A tax policy document processing method based on multi-modal analysis

Abstract

The application discloses a tax policy file processing method based on multi-modal analysis, and belongs to the technical field of tax policy file processing. The method comprises the following steps: according to a tax policy file, a region is divided to identify a region type and obtain a confidence degree; the confidence degrees are sorted from large to small to obtain a first confidence degree and a second confidence degree; when the difference between the first confidence degree and the second confidence degree is greater than or equal to a confidence threshold, it is determined that the file region is the region type corresponding to the first confidence degree; otherwise, according to the type coincidence rate of the two region types corresponding to the first confidence degree and the second confidence degree, the region type corresponding to the file region is adjusted; according to the obtained region type, data preprocessing is performed on the data in the file region, single-modal features are extracted, and multi-modal fusion is performed on the single-modal features obtained from multiple file regions in the tax policy file to obtain a knowledge graph. Through accurate division of the region, multi-modal fusion is performed to realize accurate analysis of the file.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of tax policy document processing, and particularly relates to a tax policy document processing method based on multi-modal analysis. BACKGROUND

[0002] With the deepening of the digital transformation of tax collection and management in China, the electrification and digitization of tax preferential policy documents have become an important foundation for tax management. Currently, tax preferential policy documents are mainly published in unstructured forms such as PDF and images, containing various modal data such as text, tables, and formulas. In these documents, key elements such as tax types, document numbers, preferential conditions, and applicable subjects are scattered in different modal carriers such as text, tables, and formulas, which seriously hinders the accurate delivery and efficient application of tax policies.

[0003] Early technologies mainly relied on regular rule matching, with matching rules manually formulated for text analysis, but this method had limited semantic nesting processing capability for policy texts and could not adapt to the complexity of policy expression.

[0004] With the development of OCR technology, tax systems began to use OCR technology to extract text from images and PDF files, but this technology only processed text content and ignored the relevance between tables and formulas.

[0005] In recent years, some enterprises have begun to try to use multi-modal analysis technology to process tax policy documents, but when dividing regions of tax policy documents, they lack accurate identification mechanisms for text, table, and formula modalities. When a document region meets the characteristics of two modalities at the same time (such as formula embedded in table), it cannot be dynamically adjusted according to the actual content, resulting in incorrect region type identification.

[0006] In addition, existing technologies mostly use single-modal independent analysis strategies, i.e., analyzing text, tables, and formulas separately and then simply concatenating the results, lacking mapping mechanisms for cross-modal elements and insufficient multi-modal feature fusion. Due to the above problems, existing technologies generally have low analysis accuracy when processing tax policy documents.

[0007] These problems seriously hinder the effectiveness of tax policy digitization applications, causing tax systems to be unable to accurately identify and apply preferential policies, affecting the efficiency and accuracy of policy implementation. Therefore, there is an urgent need for an analysis method that can accurately identify the region types of multi-modal tax policy documents and effectively fuse multi-modal features to solve the problems of low analysis accuracy and incomplete element correlation in tax preferential policy document analysis. SUMMARY

[0008] The application provides a tax policy document processing method based on multi-modal analysis, which solves the problems of low analysis accuracy and incomplete element correlation of tax preferential policy documents by dividing regions and fusing multi-modalities to obtain a knowledge graph for analyzing the tax policy documents.

[0009] The technical solution adopted by the application is:

[0010] A tax policy document processing method based on multi-modal analysis, comprising:

[0011] According to the collected tax policy documents, regions are divided to identify region types and obtain confidence degrees corresponding to the region types one by one, wherein the region types include text regions, table regions and formula regions;

[0012] According to the confidence degrees from high to low, first and second confidence degrees are obtained in sequence, and when the difference between the first and second confidence degrees is greater than or equal to a confidence threshold, it is determined that the file region is the region type corresponding to the first confidence degree;

[0013] Otherwise, according to the type coincidence rate of the two region types corresponding to the first and second confidence degrees, the region type corresponding to the file region is adjusted;

[0014] According to the obtained region types, data preprocessing is performed on the data in the file region, single-modal features are extracted, and the single-modal features obtained from multiple file regions in the tax policy document are fused to obtain a knowledge graph for analyzing the tax policy document.

[0015] The tax policy document processing method based on multi-modal analysis disclosed in the application has the following additional technical features:

[0016] The region types are identified, specifically:

[0017] The region division is performed on each page of the tax policy document by a modal feature detector,

[0018] wherein the text regions are identified according to the number of consecutive character blocks,

[0019] the table regions are identified according to the table boundaries and row and column structures,

[0020] and the formula regions are identified according to the digits and expressions.

[0021] The type coincidence rate of the two region types corresponding to the first and second confidence degrees is specifically:

[0022] The type coincidence rate includes a type coincidence area, a first type coincidence rate of the type coincidence area accounting for the type region corresponding to the first confidence degree,

[0023] and the type overlap area of the type corresponding to the second confidence.

[0024] According to the type overlap rate of the two region types corresponding to the first confidence and the second confidence, the region type corresponding to the file region is determined, specifically:

[0025] If the first type overlap rate and the second type overlap rate are both greater than the overlap threshold, it is judged that the file region corresponds to the two region types of the first confidence and the second confidence.

[0026] If one of the first type overlap rate and the second type overlap rate is greater than the overlap threshold, it is judged that the file region corresponds to the region type of one of the first type overlap rate and the second type overlap rate.

[0027] If the first type overlap rate and the second type overlap rate are both less than the overlap threshold, the first confidence and the second confidence are weighted and adjusted according to the first type overlap rate and the second type overlap rate, so as to determine the region type corresponding to the file region.

[0028] After obtaining the region type, the following steps are further included:

[0029] According to the order of the file region in the tax policy file, the file region is arranged,

[0030] When the region types corresponding to adjacent file regions are consistent, the terms of the two file regions are extracted to obtain a term matching degree.

[0031] The correlation relationship of a plurality of terms corresponding to the two file regions is obtained to obtain the context consistency of the two file regions.

[0032] According to the term matching degree and the context consistency, a boundary confidence is obtained, and the boundary is adjusted according to the boundary confidence.

[0033] According to the obtained region type, data preprocessing is performed on the data in the file region, and single-modal features are extracted, including:

[0034] According to the text of the text region, text standardization is performed through a tax field term dictionary, tax content is recognized, and a word vector is obtained.

[0035] According to the order of the word vector in the text region, forward and backward splicing is performed to obtain a plurality of comprehensive vectors, wherein adjacent comprehensive vectors have an overlap region.

[0036] According to the syntax of the comprehensive vector in the text of the text area, the dependency relationship between the word vectors in the comprehensive vector is recognized through a syntax tree.

[0037] Multi-modal fusion is performed, specifically:

[0038] According to the word vectors, nodes are constructed, and the edge relationship between the nodes is constructed according to the dependency relationship between the word vectors.

[0039] According to the word vectors, the single-modal features of the formula area recognition are called to fill the formula in the edge relationship.

[0040] According to the word vectors, data filling is performed with the single-modal features of the table area recognition to obtain the knowledge graph.

[0041] The application also provides a computer program product,

[0042] The computer program is executed by a processor to implement the tax policy file processing method based on multi-modal analysis.

[0043] The application also provides a storage medium,

[0044] The storage medium stores a computer program, and the computer program is executed to implement the steps of the tax policy file processing method based on multi-modal analysis.

[0045] The application further provides a processing device, comprising:

[0046] A memory is configured to store a computer program.

[0047] A processor is configured to execute the computer program to implement the steps of the tax policy file processing method based on multi-modal analysis.

[0048] Due to the adoption of the above technical solutions, the application has the following advantages:

[0049] 1. In the application, according to the collected tax policy file, the region type is identified by region division, and the confidence degree corresponding to the region type is obtained; the confidence degrees are sorted from large to small, and the first confidence degree and the second confidence degree are obtained; when the difference between the first confidence degree and the second confidence degree is greater than or equal to the confidence threshold, it is determined that the file region is the region type corresponding to the first confidence degree.

[0050] When there is a significant gap between the two high confidences, it indicates that there is a significant difference in the probability between the two identified region types, and the confidence of the region type corresponding to the first confidence is increased, and the confidence of the region type corresponding to the second confidence is reduced, so as to determine that the file region is the region type corresponding to the first confidence, and the accuracy of region type identification is improved.

[0051] However, when the first confidence and the second confidence are both greater than or both less than the fixed threshold, there is a dilemma in the judgment. In addition, due to minor interference, the confidence may fluctuate, resulting in type misjudgment.

[0052] The present application introduces a difference comparison mechanism, when there is a significant gap between the two high confidences, the region type corresponding to the highest confidence is directly used, which solves the type misjudgment problem and the dilemma problem caused by minor interference, and improves the processing efficiency and accuracy in simple scenarios.

[0053] In addition, otherwise, according to the type coincidence rate of the two region types corresponding to the first confidence and the second confidence, the region type corresponding to the file region is adjusted. When the first confidence and the second confidence are close, the traditional method often falls into a dilemma or random selection. The present application introduces type coincidence rate analysis, which accurately judges by calculating the overlap degree of the two region types in space and semantics, solves the modality recognition problem in complex scenarios, and improves the recognition accuracy.

[0054] According to the obtained region type, data preprocessing is performed on the data in the file region, single-modal features are extracted, and the single-modal features obtained from multiple file regions in the tax policy file are fused to obtain a knowledge graph for analyzing the tax policy file. Based on the accurate region type identification result, the present application first extracts single-modal features of each modality, and then performs cross-modal fusion, realizes the leap from information fragments to knowledge system, and improves the correlation completeness of tax elements. Especially in the processing of complex scenarios such as superimposed preferential policies and progressive calculation, the accuracy and integrity of the knowledge graph are significantly better than those of the traditional method, which provides reliable support for tax intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0055] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0056] Figure 1 The figure is a flowchart of the tax policy file processing method based on multi-modal analysis according to an embodiment of the present application. DETAILED DESCRIPTION

[0057] In order to more clearly illustrate the overall concept of the present application, the following detailed description will be made in conjunction with the accompanying drawings.

[0058] In the following description, a number of specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without these specific details. In other instances, well-known methods have not been described in detail in order to avoid unnecessarily obscuring the present application.

[0059] As shown in Figure 1 A tax policy document processing method based on multi-modal analysis, comprising:

[0060] S100: According to the tax policy document collected, through regional division, the region type is identified, and the confidence corresponding to the region type is obtained, wherein the region type includes text region, table region and formula region.

[0061] The core purpose of this step is to solve the technical problem of inaccurate multi-modal region identification in tax policy documents, and to provide high-quality data basis for subsequent multi-modal feature extraction and fusion.

[0062] Specifically, this step aims to accurately identify the text region, table region and formula region of the tax policy document, and solve the key problems of low tax element analysis accuracy and incomplete element association caused by incorrect region type identification in the prior art.

[0063] An improved modal feature detector (a deep learning model based on CNN+ResNet) is used to detect the multi-modal features of each page of the tax policy document to realize regional division. The region type includes text region, table region and formula region. It can be understood that the text region contains a lot of text information, which is used to identify and extract keywords and the relationship between keywords. The table region contains a lot of data information, which is used to complete the data of the keywords. The formula region contains a lot of formulas, which are used to define the relationship between keywords in detail to realize the mutual deduction and completion of the data corresponding to each keyword.

[0064] In addition, after regional division, the modal feature detector can output the confidence corresponding to each region type, for example, the text region confidence is 0.8, the table region confidence is 0, and the formula region confidence is 0.2. In order to determine the subsequent modal type through the confidence.

[0065] It should be noted that the confidence output by the above modal feature detector can also be an initial confidence, and the OCR recognition accuracy and text continuity judgment are performed on the text region to obtain a text confidence, i.e., C_text = (OCR recognition accuracy × 0.6) + (text continuity score × 0.4).

[0066] In addition, the table structure integrity and cell recognition accuracy are judged on the table region to obtain a table confidence, i.e., C_table = (table structure integrity score × 0.5) + (cell recognition accuracy × 0.5).

[0067] The mathematical symbol recognition accuracy and formula structure integrity are judged on the formula region to obtain a formula confidence, i.e., C_formula = (mathematical symbol recognition accuracy × 0.7) + (formula structure integrity × 0.3).

[0068] The initial confidence is adjusted one by one through the text confidence, the table confidence, and the formula confidence to determine the final confidence.

[0069] Accurate region type recognition is a key prerequisite for realizing multi-modal analysis of tax policy documents. Through high-quality region division results, a reliable data foundation is provided for subsequent single-modal feature extraction and cross-modal feature fusion, improving the accuracy of tax policy element analysis.

[0070] S200: According to the confidence from high to low, the first confidence and the second confidence are obtained in turn, and when the difference between the first confidence and the second confidence is greater than or equal to a confidence threshold, it is determined that the file region is the region type corresponding to the first confidence;

[0071] Otherwise, according to the type coincidence rate of the two region types corresponding to the first confidence and the second confidence, the region type corresponding to the file region is adjusted.

[0072] The core purpose of this step is to solve the ambiguity problem of multi-modal region type recognition in tax policy documents, especially when the confidence of two region types (such as text and table, table and formula, and text and formula) is close, how to make accurate judgment.

[0073] The prior art only uses a fixed threshold (such as >0.8) for region type determination, which cannot effectively handle overlapping regions with close confidence, resulting in low accuracy of tax policy document analysis. This step introduces a confidence difference comparison and type coincidence rate analysis mechanism to realize dynamic adjustment of region type recognition, ensuring high-precision region type determination even in the case of ambiguity.

[0074] Sort the three region type confidences in descending order to obtain a first confidence and a second confidence. Calculate the difference between the first confidence and the second confidence ΔC = |C_first - C_second|.

[0075] Set a confidence threshold, for example 0.15. When ΔC ≥ 0.15, determine the file region as the region type corresponding to the first confidence. When ΔC < 0.15, enter the overlapping region processing flow.

[0076] The prior art cannot make accurate judgments when the confidences are close, resulting in misjudgment of region type. This step avoids the dilemma through a difference comparison mechanism, improving the accuracy of region type determination when the confidences are close.

[0077] It should be noted that the overlapping region is defined by common features, including: spatial position overlap, semantic content similarity, modal feature matching degree, etc. The type overlap rate R_overlap = (number of common features / total number of features) × 100%.

[0078] Through the type overlap rate analysis mechanism, the system can accurately identify the true type of the overlapping region, improving the accuracy of overlapping region recognition.

[0079] In addition, it should be noted that the confidence threshold is dynamically adjusted according to historical processing data. For example, the threshold adjustment formula: T = 0.15 + (0.05 × (1 - η)). Where η is the average accuracy of the current modal in historical processing. Specifically, when the historical accuracy η > 0.85, the confidence threshold is automatically reduced to 0.10, improving the sensitivity.

[0080] Fixed threshold determination is sensitive to minor interference, resulting in a high misjudgment rate. For example, the "annual taxable income ≤ 300 million" in the table is misjudged as a text region, affecting the accurate extraction of preferential conditions. The dynamic adjustment mechanism of this step reduces the misjudgment rate and significantly improves the system robustness.

[0081] This step solves the core problem of region type recognition in tax policy documents through innovative confidence difference comparison and type overlap rate analysis mechanisms, which is a key link to realize multi-modal intelligent analysis of tax policy documents. Not only significantly improves the accuracy of region type recognition, but also provides high-quality data input for subsequent tax element structured processing, making the overall precision of tax policy document analysis improve qualitatively.

[0082] S300: Based on the obtained region type, perform data preprocessing on the data within the file region, extract unimodal features, and perform multimodal fusion using the unimodal features obtained from multiple file regions in the tax policy document to obtain a knowledge graph for parsing the tax policy document.

[0083] The core objective of this step is to address the issue of incomplete associations among multimodal elements in tax policy documents. Through precise single-modal feature extraction and an innovative multimodal fusion mechanism, it enables the intelligent association of key elements scattered across different modalities such as text, tables, and formulas in tax policy documents, thereby constructing a high-quality tax policy knowledge graph.

[0084] Existing technologies can only analyze a single modality and lack cross-modal element mapping capabilities, resulting in the inability to accurately correlate key elements. This step, through a systematic multimodal fusion process, achieves precise correlation of tax elements and structured construction of a knowledge graph, laying the technological foundation for the intelligent application of tax incentive policies.

[0085] Preprocess the text region, table region, and formula region separately to extract the corresponding text features, table features, and formula features.

[0086] Specifically, for table area preprocessing, a table topology repair algorithm (based on GCN) is applied to correct table structures that have errors in OCR recognition; key elements in the table are extracted through table cell content analysis; and the topological relationships of table rows, columns, and cells are established.

[0087] For formula preprocessing, the formulas are converted into symbol tree structures. Through dynamic parameterization technology, variables in the formulas are replaced with placeholders, and key calculation elements in the formulas are extracted.

[0088] To achieve 3D feature alignment, a cross-modal attention mechanism is constructed to align text, table, and formula 3D features. Explicit alignment between modalities is achieved through contrastive learning, addressing the semantic differences between modalities. Text feature vectors, table feature vectors, and formula feature vectors are fused using tensors, and a cross-modal fusion layer (based on a Transformer-based multi-head attention mechanism) is applied to generate a fused feature vector. This fused feature vector enables cross-modal association between elements.

[0089] It should be noted that, to evaluate the quality of fusion, the diversity index of fused features (based on the entropy value of the feature vector) is calculated; if the entropy value is <0.3, a feature enhancement mechanism (multi-model integration, domain knowledge injection, etc.) is activated to ensure the high quality and integrity of the fused features.

[0090] A dynamic knowledge graph of tax elements is constructed based on GNN, where nodes represent tax elements and edges represent the relationships between elements. It supports multi-hop reasoning and enables in-depth mining of complex element relationships.

[0091] The prior art can only parse a single modality when processing tax policy documents, cannot establish cross-modality element association, and can cause conditional recognition errors. The multi-modality fusion mechanism of this step improves the completeness rate of tax element association.

[0092] Key elements in tax policy documents are scattered in different modalities, and accurate single-modality feature extraction and fusion are needed to accurately identify them. This step improves the accuracy of key element identification through single-modality feature extraction and multi-modality fusion.

[0093] This step realizes the accurate association of scattered elements and the structured construction of knowledge graph in tax policy documents through innovative single-modality feature extraction and multi-modality fusion mechanism, which is the core link of intelligent analysis of tax policy documents. Not only does it solve the problem of incomplete multi-modality element association in existing technology, but also provides a high-quality data basis for intelligent application of tax preferential policies, improves the overall precision and practicality of tax policy document analysis, and provides key technical support for tax digital transformation.

[0094] As a preferred embodiment of the present application, the recognition area type is specifically:

[0095] Through the modality feature detector, region division is performed on each page of the tax policy file,

[0096] wherein, according to the number of continuous text blocks, a text region is identified,

[0097] According to the table boundary and row-column structure, a table region is identified,

[0098] According to the number and expression of digits, a formula region is identified.

[0099] The core purpose of this embodiment is to solve the problem of accurate multi-modality area recognition of tax policy documents, and to provide a high-quality data basis for subsequent single-modality feature extraction and multi-modality fusion.

[0100] Tax policy documents contain a large amount of professional text, complex tables and mathematical formulas, and traditional OCR technology cannot effectively distinguish these area types, leading to subsequent element analysis errors. This embodiment realizes the accurate separation of text, table and formula regions in tax policy documents through targeted modality feature detection mechanism, improves the accuracy of area recognition, and lays a reliable foundation for structured processing of tax elements.

[0101] For text area recognition, the number of continuous text blocks is detected by horizontal scanning of the page, and the number of characters in the continuous text block is counted. When the length of the continuous text block is greater than 50 characters, it is determined as a text area. Non-text content is filtered through a tax field dictionary. The misjudgment problem of dense tax policy Chinese number area is solved to avoid misjudging the number as a table.

[0102] For table area recognition, table boundary detection is adopted, and horizontal and vertical lines in the page are identified by Hough transform. Row and column structure analysis is performed based on connected component analysis to calculate the number of rows and columns, construct a table topology graph, and determine the table area. Accurate recognition of complex table structure solves the row and column misplacement problem caused by fuzzy table lines in traditional methods.

[0103] For formula area recognition, digit symbol recognition is performed based on a mathematical symbol library (containing 500+ tax formula symbols such as "<", "x", "IF", etc.) for matching. Expression structure analysis is performed to construct a formula symbol tree and analyze the logical structure of the expression to determine the formula area. The tax formula recognition problem is solved to avoid misjudging the formula as ordinary text.

[0104] Region recognition errors cause element association failure. The present embodiment realizes accurate region recognition through tax field special feature detection, improves the completeness rate of tax element association, and improves the accuracy rate of subsequent single-modal feature extraction.

[0105] The present embodiment accurately solves the separation problem of text, table, and formula areas through a modal feature detection mechanism designed for the characteristics of tax policy documents, which is a key technical link for multi-modal analysis of tax policy documents.

[0106] As a preferred embodiment of the present application, the type coincidence rate of the two region types corresponding to the first confidence and the second confidence is:

[0107] The type coincidence rate includes a type coincidence area, which accounts for a first type coincidence rate of the first confidence corresponding type area,

[0108] and the type coincidence area accounts for a second type coincidence rate of the second confidence corresponding type area.

[0109] The core purpose of the present embodiment is to solve the ambiguity problem of region type recognition in tax policy documents, especially when the confidence of two region types is close, how to accurately determine the real type of the region.

[0110] The tax policy document often contains overlapping scenarios such as "formula embedded table" (e.g., "deduction amount = IF (revenue < 5 million, revenue x 5%, 250,000 + excess part x 7%)" and "text containing formula" (e.g., "annual taxable income ≤ 3 million part reduced by 25% into taxable amount"). Traditional methods cannot accurately determine the region type due to the reliance on confidence difference.

[0111] The number of features corresponding to the region type of the first confidence is identified, the number of features corresponding to the region type of the second confidence is identified, and then the number of features belonging to both region types is identified, i.e., the type overlap region.

[0112] The first type overlap rate R_first = (type overlap feature number / first confidence corresponding type region feature number) x 100%. The second type overlap rate R_second = (type overlap feature number / second confidence corresponding type region feature number) x 100%.

[0113] Specifically, according to the type overlap rates of the two region types corresponding to the first confidence and the second confidence, the region type corresponding to the file region is adjusted and determined, specifically:

[0114] If the first type overlap rate and the second type overlap rate are both greater than the overlap threshold, it is determined that the file region corresponds to the two region types of the first confidence and the second confidence;

[0115] If one of the first type overlap rate and the second type overlap rate is greater than the overlap threshold, it is determined that the file region corresponds to the region type of one of the first type overlap rate and the second type overlap rate;

[0116] If the first type overlap rate and the second type overlap rate are both less than the overlap threshold, the first confidence and the second confidence are weighted and adjusted according to the first type overlap rate and the second type overlap rate to determine the region type corresponding to the file region.

[0117] It should be noted that the overlap threshold is set to 70% (through statistical analysis, when the overlap rate > 70%, the region type is dominant).

[0118] Both double overlap rates are > threshold (R_first > 70% and R_second > 70%), identified as double modal region (e.g., table and formula overlap region). At this time, for this file region, both need to be preprocessed and feature extracted through the method corresponding to the double modal region to avoid missing feature extraction.

[0119] In a single coincidence rate > threshold value scenario (R_first > 70% and R_second ≤ 70% or R_first ≤ 70% and R_second > 70%), the region type with a coincidence rate > threshold value is selected as the final determination. For example, R_first = 78% (table) and R_second = 65% (formula), the table region is determined. Misjudgment of the region is avoided, and the accuracy of key calculation element identification is improved.

[0120] In a double coincidence rate ≤ threshold value scenario (R_first ≤ 70% and R_second ≤ 70%), the weighted confidence new_conf_first = (R_first × conf_first) / (R_first + R_second) is calculated; the weighted confidence new_conf_second = (R_second × conf_second) / (R_first + R_second) is calculated. new_conf_first and new_conf_second are compared, and the region type corresponding to the larger value is selected. Avoiding simple averaging ensures that regions with high coincidence rates have higher weights.

[0121] The embodiment improves the accuracy of tax policy document region identification, reduces the misjudgment rate, improves the completeness of tax element association, and provides a reliable basis for accurate push of tax preferential policies.

[0122] As a preferred embodiment of the present application, after obtaining the region type, the following steps are further included:

[0123] The file regions are arranged according to their order in the tax policy file,

[0124] When the region types corresponding to adjacent file regions are consistent, the terms of the two file regions are extracted to obtain a term matching degree;

[0125] The association relationship of multiple terms corresponding to the file regions is obtained, and the context consistency of the two file regions is obtained;

[0126] According to the term matching degree and the context consistency, a boundary confidence is obtained, and the boundary is adjusted according to the boundary confidence.

[0127] The core purpose of the embodiment is to solve the problem of inaccurate identification of text region boundaries in tax policy documents, especially the problem of boundary identification error caused by tax professional terms crossing multiple text blocks.

[0128] The traditional method only relies on the OCR recognition result for boundary division, which cannot accurately identify the boundaries of tax professional terms, leading to errors in subsequent element extraction. The present embodiment introduces term matching degree and context consistency analysis to realize dynamic optimization of the region boundary, improve the accuracy of tax professional term boundary identification, and provide high-quality input basis for tax element structured processing.

[0129] Linearly order the file regions in the tax policy file according to their physical positions (from top to bottom, left to right), establish a region index table, and record the starting position and ending position of each region. Ensure that the context analysis is based on the reading order of the document to ensure the semantic coherence of the analysis results.

[0130] When the types of adjacent file regions are consistent (e.g., both are text regions), extract the key terms of the two regions and calculate the term matching degree: use the Jaro-Winkler distance algorithm to calculate the similarity of the terms of the two regions; the matching degree formula: Sim = 1 - (edit distance / max(len(term1), len(term2))); set the matching degree threshold: when Sim > 0.85, determine that it is the same professional term. Accurately identify professional terms across regions.

[0131] Based on the tax domain knowledge graph, analyze the association of multiple terms in the document and calculate the context consistency: C_context = (term association strength × 0.7) + (document logical coherence × 0.3); term association strength, query the relationship between terms (such as "small and micro profit enterprises" → "applicable subject") through the knowledge graph; document logical coherence is based on dependency syntax analysis to calculate the semantic coherence of the document. Ensure that the term identification conforms to the logic of tax policy.

[0132] Calculate the boundary confidence: C_boundary = 0.6 × term matching degree + 0.4 × context consistency. Set the boundary confidence threshold, high confidence, C_boundary > 0.85 → merge adjacent regions. Medium confidence: 0.75 ≤ C_boundary ≤ 0.85 → optimize the boundary. Low confidence: C_boundary < 0.75 → manual review. Consider term matching and context logic comprehensively to avoid misjudgment caused by a single indicator, realize dynamic optimization of boundary identification, and ensure the accuracy of professional term boundaries.

[0133] Boundary identification errors lead to incomplete association of tax elements. Through accurate boundary optimization, the completeness rate of tax element association is improved. Boundary identification is the basis for tax element extraction, and through accurate boundary optimization, the accuracy of key element (tax type, document number, applicable subject, preferential condition) identification is improved.

[0134] As a preferred embodiment of the present application, according to the obtained region type, data preprocessing is performed on the data in the file region, and single-modal features are extracted, including:

[0135] According to the text in the text region, text standardization is performed through a tax field term dictionary, tax content is recognized, and a word vector is obtained;

[0136] According to the order of the word vector in the text region, forward and backward splicing is performed to obtain a plurality of comprehensive vectors, wherein adjacent comprehensive vectors have overlapping regions;

[0137] According to the syntax of the comprehensive vector in the text in the text region, the dependency relationship between the word vectors in the comprehensive vector is recognized through a syntax tree.

[0138] The core purpose of the embodiment is to solve the problems of inaccurate professional term recognition and incomplete semantic dependency capture in tax policy documents, and to realize accurate semantic representation of tax content through a tax field self-adaptive text processing process, and to provide high-quality semantic features for tax element structured extraction.

[0139] The prior art only uses a general NLP model for text analysis when processing tax policy documents, which cannot accurately recognize professional terms and is difficult to capture long-distance semantic dependencies, resulting in errors in tax element extraction. The embodiment improves the accuracy of tax element recognition through field self-adaptive text processing and innovative comprehensive vector generation mechanism, and lays a high-quality semantic foundation for intelligent analysis of tax policy.

[0140] It can be understood that the tax field term dictionary is a field dictionary containing 2000+ tax professional terms. Through dictionary matching, tax professional terms in the text are recognized. Through field dictionary injection, the representation ability of professional terms is enhanced, and word vectors are generated. In addition, OCR recognition errors are repaired, non-tax content is removed, and irrelevant content is filtered through tax field rules.

[0141] Forward and backward splicing generates comprehensive vectors. Forward splicing sequentially splices word vectors from left to right to generate a sequence {V1+V2, V2+V3, V3+V4,..., Vn-1+Vn}; backward splicing sequentially splices word vectors from right to left to generate a sequence {Vn-1+Vn, Vn-2+Vn-1,..., V1+V2}; generate comprehensive vectors, combine the forward and backward splicing results to generate {V1+V2+V3, V2+V3+V4,..., Vn-2+Vn-1+Vn}. It can be understood that there is an overlapping part between adjacent comprehensive vectors to ensure semantic coherence and avoid long-distance dependency breakage.

[0142] The Eisner algorithm is used to build a syntax tree to solve the long-distance dependency problem. The dependency relationship between word vectors is identified through the syntax tree. In addition, the embodiment gives priority to extracting the dependency relationship between the word vectors in the comprehensive vector. It can be understood that the word vectors in the comprehensive vector are closer in content in the text and are more likely to have a corresponding dependency relationship.

[0143] As a preferred embodiment of the present embodiment, multi-modal fusion is performed, specifically:

[0144] According to the word vector, a node is constructed, and an edge relationship between nodes is constructed according to the dependency relationship between the word vectors;

[0145] According to the word vector, the monomodal feature of the formula region recognition is called to fill the formula into the edge relationship;

[0146] According to the word vector, data is filled with the monomodal feature of the table region recognition to obtain the knowledge graph.

[0147] The core purpose of the present embodiment is to solve the key technical problem of incomplete association of cross-modal elements in tax policy documents. Through the innovative node-edge relationship construction and multi-modal feature filling mechanism, the structured association of tax elements is realized.

[0148] The present embodiment constructs a high-quality tax knowledge graph by organically integrating word vectors, formula features, and table features, so that the completeness rate of tax element association is improved.

[0149] The node is constructed, and the tax elements (tax type, document number, applicable subject, and preferential condition) are extracted as the node based on the word vector as the basic unit. The node type is determined by the label sequence ({tax type, document number, subject, and condition}) output by the BiLSTM-CRF model. The node is constructed based on the professional labels in the tax field to ensure the semantic accuracy of the node and avoid semantic deviation in the general NLP model.

[0150] The edge relationship is constructed, and the dependency relationship between the word vectors is extracted based on the syntax tree analysis result. The edge relationship is constructed: source node→target node. The dependency relationship strength is calculated based on the path length and semantic similarity of the syntax tree: S_dep = 1 -(path length / maximum path length) × 0.8 + (semantic similarity × 0.2). The structured dependency relationship of the syntax tree is used to establish the logical association between elements, rather than simple semantic matching.

[0151] The formula filling is realized, and the monomodal feature (formula symbol tree, calculation logic) of the formula region recognition is called to fill the formula feature into the edge relationship. The formula feature is accurately associated with the element relationship to solve the problem of separation of dependency relationship and calculation formula.

[0152] Table data filling, calling the single-modal feature of table area recognition (table cell content, row table structure) ; filling the table data into the node, binding the specific data in the table with the element node, solving the problem that the table data cannot be associated with the element.

[0153] The embodiment realizes the accurate association of scattered elements in the tax policy document and the structured construction of the knowledge graph through the innovative node-edge relationship construction and multi-modal feature filling mechanism.

[0154] The application also provides a computer program product comprising a computer program which, when executed by a processor, implements the tax policy file processing method based on multi-modal analysis.

[0155] Therefore, any effect in the tax policy file processing method based on multi-modal analysis can be achieved, which will not be repeated here.

[0156] The application further provides a storage medium,

[0157] The storage medium has a computer program stored thereon, and the computer program, when executed, implements the steps of the tax policy file processing method based on multi-modal analysis.

[0158] Therefore, any effect in the tax policy file processing method based on multi-modal analysis can be achieved, which will not be repeated here.

[0159] The application further provides a processing device comprising:

[0160] A memory for storing a computer program;

[0161] A processor for executing the computer program to implement the steps of the tax policy file processing method based on multi-modal analysis.

[0162] Therefore, any effect in the tax policy file processing method based on multi-modal analysis can be achieved, which will not be repeated here.

[0163] The application further provides a processing device comprising:

[0164] Each embodiment in the specification is described in a progressive manner, and the same or similar parts between each embodiment can be referred to each other, and each embodiment mainly describes the difference from other embodiments.

[0165] The above merely illustrates the embodiments of the present application but should not be taken as limitations. Various changes and modifications can be made by those skilled in the art. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of claims of the present application.

Claims

1. A method for processing tax policy documents based on multimodal analysis, characterized in that, include: Based on the collected tax policy documents, the region types are identified by dividing the region into regions, and a confidence level corresponding to each region type is obtained. The region types include text regions, table regions, and formula regions. Based on the confidence levels sorted from highest to lowest, a first confidence level and a second confidence level are obtained sequentially. When the difference between the first confidence level and the second confidence level is greater than or equal to the confidence threshold, the file region is determined to be the region type corresponding to the first confidence level. Otherwise, the region type corresponding to the file region is adjusted and determined based on the type overlap rate of the two region types corresponding to the first confidence level and the second confidence level. The type overlap rate includes type overlap regions, a first type overlap rate occupying the type region corresponding to the first confidence level, and a second type overlap rate occupying the type region corresponding to the second confidence level. If both the first type overlap rate and the second type overlap rate are greater than an overlap threshold, the file region is determined to correspond to the two region types of the first confidence level and the second confidence level. If either the first type overlap rate or the second type overlap rate is greater than an overlap threshold, the file region is determined to correspond to the region type of either the first type overlap rate or the second type overlap rate. If both the first type overlap rate and the second type overlap rate are less than an overlap threshold, the first confidence level and the second confidence level are weighted and adjusted according to the first type overlap rate and the second type overlap rate to determine the region type corresponding to the file region. Based on the obtained region type, data preprocessing is performed on the data within the file region to extract unimodal features. Multimodal fusion is then performed using the unimodal features obtained from multiple file regions within the tax policy document to obtain a knowledge graph for parsing the tax policy document. Specifically, Based on the text in the text region, tax content is identified, and word vectors are obtained. These word vectors are then concatenated forward and backward according to their order in the text region to obtain multiple composite vectors. The grammar of these composite vectors within the text region is parsed to identify the dependencies between word vectors. Nodes are constructed based on the word vectors, and edge relationships are established between these nodes based on the dependencies between them. The single-modal features identified in the formula region are used to fill in the formulas for these edge relationships. Finally, data is filled in using the word vectors and the single-modal features identified in the table region to obtain the knowledge graph.

2. The tax policy document processing method based on multimodal parsing according to claim 1, characterized in that, Identify the region type, specifically: The tax policy document is divided into regions for each page using a modal feature detector. Among these methods, the detection of the number of consecutive text blocks is used to identify text regions. Identify table regions based on table boundaries and row / column structure. The formula area is identified based on the numerical symbols and expressions.

3. The tax policy document processing method based on multimodal parsing according to claim 1, characterized in that, After obtaining the region type, it also includes: The documents are arranged according to their order within the tax policy documents. When adjacent file regions correspond to the same region type, extract the terms from the two file regions to obtain the term matching degree; The association relationships between multiple terms in the file region are obtained, and the contextual consistency between the two file regions is obtained. Based on the term matching degree and the contextual consistency, the boundary confidence is obtained, and the boundary is adjusted based on the boundary confidence.

4. The tax policy document processing method based on multimodal parsing according to claim 1, characterized in that, Based on the obtained region type, data preprocessing is performed on the data within the file region to extract single-modal features, including: Based on the text in the text region, text standardization is performed using a tax terminology dictionary to identify tax-related content and obtain word vectors; Based on the order of word vectors in the text region, they are concatenated forward and backward to obtain multiple composite vectors, wherein adjacent composite vectors have overlapping regions; The syntax of the text in the text region is parsed based on the comprehensive vector, and the dependencies between word vectors within the comprehensive vector are identified through a syntax tree.

5. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the tax policy document processing method based on multimodal parsing as described in any one of claims 1 to 4.

6. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed, implements the steps of the tax policy document processing method based on multimodal parsing as described in any one of claims 1 to 4.

7. A processing apparatus, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the steps of the tax policy document processing method based on multimodal parsing as described in any one of claims 1 to 4 when executing the computer program.

Images