Tax policy document processing method based on multi-modal analysis

Abstract

The invention discloses a tax policy document processing method based on multi-modal analysis, and belongs to the technical field of tax policy document processing, and the method comprises the steps: recognizing the type of a region through region division according to a tax policy document, and obtaining a confidence coefficient; according to the descending order of the confidence coefficients, obtaining a first confidence coefficient and a second confidence coefficient, and when a difference value between the first confidence coefficient and the second confidence coefficient is greater than or equal to a confidence threshold value, determining that the file region is a region type corresponding to the first confidence coefficient; otherwise, adjusting and determining the region type corresponding to the file region according to the type coincidence rate of the two region types corresponding to the first confidence coefficient and the second confidence coefficient; and according to the obtained region type, performing data preprocessing on the data in the file region, extracting single-mode features, and performing multi-mode fusion through the single-mode features obtained by the plurality of file regions in the tax policy file to obtain a knowledge graph. And multi-modal fusion is executed through accurate division of regions, so that accurate analysis of the file is realized.

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: A tax policy document processing method based on multi-modal analysis, comprising: According to the collected tax policy documents, region division is performed 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; 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; 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 to be determined; 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.

[0010] The tax policy document processing method based on multi-modal analysis disclosed in the application has the following additional technical features: The region types are identified, specifically: Region division is performed on each page of the tax policy document by a modal feature detector, wherein the text regions are identified according to the number of consecutive character blocks, the table regions are identified according to the table boundaries and row and column structures, and the formula regions are identified according to the numerical characters and expressions.

[0011] The type coincidence rate of the two region types corresponding to the first and second confidence degrees is specifically: 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, and a second type coincidence rate of the type coincidence area accounting for the type region corresponding to the second confidence degree.

[0012] 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 to be determined, specifically: If the first type coincidence rate and the second type coincidence rate are both greater than the coincidence threshold, it is judged that the file region corresponds to two region types of the first confidence and the second confidence; If one of the first type coincidence rate and the second type coincidence rate is greater than the coincidence threshold, it is judged that the file region corresponds to the region type of one of the first type coincidence rate and the second type coincidence rate. If the first type coincidence rate and the second type coincidence rate are both less than the coincidence threshold, the first confidence and the second confidence are weighted and adjusted according to the first type coincidence rate and the second type coincidence rate, so as to determine the region type corresponding to the file region.

[0013] After obtaining the region type, the following steps are further included: According to the order of the file region in the tax policy file, the file region is arranged, When the region types corresponding to the adjacent file regions are consistent, the terms of the two file regions are extracted to obtain a term matching degree; The context consistency of the two file regions is obtained according to the correlation relationship of the plurality of terms corresponding to the file regions; 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.

[0014] According to the obtained region type, data preprocessing is performed on the data in the file region, and single-modal features are extracted, including: According to the text of the text region, text standardization is performed on the text of the text region through a tax field term dictionary, tax content is recognized, and a word vector is obtained; 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; According to the syntax of the comprehensive vector in the text of the text region, the dependency relationship between the word vectors in the comprehensive vector is recognized through a syntax tree.

[0015] Multi-modal fusion is performed, specifically: 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; According to the word vector, the single-modal features of the formula region recognition are called to fill the edge relationship with the formula; According to the word vector and the single-modal features of the table region recognition, data filling is performed to obtain the knowledge graph.

[0016] The application also provides a computer program product, The computer program is executed by a processor to implement the tax policy file processing method based on multi-modal analysis.

[0017] The application further provides a storage medium, 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.

[0018] The application further provides a processing device, comprising: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the tax policy file processing method based on multi-modal analysis.

[0019] Due to the above technical solutions, the application has the following beneficial effects: 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 a confidence threshold, it is determined that the file region is the region type corresponding to the first confidence degree.

[0020] When there is a significant gap between the two high confidence degrees, it indicates that there is a significant difference in the probability between the two identified region types, the credibility of the region type corresponding to the first confidence degree is increased, and the credibility of the region type corresponding to the second confidence degree is reduced, so that the file region is determined to be the region type corresponding to the first confidence degree, and the accuracy of region type identification is increased.

[0021] However, when the first confidence degree and the second confidence degree are both greater than or less than the fixed threshold (such as >0.8), there is a dilemma in distinguishing the region type. Moreover, due to slight interference, the confidence degree may fluctuate, leading to type misjudgment.

[0022] The application introduces a difference comparison mechanism, when there is a significant gap between the two high confidence degrees, the region type corresponding to the highest confidence degree is directly used, which solves the type misjudgment problem and the dilemma problem caused by slight interference, and improves the processing efficiency and accuracy in a simple scenario.

[0023] In addition, otherwise, according to the type coincidence rate of 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 makes a random choice. The present application introduces type coincidence rate analysis, accurately judges by calculating the overlap degree of two region types in space and semantics, solves the modality recognition problem in complex scenes, and improves the recognition accuracy.

[0024] 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 subjected to multi-modal fusion to obtain a knowledge graph for analyzing the tax policy file. Based on the accurate region type recognition 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 rate of tax elements. Especially in dealing with complex scenes 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

[0025] 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 serve to explain the present application, and do not constitute an improper limitation on the present application. In the drawings: Figure 1 The flowchart of the tax policy file processing method based on multi-modal analysis according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0026] In order to more clearly explain the overall concept of the present application, the following detailed description is made with reference to the accompanying drawings.

[0027] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0028] As shown in Figure 1 A tax policy file processing method based on multi-modal analysis, comprising: S100: According to the tax policy file obtained by acquisition, the region type is identified by region division, and the confidence corresponding to the region type is obtained, wherein the region type includes text region, table region and formula region.

[0029] The core purpose of this step is to solve the technical problem of inaccurate multi-modal region identification in tax policy documents, providing a high-quality data basis for subsequent multi-modal feature extraction and fusion.

[0030] Specifically, this step aims to accurately identify text, table, and formula regions in tax policy documents, solving the key problems of low tax element analysis accuracy and incomplete element association caused by incorrect region type identification in existing technologies.

[0031] An improved modal feature detector (a deep learning model based on CNN+ResNet) is used to detect multi-modal features of each page of the tax policy document to achieve region division. The region types include text, table, and formula regions. 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 achieve mutual deduction and completion of data corresponding to each keyword.

[0032] In addition, after region division, the modal feature detector can output the confidence of each region type, such as text region confidence 0.8, table region confidence 0, and formula region confidence 0.2. This allows the confidence to be used to determine the subsequent modal type.

[0033] It should be noted that the confidence output by the above modal feature detector can also be the initial confidence. For the text region, the OCR recognition accuracy and text continuity judgment are performed to obtain the text confidence, i.e., C_text = (OCR recognition accuracy × 0.6) + (text continuity score × 0.4).

[0034] In addition, for the table region, the table structure integrity and cell recognition accuracy are judged to obtain the table confidence, i.e., C_table = (table structure integrity score × 0.5) + (cell recognition accuracy × 0.5).

[0035] For the formula region, the mathematical symbol recognition accuracy and formula structure integrity are judged to obtain the formula confidence, i.e., C_formula = (mathematical symbol recognition accuracy × 0.7) + (formula structure integrity × 0.3).

[0036] The text confidence, table confidence, and formula confidence are used to adjust the initial confidence one by one to determine the final confidence.

[0037] Accurate region type recognition is the key prerequisite for implementing 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.

[0038] S200: According to the confidence from large to small, the first confidence and the second confidence are obtained in turn. 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. Otherwise, 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 adjusted.

[0039] 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, text and formula) is close, how to make accurate judgment.

[0040] 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 overlap rate analysis mechanism to realize dynamic adjustment of region type recognition, ensuring high-precision region type determination in the case of fuzzy boundaries.

[0041] The three region type confidences calculated are sorted from high to low to obtain the first confidence and the second confidence. The difference between the first confidence and the second confidence is calculated as ΔC = |C_first - C_second|.

[0042] Set a confidence threshold, for example 0.15. When ΔC ≥ 0.15, it is determined that the file region is the region type corresponding to the first confidence. When ΔC < 0.15, enter the overlapping region processing flow.

[0043] The prior art cannot make accurate judgments when the confidence is close, resulting in misjudgment of region type. This step uses a difference comparison mechanism to avoid the dilemma, improving the accuracy of region type determination when the confidence is close.

[0044] It should be noted that overlapping regions are 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% is defined.

[0045] Through the type coincidence rate analysis mechanism, the system can accurately identify the true type of the overlapping area, and the accuracy of the overlapping area recognition is improved.

[0046] 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 x (1 - η)). Where η is the average accuracy of the current mode in historical processing. Specifically, when the historical accuracy η > 0.85, the confidence threshold is automatically reduced to 0.10, improving the sensitivity.

[0047] Fixed threshold determination is sensitive to small disturbances, resulting in a high rate of false positives. For example, the "annual taxable income ≤ 300 million" in the table is misjudged as a text area, affecting the accurate extraction of preferential conditions. The dynamic adjustment mechanism of this step reduces the false positive rate and significantly improves the system robustness.

[0048] This step solves the core problem of region type recognition in tax policy documents through innovative confidence difference comparison and type coincidence rate analysis mechanism, which is the key link to realize the 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 structuring processing, making the overall precision of tax policy document analysis improve qualitatively.

[0049] S300: According to the obtained region type, the data in the file region is preprocessed, the single modal feature is extracted, and the single modal feature obtained from the multiple file regions in the tax policy file is fused to obtain the knowledge graph of the tax policy file.

[0050] The core purpose of this step is to solve the problem of incomplete correlation of multi-modal elements in tax policy documents. Through accurate single modal feature extraction and innovative multi-modal fusion mechanism, the intelligent correlation of key elements scattered in different modalities such as text, table, formula, etc. in tax policy documents is realized, and a high-quality tax policy knowledge graph is constructed.

[0051] The existing technology can only analyze a single modality, lacks cross-modal element mapping capability, and the key elements cannot be accurately correlated. This step realizes the accurate correlation of tax elements and the structured construction of knowledge graph through a systematic multi-modal fusion process, laying a technical foundation for the intelligent application of tax preferential policies.

[0052] The text area, table area, and formula area are preprocessed respectively to extract the corresponding text features, table features, and formula features.

[0053] Among them, for table area preprocessing, the table topology repair algorithm (based on GCN) is applied to correct the table structure error of OCR recognition; through table cell content analysis, the key elements in the table are extracted; the topological relationship of table rows, columns and cells is established.

[0054] For formula area preprocessing, the formula is converted into a symbol tree structure, the variables in the formula are replaced with placeholders through dynamic parameterization technology, and the key calculation elements in the formula are extracted.

[0055] In order to realize the alignment of three-dimensional features, a cross-modal attention mechanism is constructed to realize the alignment of text-table-formula three-dimensional features; through comparative learning, explicit alignment between modalities is realized to solve the problem of semantic difference between modalities. The text feature vector, table feature vector and formula feature vector are fused into a tensor, and a cross-modal fusion layer (based on the multi-head attention mechanism of Transformer) is applied to generate a fusion feature vector, and the cross-modal correlation between elements is realized through the fusion feature vector.

[0056] It should be noted that the fusion quality is evaluated, and the diversity index (based on the entropy value of the feature vector) of the fusion feature is calculated; if the entropy value is <0.3, the feature enhancement mechanism (multi-model integration, domain knowledge injection, etc.) is started to ensure the high quality and integrity of the fusion feature.

[0057] Based on GNN, a dynamic knowledge graph of tax elements is constructed, the nodes represent tax elements, and the edges represent the correlation between elements, supporting multi-hop reasoning and realizing deep mining of complex element relationships.

[0058] The prior art can only analyze a single modality when processing tax policy documents, and cannot establish cross-modal element correlation, which may lead to conditional recognition errors. The multi-modal fusion mechanism of this step improves the completeness rate of tax element correlation.

[0059] The key elements in the tax policy document are scattered in different modalities, and accurate single-modal feature extraction and fusion are needed to accurately identify them. Through single-modal feature extraction and multi-modal fusion, the accuracy of key element identification is improved in this step.

[0060] Through the innovative single-modal feature extraction and multi-modal fusion mechanism, this step realizes the accurate correlation of scattered elements in the tax policy document and the structured construction of the knowledge graph, which is the core link of intelligent analysis of tax policy documents. Not only does it solve the problem of incomplete multi-modal element correlation in the prior art, but it also provides a high-quality data basis for the 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.

[0061] As a preferred embodiment of the present application, the type of the identified area is recognized, specifically: Through the modal feature detector, region division is performed on each page of the tax policy file, wherein, according to the number of continuous character blocks, a text region is identified, according to the table boundary and row and column structure, a table region is identified, and according to the number and expression, a formula region is identified.

[0062] The core purpose of the embodiment is to solve the precision problem of multi-modal region identification of tax policy documents, and to provide a high-quality data basis for subsequent single-modal feature extraction and multi-modal fusion.

[0063] The tax policy document contains a large amount of professional text, complex table and mathematical formula, and the traditional OCR technology cannot effectively distinguish these region types, resulting in subsequent element analysis errors. The embodiment realizes the precise separation of text, table and formula regions in the tax policy document through the targeted modal feature detection mechanism, improves the region identification accuracy, and lays a reliable foundation for tax element structured processing.

[0064] Among them, for text region identification, a character block continuous number detection algorithm is used to scan the page horizontally and count the number of characters in continuous character blocks; when the length of the continuous character block is > 50 characters, it is determined as a text region; non-text content is filtered through a tax field dictionary. Solve the misjudgment problem of tax policy number dense area, avoid misjudging the number as a table.

[0065] For table region identification, table boundary detection is used to identify horizontal and vertical lines in the page through Hough transform; row and column structure analysis is used to calculate the number of rows and columns based on connected component analysis, and a table topology graph is constructed to determine the table region. Accurately identify complex table structure, solve the row and column misplacement problem caused by fuzzy table lines in traditional methods.

[0066] For formula region identification, number symbol recognition is used to match based on a mathematical symbol library (containing "≤", "×", "IF", etc. 500+ tax formula symbols); expression structure analysis is used to construct a formula symbol tree, analyze the logical structure of the expression, and determine the formula region. Solve the tax formula recognition problem, avoid misjudging the formula as ordinary text.

[0067] Region identification errors can cause element association failure. The embodiment realizes accurate region identification through tax field special feature detection, improves the completeness rate of tax element association, and improves the accuracy of subsequent single-modal feature extraction.

[0068] The embodiment accurately solves the separation problem of text, table and formula regions through the 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.

[0069] 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 specifically: The type coincidence rate includes a type coincidence region, a first type coincidence rate of the type coincidence region in the first confidence corresponding type region, and a second type coincidence rate of the type coincidence region in the second confidence corresponding type region.

[0070] The core purpose of the embodiment is to solve the ambiguity problem of region type identification in tax policy documents, especially when the confidence of two region types is close, how to accurately judge the real type of the region.

[0071] The tax policy document often has overlapping scenes such as "formula embedded table" (such as "reduction amount = IF (revenue < 5 million, revenue x 5%, 250,000 + excess part x 7%)") and "text containing formula" (such as "annual taxable income ≤ 300,000 part reduces by 25% into taxable amount"). The traditional method cannot make accurate judgments only relying on the confidence difference, resulting in misjudgment of region type. The embodiment realizes accurate judgment of overlapping regions through the type coincidence rate calculation mechanism, which significantly improves the accuracy of multi-modal analysis of tax policy documents.

[0072] The number of features of the first confidence corresponding region type is identified, the number of features of the second confidence corresponding region type is identified, and then the number of features belonging to both region types, i.e. the type coincidence region, is identified.

[0073] The first type coincidence rate R_first = (type coincidence feature number / first confidence corresponding type region feature number) x 100%. The second type coincidence rate R_second = (type coincidence feature number / second confidence corresponding type region feature number) x 100%.

[0074] Specifically, 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 and determined, specifically: If the first type coincidence rate and the second type coincidence rate are both greater than the coincidence threshold, it is judged that the file region corresponds to the two region types of the first confidence and the second confidence; If one of the first type coincidence rate and the second type coincidence rate is greater than the coincidence threshold, it is judged that the file region corresponds to the region type of one of the first type coincidence rate and the second type coincidence rate; If the first type coincidence rate and the second type coincidence rate are both less than a coincidence threshold, the first confidence and the second confidence are adjusted according to the first type coincidence rate and the second type coincidence rate to determine the region type corresponding to the file region.

[0075] It should be noted that the coincidence threshold is set to 70% (through statistical analysis, when the coincidence rate is greater than 70%, the region type is dominant).

[0076] Both double coincidence rates are greater than the threshold (R_first> 70% and R_second> 70%), which is identified as a double-mode region (such as a table and formula overlap region). At this time, for this file region, it is necessary to pre-process and extract features through the method corresponding to the double-mode region to avoid missing features.

[0077] Single coincidence rate> threshold scenario (R_first> 70% and R_second≤ 70% or R_first≤ 70% and R_second> 70%), select the region type with coincidence rate> threshold as the final determination. For example, R_first=78% (table), R_second=65% (formula), and the table region is determined. Avoid misjudgment of the region to improve the accuracy of identifying key calculation elements.

[0078] Both double coincidence rates are less than the threshold (R_first≤ 70% and R_second≤ 70%), calculate the weighted confidence new_conf_first=(R_first×conf_first) / (R_first+R_second); calculate the weighted confidence new_conf_second=(R_second×conf_second) / (R_first+R_second). Compare new_conf_first and new_conf_second, and select the region type corresponding to the larger value. Avoid simple averaging to ensure that regions with high coincidence rates have higher weights.

[0079] The present 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.

[0080] As a preferred embodiment of the present application, after obtaining the region type, it further includes: arranging according to the order of the file region in the tax policy file, when the region types corresponding to the adjacent file regions are consistent, extracting the terms of the two file regions to obtain the term matching degree; obtain a plurality of terms corresponding to the association relationship of the file area, and obtain the context consistency of the two file areas; 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.

[0081] The core purpose of the embodiment is to solve the problem of inaccurate identification of text area boundaries in tax policy documents, especially the problem of boundary identification error caused by tax professional terms crossing multiple text blocks.

[0082] The traditional method only relies on the OCR recognition result for boundary division, and cannot accurately identify the boundaries of tax professional terms, leading to errors in subsequent element extraction. The embodiment introduces term matching degree and context consistency analysis to realize dynamic optimization of area boundaries, improve the accuracy of tax professional term boundary identification, and provide high-quality input basis for tax element structured processing.

[0083] Linearly sort the file areas in the tax policy document according to their physical positions (from top to bottom, from 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, and ensure the semantic coherence of the analysis results.

[0084] When the types of adjacent file areas are consistent (such as both being text areas), extract the key terms of the two areas and calculate the term matching degree: use the Jaro-Winkler distance algorithm to calculate the similarity of the terms of the two areas; the matching degree formula: Sim = 1 - (edit distance / max(len(term1), len(term2))); set the matching degree threshold: when Sim>0.85, it is determined that it is the same professional term. Accurately identify professional terms across regions.

[0085] Based on the tax domain knowledge graph, analyze the association relationship 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.

[0086] Compute boundary confidence: C_boundary = 0.6 x term matching degree + 0.4 x 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 to avoid false positives caused by a single indicator, achieve dynamic optimization of boundary recognition, and ensure accurate professional term boundaries.

[0087] Boundary recognition errors lead to incomplete tax element association. Through accurate boundary optimization, the completeness rate of tax element association is improved. Boundary recognition 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) recognition is improved.

[0088] 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: According to the text of the text region, text standardization is performed through a tax field term dictionary, tax content is identified, and a word vector is obtained; 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; According to the syntax of the comprehensive vector in the text of the text region, the dependency relationship between the word vectors in the comprehensive vector is identified through a syntax tree.

[0089] The core purpose of the present embodiment is to solve the problem of inaccurate professional term recognition and incomplete semantic dependency capture in tax policy documents. Through a tax field adaptive text processing process, accurate semantic representation of tax content is achieved, and high-quality semantic features are provided for tax element structured extraction.

[0090] The prior art only uses general NLP models for text analysis when processing tax policy documents, which cannot accurately identify professional terms and is difficult to capture long-distance semantic dependencies, resulting in tax element extraction errors. The present embodiment improves the accuracy of tax element recognition through field adaptive text processing and innovative comprehensive vector generation mechanism, and lays a high-quality semantic foundation for intelligent analysis of tax policy.

[0091] 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 identified. 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.

[0092] Forward and backward splicing to generate comprehensive vectors. Forward splicing: sequentially splice word vectors from left to right to generate a sequence {V1+V2, V2+V3, V3+V4,..., Vn-1+Vn}; backward splicing: sequentially splice word vectors from right to left to generate a sequence {Vn-1+Vn, Vn-2+Vn-1,..., V1+V2}; generate comprehensive vectors, merge 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.

[0093] Based on the Eisner algorithm, a syntax tree is constructed to solve the long-distance dependency problem. Through the syntax tree, the dependency relationship between word vectors is identified. Moreover, the embodiment preferentially extracts the dependency relationship between word vectors within the comprehensive vector. It can be understood that the word vectors within the comprehensive vector are closer in content in the text and are more likely to have corresponding dependency relationships.

[0094] As a preferred embodiment under the present embodiment, multi-modal fusion is performed, specifically: 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; According to the word vector, a single-modal feature of formula region recognition is called to fill the formula in the edge relationship; According to the word vector, data filling is performed with the single-modal feature of table region recognition to obtain the knowledge graph.

[0095] 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.

[0096] The present embodiment organically integrates word vectors, formula features, and table features to construct a high-quality tax knowledge graph, thereby improving the completeness rate of tax element association.

[0097] The node is constructed, and a tax element (tax type, document number, applicable subject, and preferential condition) is extracted as a basic unit of a word vector; and a label sequence ({tax type, document number, subject, and condition}) output by a BiLSTM-CRF model is used to determine the node type. The node is constructed based on a professional tax field label, so that the semantic accuracy of the node is ensured, and semantic deviation in a general NLP model is avoided.

[0098] The edge relationship is constructed, the dependency relationship between the word vectors is extracted based on the parsing result of the syntax tree, and the edge relationship is constructed: source node→target node. The dependency relationship strength is calculated based on the path length of the syntax tree and the semantic similarity: S_dep = 1-(path length / maximum path length) x 0.8 + (semantic similarity x 0.2). The structured dependency relationship of the syntax tree is used to establish the logical association between the elements, instead of simple semantic matching.

[0099] The formula filling is realized, the monomodal features (formula symbol tree and calculation logic) of the formula region recognition are called, and the formula features are filled into the edge relationship. The formula features are accurately associated with the element relationship, and the problem that the dependency relationship and the calculation formula are separated is solved.

[0100] The table data filling is realized, the monomodal features (table cell content and row table structure) of the table region recognition are called, and the table data is filled into the node. The specific data in the table is bound to the element node, and the problem that the table data cannot be associated with the element is solved.

[0101] The embodiment realizes the accurate association of the scattered elements in the tax policy document and the structured construction of the knowledge graph by using the innovative node-edge relationship construction and the multi-modal feature filling mechanism.

[0102] The application also provides a computer program product, which comprises a computer program.

[0103] Therefore, any effect in the multi-modal analysis-based tax policy document processing method can be realized, and details are not described herein.

[0104] The application further provides a storage medium, The storage medium stores a computer program, and the computer program is executed to realize the steps of the multi-modal analysis-based tax policy document processing method.

[0105] Therefore, any effect in the multi-modal analysis-based tax policy document processing method can be realized, and details are not described herein.

[0106] The application further provides a processing device, which comprises: a memory for storing a computer program; a processor for implementing the steps of the tax policy document processing method based on multi-modal resolution when executing the computer program.

[0107] Therefore, any effect in the tax policy document processing method based on multi-modal resolution can be achieved, which is not described here.

[0108] The places not mentioned in the present application can be realized by using or referring to the existing technology.

[0109] 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.

[0110] The above only describes the embodiments of the present application and is not used to limit the present application. For those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application shall be included in the scope of claims of the present application.

Claims

1. A tax policy document processing method based on multimodal analysis, characterized in that: include: Based on the collected tax policy documents, regional division is performed to identify regional types and obtain confidence scores corresponding to the regional types, wherein the regional types include text areas, table areas, and formula areas; Sorting the confidence levels in descending order to obtain a first confidence level and a second confidence level, and determining that the file region is a region type corresponding to the first confidence level when a difference between the first confidence level and the second confidence level is greater than or equal to a confidence threshold; Otherwise, adjusting and determining the region type corresponding to the file region according to the type overlap ratio of the two region types corresponding to the first confidence level and the second confidence level; According to the obtained area type, data preprocessing is performed on the data in the file area to extract unimodal features. The unimodal features obtained from multiple file areas in the tax policy file are then multimodally fused to obtain a knowledge graph for parsing the tax policy file.

2. The tax policy document processing method based on multimodal analysis according to claim 1 is characterized in that: Identify the region type, specifically: By using a modal feature detector, each page of the tax policy document is divided into regions. Among them, the text area is identified based on the continuous number of text blocks. Identify table areas based on table boundaries and row and column structures. Identify formula areas based on numeric characters and expressions.

3. The tax policy document processing method based on multimodal analysis according to claim 1 is characterized in that: The type overlap rate of the two area types corresponding to the first confidence level and the second confidence level is specifically: The type overlap rate, including the type overlap area, accounts for the first type overlap rate of the type area corresponding to the first confidence level, And the type overlap area accounts for the second type overlap rate of the type area corresponding to the second confidence level.

4. The method for processing tax policy documents based on multimodal analysis according to claim 3 is characterized in that: 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, specifically: If the first type overlap rate and the second type overlap rate are both greater than the overlap threshold, determining the file area as two area types corresponding to the first confidence level and the second confidence level; If one of the first type overlap rate and the second type overlap rate is greater than an overlap threshold, determining that the file area corresponds to the area type of the first type overlap rate and the second type overlap rate; If the first type overlap rate and the second type overlap rate are both less than the overlap threshold, the first confidence level and the second confidence level are weightedly adjusted according to the first type overlap rate and the second type overlap rate to determine the area type corresponding to the file area.

5. The tax policy document processing method based on multimodal analysis according to claim 1 is characterized in that: After getting the region type, it also includes: Arrange the file areas according to the order in which they appear in the tax policy file. When the region types corresponding to the adjacent file regions are consistent, extracting terms from the two file regions to obtain a term matching degree; Obtaining association relationships between multiple terms in the document regions, and obtaining context consistency between the two document regions; A boundary confidence is obtained according to the term matching degree and the context consistency, and the boundary is adjusted according to the boundary confidence.

6. The tax policy document processing method based on multimodal analysis according to claim 1 is characterized in that: According to the obtained region type, data preprocessing is performed on the data in the file region to extract single-modal features, including: Based on the text in the text area, the tax domain term dictionary is used to perform text standardization, identify tax content, and obtain word vectors; According to the order of the word vectors in the text area, forward and backward splicing is performed to obtain multiple comprehensive vectors, wherein adjacent comprehensive vectors have overlapping areas; The grammar of the text in the text region of the comprehensive vector is parsed, and the dependency relationship between word vectors in the comprehensive vector is identified through a grammar tree.

7. The method for processing tax policy documents based on multimodal analysis according to claim 6 is characterized in that: Perform multimodal fusion, specifically: Constructing nodes based on the word vectors, and constructing edge relationships between nodes based on the dependency relationships between the word vectors; Based on the word vector, calling the unimodal features identified in the formula area to fill in the formula of the edge relationship; Data is filled based on the word vector and the unimodal features identified in the table area to obtain the knowledge graph.

8. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, it implements the tax policy document processing method based on multimodal analysis as described in any one of claims 1 to 7.

9. 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 analysis as described in any one of claims 1 to 7.

10. A processing device, characterized in that: include: Memory for storing computer programs; A processor is used to implement the steps of the tax policy document processing method based on multimodal analysis as described in any one of claims 1 to 7 when executing the computer program.

Images