closed-loop iteration: Collect manually reviewed and modified samples, update deep learning model parameters and pairing weight parameters, and continuously optimize the system matching and subject recognition effects.
[0045] It should be understood that this invention is not limited to the embodiments described above. Any conventional adjustments to the deep learning model type, matching algorithm, threshold parameters, and interface form without departing from the core concept of this invention are all within the scope of protection of this invention.
Claims
1. A multi-source data automatic accounting system based on deep learning, characterized in that, include: The system includes a multi-source data access standardization module, a text semantic encoding module, an intelligent ticket flow matching module, an accounting subject recommendation and voucher generation module, and a feedback iteration module. The multi-source data access standardization module is used to collect multi-source structured data such as electronic invoices, bank statements, and electronic receipts, and to complete text cleaning, format unification, invalid data removal and deduplication. It outputs standardized invoice datasets and standardized transaction datasets, and pushes them to the text semantic encoding module and the invoice flow intelligent matching module. The text semantic encoding module uses a Sentence-BERT deep learning model fine-tuned with a financial and tax corpus to convert invoice product descriptions and bank transaction summaries into fixed-dimensional semantic vectors; a pre-built accounting subject vector index library is used for accounting subject semantic retrieval. The intelligent ticket flow matching module receives standardized tickets, transaction data, and semantic vectors, calculates the comprehensive confidence score of candidate pairings of tickets and transactions, constructs a bipartite graph with tickets as the first node set and transactions as the second node set, uses the confidence score as the edge weight, solves the maximum weight matching of the bipartite graph, outputs the globally optimal ticket-fund pairing relationship, and marks low-confidence abnormal pairings. The accounting subject recommendation and voucher generation module obtains candidate accounting subjects based on invoice semantic vector retrieval, and calculates the subject confidence level by weighting the paired fund information; based on the subject results and invoice flow pairing relationship, it automatically generates accounting voucher drafts, and distributes them to the automatic accounting queue or manual review queue according to the confidence level. The feedback iteration module collects operation records of financial personnel manually adjusting pairing relationships and modifying accounting subjects to form labeled samples, and periodically updates the semantic vector model and pairing confidence calculation parameters to form a closed-loop optimization.
2. The multi-source data automatic accounting system based on deep learning according to claim 1, characterized in that, The system is configured with a cold start alternative operation plan; when there are insufficient historical accounting samples for the enterprise, it switches to the zero-sample inference mode of the large language model, directly outputting matching suggestions and recommended accounting subjects by inputting multi-source data text information; when the manually labeled samples accumulate to the threshold, it can smoothly switch to the basic solution of bipartite graph matching + semantic vector retrieval.
3. The multi-source data automatic accounting system based on deep learning according to claim 1, characterized in that, The comprehensive confidence score is calculated by weighted fusion of transaction entity similarity, transaction time difference, transaction amount difference, and text semantic similarity.
4. The multi-source data automatic accounting system based on deep learning according to claim 1, characterized in that, The accounting subject recommendation and voucher generation module is configured with hierarchical and tiered rules: vouchers with a confidence level higher than the first threshold enter the automatic accounting queue; vouchers with a confidence level between the first and second thresholds enter the manual review queue; vouchers with a confidence level lower than the second threshold are subject to mandatory warnings.
5. A multi-source data automatic accounting method based on deep learning, characterized in that, Applied to the system according to any one of claims 1 to 4, comprising the following steps: S1 Multi-Source Data Acquisition Standardization: Acquire multi-source structured data such as electronic invoices, bank statements, and electronic receipts, perform cleaning, deduplication, and format normalization to obtain standardized invoice datasets and standardized transaction datasets; S2 Text Semantic Vectorization: Using a fine-tuned Sentence-BERT deep learning model to encode invoice product names and bank transaction summaries, a corporate accounting subject vector index library is constructed. S3 Global Optimal Pairing of Invoices and Transactions: Calculate the multidimensional features and comprehensive confidence scores of candidate pairings of invoices and transactions, construct a bipartite graph and solve for the maximum weighted matching, output the optimal pairing results, and identify and mark low-confidence abnormal pairings; S4 intelligently recommends accounting subjects and generates vouchers: it uses vector retrieval to obtain candidate accounting subjects, integrates fund matching information to calculate the final subject confidence level, automatically generates draft accounting vouchers, and distributes them according to confidence level. S5 closed-loop iteration: Collect manually reviewed and modified samples, update semantic vector model parameters and pairing weight parameters, and continuously optimize the system's matching accuracy and subject recognition effect.
