Tax policy semantic analysis and automatic adaptation engine for scientific research business

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a scientific research business-oriented tax policy semantic analysis and automatic adaptation engine. The engine executes the following steps: step 1, performing standardized analysis on a tax policy text to form policy clause structure representation taking clause nodes as a center; 2, aligning the policy term terms with the scientific research business terms to generate term mapping indexes; step 3, refining clause obligations into an obligational slot set, converting business events into an evidence unit set, and executing a stable wedding process according to a bidirectional preference sequence to complete pairing; and step 4, generating an action list and a process sticking point rule according to a pairing result, and solidifying to generate a clause-evidence-action link containing an embedded box body and a hyperbolic embedded coordinate snapshot. According to the invention, policy term automatic identification, business activity automatic matching and whole-course traceability of the compliance process can be realized, and the applicable accuracy and execution transparency of the scientific research project tax preferential policy are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a semantic parsing and automatic adaptation engine for tax policies for scientific research. Background Technology

[0002] Matching financial accounting for scientific research activities with tax policies has become one of the core issues facing research institutions and technology-based enterprises. The scope of research funding use, pre-tax deduction items, VAT deduction conditions, and the applicability of the R&D expense super-deduction policy all require matching the specific semantics of policy clauses with the actual processes of enterprise research operations. Traditional policy adaptation methods mainly rely on manual interpretation and static rule coding, lacking automatic semantic understanding capabilities, resulting in low policy application efficiency, high error rates, and difficulty in coping with the complexity and frequent updates of policy clauses.

[0003] Currently, existing technical approaches mainly fall into two categories: one is policy parsing systems based on keyword matching and rule engines, and the other is policy text understanding systems based on natural language processing models. The former achieves matching of clauses with business operations by constructing policy dictionaries, rule templates, and condition trees. For example, it uses regular expression matching to identify keywords such as "additional deduction," "R&D project," and "technology contract" and performs logical judgments. This method is simple and easy to implement in the initial deployment, but as policies become more detailed and cross-domain, the number of rules grows exponentially, maintenance costs rise sharply, and it struggles to identify clauses with similar semantics but different expressions. The latter utilizes machine learning or deep learning models for policy semantic understanding, such as using pre-trained language models to identify clause obligations, applicable conditions, and tax rate parameters. This type of method can handle semantic diversity to some extent, but due to the lack of a unified terminology semantic space and domain logical constraints, its accuracy in cross-clause semantic alignment and structured mapping remains unsatisfactory. Summary of the Invention

[0004] The main objective of this invention is to provide a tax policy semantic parsing and automatic adaptation engine for scientific research operations. This engine performs the following steps: Step 1: Standardize and parse the tax policy text to form a policy clause structure centered on clause nodes; Step 2: Align policy terms with scientific research business terms to generate a term mapping index. The alignment process includes: constructing embedding boxes to represent the scope of application of terms, constructing hyperbolic embedding coordinates to represent the hierarchical relationship of terms, performing intersection retrieval on the embedding boxes and nearest neighbor filtering on the hyperbolic embedding coordinates to jointly confirm mapping pairs, and repairing semantic differences across contexts through Sheaf consistency diagnosis. Step 3: Break down the terms and obligations into a set of obligation slots, transform business events into a set of evidence units, and complete the pairing by executing a stable matching process based on the order of mutual preferences. Step 4: Generate an action list and process checkpoint rules based on the pairing results, and solidify the generated terms-evidence-action chain that includes the embedded box and hyperbolic embedded coordinate snapshots.

[0005] Furthermore, the standardization and parsing steps for policy clauses specifically include: establishing a domain terminology and a list of node types; based on sentence delimitation and domain template identification, extracting the three-part structure of object, condition, and obligation, while also extracting exceptions, time, and scope, and marking cross-sentence references within the same clause to generate a set of clause fragments; mapping the set of clause fragments into a policy clause structure representation centered on clause nodes and connecting object, condition, obligation, exception, time, and scope nodes.

[0006] Furthermore, the construction of the embedded box specifically includes: performing discrete value statistics along a fixed axis system composed of object category, activity stage, expense category, voucher type, regional identifier, time limit marker, range marker and reference relationship, and determining the coverage start point and coverage end point on each axis to form the embedded box; for axes with exception markers, the coverage set of the axis is differentially processed using the exception markers.

[0007] Furthermore, the construction of hyperbolic embedding coordinates specifically includes: generating a term hierarchy tree and a relational adjacency table based on the domain hierarchy and the hierarchical relationship of terms; assigning hyperbolic embedding coordinates to each term with the level depth as the radial index and the same-level lexicographical order as the angular index.

[0008] Furthermore, the intersection retrieval and nearest neighbor selection specifically include: calculating the intersection state of each axis for each pair of terms and generating three types of intersection judgment labels: fully compatible, partially compatible, and incompatible; for candidate pairs that are fully compatible and partially compatible, searching the hyperbolic nearest neighbor table, and if the reference set has items with the same superordinate terms on both sides, then the candidate pair is marked as structural adjacency matching; the intersection judgment label and the structural adjacency matching label are combined as the basis for alignment confirmation.

[0009] Furthermore, the Sheaf consistency diagnosis specifically includes: establishing a base space graph composed of clause nodes, reference nodes, and version nodes; establishing local fibers containing embedded boxes and hyperbolic embedded coordinates on the coverage sets such as clause contexts and interpretive attachments; defining constraint mappings for semantic equivalence merging and attribute projection; pushing fiber content along each loop of the base space graph according to the constraint mappings and comparing them to detect and record consistency differences; and repairing consistency difference records sequentially through semantic equivalence merging, embedded box expansion, or segmentation.

[0010] Furthermore, refining the obligation slot set and transforming the evidence unit set specifically includes: traversing the clause nodes in the clause list, refining the obligation nodes, and generating an obligation slot set containing information on the required set of voucher types and the set of allowed fee categories; accessing event records, image records, and text records from the business event flow and evidence warehouse, transforming business-side terms into policy-side terms based on the terminology mapping index, and binding records associated with the same event to form an evidence unit set.

[0011] Furthermore, the process of executing stable matchmaking specifically includes: establishing an evidence preference order for obligation slots and an obligation preference order for evidence units; adopting a delayed acceptance process, whereby unmatched obligation slots initiate proposals according to their own preference order until no new proposals are made, thus forming a collection of matchmaking lists.

[0012] Furthermore, the specific rules for generating action lists and process checkpoints include: summarizing the paired items in the collection and matching list to generate a list of required data, and generating a gap list for unmatched slots; and generating three types of process checkpoint rules for each collection batch based on the condition nodes and scope nodes in the clause list, namely, pre-submission verification, inter-node transfer verification, and pre-archiving verification.

[0013] Furthermore, the solidified clause-evidence-action link specifically includes: generating a link record for each pairing item containing a clause identifier, evidence number, and action list item number; attaching an embedding box and hyperbolic embedding coordinate snapshot to each link record; and writing the link record to audit storage.

[0014] This invention presents a tax policy semantic parsing and automatic adaptation engine for scientific research operations, offering the following advantages: Firstly, it achieves end-to-end fusion of semantic, alignment, and execution layers in the field of intelligent tax policy parsing and automatic adaptation for scientific research operations. This enables the establishment of a traceable logical mapping between complex policy clauses and actual R&D activities in a structured manner. Secondly, by employing a terminology alignment method combining embedded boxes and hyperbolic embedded coordinates, this invention not only establishes stable correspondences between synonyms, hierarchical items, and cross-document expressions but also maintains consistency during policy updates or terminology system expansions, eliminating the reliance on manual rule maintenance for policy application judgments. Thirdly, combined with the Sheaf consistency diagnostic mechanism, this invention ensures global consistency in compliance terminology across documents and contexts, thereby eliminating the ambiguity accumulation problem caused by context drift in traditional semantic matching. Finally, by refining clause obligations into a set of obligation slots and adopting a stable matching aggregation strategy, this invention achieves a one-to-one mapping between evidentiary materials and clause obligations, making the aggregation results deterministic and interpretable, avoiding the randomness and opacity caused by relying on weighted models. At the execution layer, this invention, through the coordinated generation of action lists and process checkpoint rules, enables policy requirements to be embedded into the entire scientific research process as verifiable execution nodes, forming a complete closed loop from clauses and evidence to actions. In particular, the embedded boxes and hyperbolic embedded coordinate snapshots fixed in the clause-evidence-action chain provide time-reproducible and structurally replayable evidence for tax audits and policy tracing. Overall, this invention surpasses existing technologies in semantic accuracy, execution traceability, and system adaptability. It achieves unified and automated processing of intelligent tax policy identification, meaning matching, evidence collection, and compliance verification, significantly reducing manual comparison costs and improving the transparency and consistency of policy implementation. It is applicable to various scenarios involving additional deductions for scientific research projects and the collection of R&D expenses. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of an eight-dimensional embedded box structure provided in an embodiment of the present invention; Figure 2 A schematic diagram of the semantic space distribution of the three-part clause structure (object-condition-obligation) provided for embodiments of the present invention; Figure 3 A schematic diagram showing the accuracy comparison curves of different terminology alignment methods provided in the embodiments of the present invention. Detailed Implementation

[0016] The method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0017] A semantic parsing and automatic adaptation engine for tax policies geared towards scientific research operations. This engine performs the following steps: Step 1: Standardize and parse the tax policy text to form a policy clause structure centered on clause nodes.

[0018] Specifically, policy texts are sourced from three channels: official gazette pages, downloadable versions of normative documents, and plain text files exported from authorized policy compilation systems. Each channel registers its source identifier and publication date before entering the process to ensure the latest version is maintained in chronological order when duplicate texts appear. All inputs are standardized with a unified encoding and consistent use of full-width and half-width punctuation; for example, Chinese commas and English semicolons are not mixed, nor are Chinese book titles and English quotation marks. This initial step is crucial because subsequent sentence delimitation relies on punctuation stability. Mixed punctuation can cause sentence delimitation to drift, leading to incorrect segmentation of the object, condition, and obligation triad. Most policy texts contain hierarchical headings, such as Chapter 1, Section 2, Article 3, Paragraph 1, and Item 1. These hierarchical markers are identified and numbered in descending order to form a chapter hierarchy tree. A top-down scanning strategy is employed: upon encountering a section-level marker, a section-level node is immediately created; upon encountering a paragraph-level marker, a paragraph-level node is attached to the current section-level node, and so on. The advantage of this approach is that it preserves the boundaries of the internal hierarchy, providing a clear reference point for subsequent cross-sentence referencing, such as the preceding clause or item. Remove duplicate organization names and publication dates from headers, footers, page numbers, and watermarks. To avoid accidental deletion of text, the cleanup rules are limited to fixed phrases that appear consecutively multiple times at the beginning or end of a line, with a length not exceeding 40 characters. After cleanup, re-verify the integrity of the chapter hierarchy tree. If cleanup results in missing levels, roll back the cleanup action and mark the paragraph as requiring manual review.

[0019] The initial segmentation is performed based on periods, semicolons, colons, and line breaks. Lines containing enumeration structures are then segmented a second time by item-level markers to ensure each enumeration item forms an independent sentence. Enumeration items are separated because obligations are typically developed through enumeration; keeping them within the same sentence would broaden the scope of subsequent obligation location. Each sentence is identified into three categories: normative sentences, defining sentences, and explanatory sentences. Normative sentences have explicit behavioral requirements or restrictions and often contain verbs or adverbs such as "should," "need," "may," and "must not." Defining sentences define noun concepts and often contain structures such as "refers to," "includes," and "belongs to." Explanatory sentences explain background or provide examples. This classification aims to ensure that the three-part structure extraction process is limited to normative sentences, thereby reducing false positives.

[0020] Establish a terminology list covering object categories, activity stages, expense categories, voucher types, geographical identifiers, time-limited markers, scope markers, exception markers, and reference markers. Examples of object categories include R&D projects, R&D activity entities, and external commissioning units; examples of expense categories include material costs, testing fees, and labor costs; examples of voucher types include invoices, contracts, test reports, and experimental records; examples of geographical identifiers include national, provincial, and municipal; examples of time-limited markers include annual, quarterly, and fiscal year; examples of scope markers include limited to, not exceeding, and not less than; examples of exception markers include but, excluding, and not including; and examples of reference markers include in accordance with this method, this article, and this clause. The terminology list is imported once at the start of the project and allows the addition of equivalent terms and aliases. All equivalence relationships are explicitly registered in the list to avoid ambiguity in subsequent stages. Set up three sets of trigger word tables: object guiding terms, condition trigger terms, and obligatory verbs. Examples of object-introducing words include "to," "targeting," "applicable to," and "oriented towards"; examples of conditional triggering words include "if," "when," "in," "compliant with," and "satisfied"; examples of obligatory verbs include "should," "need," "should be," "must," "provide," "collect," "calculate," "verify," and "fill in." To address the situation where multiple triggering words coexist in the same sentence, a fixed recognition order is set for the three groups of triggering words: first, identify obligatory verbs, then identify conditional triggering words, and finally identify object-introducing words. This order is based on the fact that the core of a standard sentence lies in behavioral requirements. Once the obligation is identified, the object and condition can be collected to the left and right using the obligation as an anchor point, reducing misjudgments in scope.

[0021] The system scans from left to right within the standard sentence. Upon encountering an obligatory verb, an obligatory node is immediately created, recording the verb text and its start and end position indices. If a sentence contains multiple obligatory verbs, each verb forms a separate obligatory node, and the sentence is segmented into multiple fragments based on the verb's position, allowing a single sentence to generate multiple clause fragments. This approach is necessary because multiple obligations often point to different lists of data or different actions; forcibly merging them would lead to inaccurate subsequent action generation. The system backtracks from the obligatory node to the left until it encounters the beginning of a sentence, a semicolon, or another obligatory node. During this backtracking process, conditional trigger words are identified, and the content following the trigger word up to the obligatory node is collected as the condition range. To avoid mistakenly collecting background information as conditions, two filtering rules are set: first, if a condition paragraph begins with words such as "for example," "for example but not limited to," or "example," it is marked as a residual explanatory sentence and removed; second, if a condition paragraph does not contain any item from the terminology list, it is considered a generalized description and handled by the optional implementation method. When multiple conditional trigger words appear within the condition range, clauses are not split; they are recorded as a single condition node for later refinement. Looking forward from the obligation node to the right until a semicolon, period, or enumeration terminator is encountered. During this forward look, object-introducing words are identified, and the object phrases following these words are collected as object ranges. If there is no object-introducing word in the sentence, object nouns are searched between the beginning of the sentence and the obligation node, prioritizing the noun phrase or group of noun phrases closest to the obligation node, and validating them against the object category in the terminology list. Selecting object phrases close to the obligation node as associated objects maximizes the consistency between predicate and agent, avoiding misattachment of distant nouns to obligations. An obligation node, an object range, and a condition range are combined into a clause fragment. If a range is missing, it is explicitly recorded as empty within the fragment. The clause fragment records the start and end indexes of the text and the corresponding chapter / level path to ensure accurate positioning during subsequent cross-sentence processing.

[0022] Expand up to two sentences to the left and right of the clause fragment to look for exception markers. If an exception marker appears in the same enumeration group as the clause fragment, bind that exception marker to the clause fragment first. The binding rule follows the nearest-neighbor principle, that is, prioritize the exception phrase closest to the clause fragment. The reason for adopting the nearest-neighbor principle is that each item within an enumeration group usually only affects that item, while exceptions far away are more likely to be general restrictions on the entire clause and need to be handled separately. Scan the clause fragment and its parent clauses to identify time phrases such as year, quarter, fiscal year, start and end dates. If there is no time marker in the clause, trace back to whether there is a uniform time limit in the chapter or section containing this clause, such as in the current year or period. When multiple time phrases exist, prioritize the phrase that contains a more specific start and end range, because a clear time limit is more operational in the aggregation process. Scan the clause fragment and its parent clauses to identify time phrases such as year, quarter, fiscal year, start and end dates. If there is no time marker within the clause, trace back to the chapter or section containing this clause to see if there is a unified time limit, such as within the current year or period. When multiple time phrases exist, prioritize the phrase containing the more specific start and end range, as clear time boundaries are more operational in the aggregation process. Identify citation phrases such as "in accordance with this method," "this clause," "this paragraph," and "refer to superior laws." For citations on the same page, directly record the citation mark and location index. For cross-document citations, record the citation name and possible number; external links will be added later during knowledge alignment or manual review. The citation mark needs to be recorded independently because when generating clause nodes later, the citation needs to be linked to form a relationship edge for easy source tracing.

[0023] For sentences containing pronouns such as "the above," "this," "the preceding paragraph," "the preceding item," "the preceding item," and "the above expenditures," a backtracking approach is used in conjunction with the chapter hierarchy tree. The backtracking order is item level, clause level, sub-item level, section level, and chapter level, searching for the nearest referential noun phrase or clause fragment at each level. This fine-to-coarse backtracking order is adopted because referents usually preferentially point to objects at the same or nearby levels; only when the same level cannot be resolved does the scope need to be expanded. The pronouns are replaced with their actual references, and a citation marker is recorded in the clause fragment to indicate the source of the replacement process. If a sentence references multiple enumerated items through a pronoun, it is expanded item by item to generate multiple clause fragments, and the original sentence's position index is retained in each clause fragment to ensure traceability. Enumeration is commonly used for object scope or obligatory actions. For example, the object scope might be written as R&D project, technological transformation project, or pilot project. When processing enumerations, each item is saved in list form without sacrificing semantic clarity, and the enumeration sequence number is marked in the clause fragment. When enumerated items have the same conditions or the same scope markers, the common conditions and scope markers are promoted to shared attributes of the clause fragments to save subsequent processing.

[0024] Check if the clause fragment contains all three types of information: object, condition, and obligation. If only the condition is missing, but the object and obligation are clear, retain it and mark the condition field as empty. If either the object or obligation is missing, it is considered an incomplete clause fragment and its entry into the clause node generation queue is temporarily suspended; instead, it is placed in the completion pool. The completion pool attempts to complete the clause fragment again after cross-sentence reference is completed. After removing stop words, the object scope, condition scope, and obligation text of the clause fragment are compared with other clause fragments in the same clause for similarity. If all three are identical and the position index differs by no more than 20 characters, it is considered a duplicate, and the one appearing earlier is retained. The 20-character threshold is set because layout conversion often introduces very short offsets; exceeding this range usually indicates a genuine duplication rather than layout jitter.

[0025] Clause node numbering follows a hierarchical path plus fragment order rule. An example structure is chapter number - clause number - fragment number. Once generated, the number is written to the clause node and does not change with subsequent editing. Stable numbering facilitates consistent referencing by external systems or manual reviewers, preventing data chain breaks during data linkage. A clause node is generated for each clause fragment. The clause node stores the following: object scope text and its location index, condition scope text and its location index, obligation text and its location index, exception marker, time marker, scope marker, citation marker, hierarchical path, source document identifier, source publication time, and clause node number. All fields are directly readable combinations of original text fragments and indexes, avoiding black-box representation. Under the same clause level, if multiple clause nodes share the same condition scope, a common-origin relationship is established between the clause nodes. For clause nodes with citation markers, an external citation relationship pointing to the cited clause or document is established. A subordinate relationship is established between clause nodes and clause-level nodes in the chapter hierarchy tree for subsequent retrieval and display.

[0026] The clause nodes and their relationships are stored in a clause-centric graphical structure on the disk. The stored objects include clause nodes, object nodes, condition nodes, obligation nodes, and their corresponding marker nodes. Each node and edge carries a location index and source identifier. This clause-centric structure allows for rapid context recovery during subsequent matching phases, using clause nodes as entry points and avoiding repeated scanning of long texts. A comprehensive validation is performed after storage. Validation includes verifying the uniqueness of clause node numbers, whether location indices are within the range of the source text, whether bidirectional records exist for citation relationships, and whether item-level order is maintained during enumeration and merging. If duplicate clause node numbers are found, the entire storage of this clause is immediately rolled back and fragment numbers are regenerated; if location indices are out of bounds, the problematic clause is recorded and a validation report is generated for manual correction.

[0027] In standard sentences, obligatory verbs are the syntactic and semantic focus. Converging around the obligation center stably defines the boundaries between predicates and actions, and then collecting conditions and objects from both sides significantly reduces the probability of false inclusion. Conversely, if collection starts with objects, common background noun phrases are easily over-included, leading to an overly broad object range in subsequent matching. Time markers and scope markers are usually adjacent to their modified objects or modified actions. Using a cross-sentence global search can easily mis-bind general limitations to local actions, resulting in over-restriction. The proximity modification strategy, by limiting the search radius and priority, binds markers to the most likely entities, reducing bias and facilitating manual verification. Breaking up enumerated items into independent sentences and clause fragments allows for separate verification of each item during subsequent evidence collection and prevents the lack of evidence for one item from affecting the compliance assessment of other items.

[0028] In another optional implementation, when two policy texts exist on the same topic at the national and local levels, a source region identifier can be added during the initial acquisition stage. When generating clause nodes, the region identifier is written into the metadata field of the clause node. Subsequent use can be filtered according to the region identifier to ensure matching the correct scope of application. This method does not change the specific steps of the three-part structure extraction and modification item location, resulting in lower implementation costs. If a Chinese and English version exists, a unique English translation can be registered for each Chinese term in the terminology list. After the three-part structure extraction is completed, a cross-language index is built by aligning the object scope, condition scope, and obligation text sentence by sentence. The cross-language index uses the position index as the key and also saves the start and end offsets of the Chinese and English segments. The advantage of this is that when internal enterprise processes require outputting English materials, the English translation of the same clause node can be directly referenced, avoiding further manual translation that could lead to discrepancies. For policy documents with complex tables and multi-column layouts, common issues include sentences being broken across columns or column headings and content being separated into different lines within tables. After cleaning up layout noise, a row-block rearrangement step can be added: first, detect the start and end positions of columns within the same page, collect row blocks from top to bottom column by column, and then concatenate them according to column order. For tables, use column headers as candidates for object ranges, merge cell content into complete sentences row by row before proceeding to sentence delimiter segmentation. This method can significantly reduce broken sentences caused by manual typesetting and improve the completeness of clause fragments.

[0029] Step 2: Align policy terms with scientific research business terms to generate a term mapping index. The alignment process includes: constructing embedding boxes to represent the scope of application of terms, constructing hyperbolic embedding coordinates to represent the hierarchical relationship of terms, performing intersection retrieval on the embedding boxes and nearest neighbor filtering on the hyperbolic embedding coordinates to jointly confirm mapping pairs, and repairing semantic differences across contexts through Sheaf consistency diagnosis.

[0030] In one specific implementation, input one is a policy clause structure representation, including clause nodes, object nodes, condition nodes, obligation nodes, exception markers, time markers, scope markers, reference markers, and location indexes. Input two is a scientific research business terminology system, including an object category table, an activity stage table, a cost category table, a voucher type table, a regional identifier table, a time range table, and a necessary list of aliases. Input three is a terminology equivalence table, used to register synonyms and common aliases, such as R&D projects and scientific research projects belonging to the same object category. Equivalence relations are imported once at startup, and only new entries are allowed to be added during operation; registered entries are not overwritten. To ensure the determinism of subsequent steps, a unified embedded axis system is established for the semantic dimensions required for alignment. The set of axes is fixed as object category axis, activity stage axis, cost category axis, voucher type axis, regional identifier axis, time range axis, scope restriction axis, and exception restriction axis, for a total of 8 axes. The value of each axis comes from the enumerated set of the terminology list, and blank entries are not allowed. If a clause does not have a value that appears directly on a certain axis, then the global coverage value of that axis is used as a placeholder so that subsequent calculations are not interrupted.

[0031] A sequence number table is established for each axis. For example, the geographic identifier axis sequentially records four levels: national, provincial, municipal, and county; the time range axis sequentially records four levels: year, quarter, month, and fiscal year; and the expense category axis records items such as material costs, testing costs, labor costs, depreciation, and amortization of intangible assets. The sequence number table is fixed during system initialization and does not change with the text. The existence of the sequence number table allows for discrete processing of range comparison and adjacency judgment using the sequence number range, avoiding fuzzy judgments. For each clause node, terms associated with the eight axes are retrieved from the text of the object node, condition node, and obligation node. The retrieval order is: object category axis, expense category axis, voucher type axis, activity stage axis, geographic identifier axis, time range axis, range restriction axis, and exception restriction axis. When a search is successful, the matched terms are transcribed into the enumerated values ​​under that axis; the position index is also recorded for traceability. When a search fails, the full coverage value of that axis is used as a placeholder. When a clause node contains an exception flag, the value of the exception restriction axis is applied to other related axes. For example, if the exception constraint axis matches depreciation for non-R&D dedicated equipment, then depreciation entries are excluded from the current coverage set on the expense category axis, forming a differentiated expense category set. This processing allows subsequent intersection judgments to naturally exclude entries removed by exceptions, avoiding additional exclusions later. When clauses contain scope phrases such as "limited" or "must not exceed," a binding relationship is established between the scope constraint axis and the modified axis, forming a finer-grained coverage set on the modified axis. For example, if the time range axis matches the current year and the scope constraint axis matches "limited," then the entry on the time range axis is fixed as an annual entry, not extending to quarterly or monthly entries. This binding ensures that downstream mappings between different granularities do not skip levels. The value sets on the eight axes are merged into a single embedded box. The embedded box records the coverage set for each axis with eight fields, the field content being a list of enumerated sequence numbers for easy comparison. Each embedded box includes a clause node number and position index to ensure traceability. For business-side terms, the same method is used to generate embedded boxes for business terms. The policy side and the business side use the same axis set and the same sequence number table to ensure that the two have a basis for direct comparison.

[0032] like Figure 1 The diagram shown is a schematic representation of the eight-dimensional embedded box structure of the present invention. This diagram illustrates the eight dimensional axes of the embedded box constructed during the alignment of policy terms and business terminology, and their interrelationships. Figure 1In the center, a cube represents the core structure embedded within the box. Eight dimensional axes extend outwards from this center: object category axis, expense category axis, voucher type axis, geographic identifier axis, time range axis, scope restriction axis, exception restriction axis, and activity stage axis. The object category axis extends vertically upwards, with scales indicating object categories such as R&D projects, technological upgrades, and pilot projects; the expense category axis extends horizontally to the right, with scales indicating expense categories such as material costs, testing costs, and labor costs; the voucher type axis extends vertically downwards, with scales indicating voucher types such as invoices, contracts, and test reports; and the geographic identifier axis extends horizontally to the left, with scales indicating geographic levels such as national, provincial, and municipal. The other four dimensional axes are distributed diagonally: the time range axis extends from the center to the upper right, indicating time granularities such as year and quarter; the scope restriction axis extends to the upper left, indicating restriction types such as "only allowed" and "must not exceed"; the exception restriction axis extends to the lower left, indicating exceptions such as "except" and "but"; and the activity stage axis extends to the lower right, indicating stages such as project initiation, execution, and completion. Figure 1 In the diagram, the center of the embedded box is marked as the origin. Each clause node corresponds to an enumerated set of values ​​in eight dimensions. These sets of values ​​together constitute the embedded box representation of that clause. The dashed ellipse in the diagram indicates the scope of the coverage set, representing the coverage of a clause in multidimensional space. The core function of this structure is that by performing set intersection operations in eight dimensions, the compatibility between policy-side terminology and business-side terminology can be quickly determined, thereby achieving precise terminology alignment.

[0033] Based on authoritative classification standards and internal corporate governance documents, four hierarchical trees are established for object categories, expense categories, voucher types, and regional identifiers. Each tree follows a single-root, progressively detailed structure, with the root node representing the broadest concept and leaf nodes representing the most specific entries. For example, the root of the expense category tree is "expense," the first level includes R&D-related expenses and non-R&D-related expenses, and the second level further refines R&D-related expenses into material costs, testing costs, labor costs, etc. Nodes in the hierarchical trees are stored in their full Chinese names, without any abbreviations. Each term is assigned a hyperbolic embedded coordinate system. The radial index equals the depth of the level in which the term resides, starting from 1 and increasing layer by layer; the angular index equals the term's position within that level, ordered lexicographically from left to right. For example, if "material cost" is located in the second level and ranks 3rd in the pinyin order within that level, its radial index is 2 and its angular index is 3. Using this coordinate allocation, the proximity of any two terms can be determined by the position of their common superior term and their adjacency within the same level, without the need for continuous parameters. Based on hyperbolic embedding coordinates, a hyperbolic nearest neighbor table is built for each term. The nearest neighbor table contains three concentric levels: two adjacent terms at the same level, the direct parent term of the upper level, and all direct child terms of the lower level. These three types of nearest neighbors form a directed adjacency list. These three types of nearest neighbors are chosen because alignment needs to consider both alternative expressions at the same level and preserve the shortest transformation paths for upward generalization and downward refinement; the adjacency list ensures that these three most common structural adjacencies are covered. The policy side and the business side are defined using the same order and level in the four hierarchical trees. If a unique entry exists on the enterprise side, it is appended as a leaf node under the nearest parent term, with its source noted. The tree definition is frozen once confirmed and does not fluctuate with new text during runtime.

[0034] For each policy-side term's embedding box, iterate through the business-side term's embedding boxes. Compare the coverage sets one by one for each of the eight axes. The comparison method is to intersect the sets along each axis, with the set element being the enumeration index of the corresponding axis. The comparison order is: object category axis, expense category axis, voucher type axis, geographic identifier axis, time range axis, range restriction axis, exception restriction axis, and activity stage axis. This order is fixed to facilitate process reproduction during log auditing. When all eight axes have a non-empty intersection, it is marked as fully compatible; when the object category axis and expense category axis have a non-empty intersection, and all other axes have non-empty intersections or full coverage, it is marked as partially compatible; otherwise, it is marked as incompatible. Full coverage is treated as a wildcard in this method, participating in the comparison but not used to negate intersection. If an exception restriction axis results in an empty intersection of an axis, the comparison is directly marked as incompatible. Add fully compatible and partially compatible term pairs to the candidate set, and assign an intersection judgment label to each pair of records. The candidate set enters the nearest neighbor filtering in the next step. Incompatible term pairs are discarded and not included in subsequent calculations.

[0035] For each pair of terms in the candidate set, a common superordinate term is searched from bottom to top in the corresponding hierarchical tree. The search method is to trace up the parent pointer layer by layer until both point to the same node. The hierarchical depth of the common superordinate term and the number of steps from both to that node are recorded. When the common superordinate term is in the first three layers, it is marked as strongly close. When the common superordinate term is in the fourth and fifth layers, it is marked as moderately close. When the common superordinate term is in a deeper layer, it is marked as weakly close. The discrete division of closeness makes the screening criteria stable and does not depend on adjustable parameters. If two terms are in the same layer and the difference in their angular indices does not exceed 2, they are additionally marked as adjacent at the same layer. This can capture the relationship of alternative expressions at the same layer, which is very common in industry terms. The closeness and adjacent at the same layer information are combined to form a structural adjacency matching label. Strong closeness or adjacent at the same layer are recorded as positive labels; moderate closeness is recorded as a positive label when both the object category axis and the cost category axis are positive; weak closeness is recorded as a reserved label and will be further processed by consistency diagnosis later.

[0036] When the intersection decision label of a pair of terms is fully compatible and the structural adjacency matching label is positive, the pair is directly confirmed as a mapping pair. When the intersection decision label is partially compatible and the structural adjacency matching label is positive, check for conflict markers between the scope limitation axis and the exception limitation axis. If there is no conflict, the pair is confirmed as a mapping pair. When the structural adjacency matching label is a reserved label, it is not directly confirmed, and the consistency diagnosis process is initiated. A conflict on the scope limitation axis is defined as a situation where the policy-side limitation and the business-side limitation are opposite in direction on the same axis, for example, the policy side specifies "only" while the business side specifies "not less than". A conflict on the exception limitation axis is defined as an item excluded by the policy side appearing as a necessary item on the business side. Once a conflict is identified, a clear mark is left on the mapping pair for subsequent repair. Each confirmed mapping pair records the policy term identifier, business term identifier, embedding box snapshot, hyperbolic embedding coordinates, intersection decision label, structural adjacency matching label, common superior term and its hierarchy depth. All fields are traceable to the source text and hierarchy tree, ensuring that the process can be replayed step by step during debugging and auditing.

[0037] The nodes in the base space graph include clause nodes, reference nodes, and version nodes. Edges represent the reference relationships between clauses and the evolutionary relationships between versions. Coverage sets include clause context, clause explanation appendices, Q&A appendices, and caliber descriptions. Each coverage set is bound to a set of clause nodes, and the source and time are recorded in the binding relationship. Fibers are created on each coverage set, recording the terms appearing within the coverage, their corresponding embedding boxes, hyperbolic embedding coordinates, and naming terms. Fibers serve as a local semantic view of the coverage, and subsequent consistency comparisons are performed within the fibers. Restriction mapping is defined for coverages with inclusion relationships. Restriction mapping includes two deterministic operations. The first is a semantic equivalence merge operation, used to unify aliases and abbreviations to the base terms in the terminology equivalence table, generating updated naming terms. The second is an attribute projection operation, used to retain the common set of axis values ​​between parent and child coverages and remove local values ​​that only appear in child coverages. Restriction mapping is performed in a fixed order: merge first, then projection. This order is fixed in the project documentation and cannot be changed. Push checks are performed along loops in the base space graph. Select the loop start cover, and push the fiber content of the start cover to the next cover according to the constraint mapping; push edge by edge until returning to the start cover; compare the pushback result with the current fiber of the start cover item by item. Item by item comparison includes object category axis, cost category axis, voucher type axis, region identifier axis, time range axis, range restriction axis, and exception restriction axis. When completely consistent, it is marked as a consistent unit; when any difference occurs, a consistency difference record is generated.

[0038] For each consistency discrepancy record, a repair sequence is executed. The first step checks if the inconsistency is due to naming differences. If so, a new equivalent entry is added to the terminology equivalence table, and the loop check is re-executed. The second step checks if the inconsistency is due to different axis coverage granularity. If so, expansion or splitting is performed on the embedded box. Expansion involves including adjacent values ​​under the same parent term on the axis, causing the two coverage sets to intersect; splitting involves dividing a coverage set into two non-overlapping subsets, creating a one-to-one correspondence with the opposite coverage set. After repair, the loop is re-executed until a consistent unit is obtained. All repair actions are recorded in an operation log, including the time, operator, coverage identifier, and changes. For candidate pairs with reserved labels for structural adjacency matching, if all involved coverage sets reach a consistent unit after Sheaf consistency diagnosis, the candidate pair is promoted to a mapping pair; if discrepancies still exist, the candidate pair is retained in the observation area of ​​the terminology mapping index, awaiting further diagnosis after subsequent clause or business term updates.

[0039] The terminology mapping index uses policy terminology identifiers and business terminology identifiers as a joint key, storing intersection judgment labels, structural adjacency matching labels, common superior terms, embedding box snapshots, hyperbolic embedding coordinates, consistency status, coverage set list, and last verification time. Entries in the observation area are distinguished by individual status identifiers to prevent misuse in downstream processes. When new policy texts or business terms enter the system, incremental alignment is performed only on the affected entries. Incremental alignment is performed using the clause node number and hierarchical tree position as anchors, without overwriting existing mapping pairs. When an equivalence table expansion occurs, Sheaf consistency diagnostics are triggered and automatically replayed once on the relevant loop, updating the consistency status and repair logs. When downstream processes read the terminology mapping index, consistency status is used as the first filtering condition, selecting only mapping pairs with a consistent status; if business needs require, observation area entries can be displayed but not involved in executable decisions. This ensures the stability and interpretability of the process's terminology.

[0040] In one alternative implementation, when different systems within an enterprise use different names for the same voucher type, source priority can be introduced in the semantically equivalent merge operation. Source priority is given during initialization, for example, using the archive system name as the base name and the expense system name as the secondary name. During merge, secondary names are prioritized to be unified to the base name to avoid multiple base candidates in the index. When policy text only contains the fiscal year while business systems only contain the calendar year, an interleaving completion step can be inserted before the attribute projection operation. Interleaving completion uses the month sequence as the smallest granularity, aligning the fiscal year and calendar year at the month level, and then promoting them to annual or fiscal year entries. This refinement-then-promotion approach can establish a verifiable mapping without introducing continuous parameters. When data submitted by the business system only reaches the prefecture-level city level, while policy provisions take effect at the provincial level, a roll-up strategy is adopted for the regional identifier axis. The roll-up strategy allows prefecture-level entries to be rolled up to provincial entries on the regional identifier axis to ensure that intersection searches can form a non-empty intersection. Roll-up is only enabled when the policy side has a higher level, avoiding meaningless roll-down.

[0041] Step 3: Refine the terms and obligations into a set of obligation slots, transform business events into a set of evidence units, and complete the pairing by executing a stable matching process according to the order of mutual preferences.

[0042] In one specific implementation, the obligation text is read from the clause node. The obligation text is split into several obligation actions according to the order of appearance of verb phrases, such as submitting, providing, collecting, verifying, registering, and retaining. If verb phrases are connected by coordinating conjunctions, they are split into multiple obligation actions; if they appear as gerunds and are separated by commas, they are split into multiple obligation actions enumerated. After splitting, each obligation action is processed separately, ensuring that one obligation action generates a set of obligation slots. The object phrases of the obligation actions are traversed to identify the voucher type and content elements. Examples of voucher types include invoices, contracts, orders, test reports, experimental records, technical solutions, and meeting minutes. Examples of content elements include issuance date, payee, amount field, project number, and signature status. For each voucher type, the terminology mapping index is queried to obtain the standard name and possible synonyms. If the obligation text contains phrases such as "equivalent materials" or "equivalent data," the parent term of that voucher type is searched in the terminology mapping index, and its direct sub-items are appended to the list of acceptable alternatives. The advantage of this approach is that the alternatives come from the standardization level, providing a basis for review.

[0043] The expense categories, object categories, and obligations are linked together in the clause nodes. If the obligation text includes expressions such as "collection of material costs" or "retention of testing fees," a separate obligation slot entry is generated for each expense category. If the obligation text does not specify an expense category, but the clause node's conditions include a list of expense categories, the list is expanded, and each item generates a separate slot. This expansion prevents a voucher from being mistakenly used for an unrelated expense category. Unless further specific limitations are specified, the obligation slot inherits the time range and geographical identifier of the clause node. When the upper-level clause of the clause node has a unified time limit, such as "this year" or "this fiscal year," the obligation slot uses that limit. If the obligation text includes time expressions such as "at the time of declaration" or "at the time of final settlement," a submission time reminder field is also recorded for subsequent gap list reminders.

[0044] If the obligation text contains scope restrictions such as "limited," "at least," or "at most," then that restriction is bound to the entity it modifies. For example, "limited to expenditures incurred after project initiation" is bound to a time frame; "at least includes experimental records and test reports" is expanded into multiple parallel slots, each bound to a voucher type. If the clause node contains exceptions, such as depreciation of non-R&D dedicated equipment, then that category is removed from the slot set for the corresponding expense category or an explicit exclusion flag is generated for that category. Explicit exclusion prevents subsequent use of evidence units to satisfy slots that are not applicable. For scenarios where the same obligation action requires multiple vouchers, such as providing invoices, contracts, and experimental records, it is expanded into multiple independent obligation slots, each containing only one voucher type requirement. This allows for one-to-one verification of fulfillment while avoiding duplicate counting of a single evidence unit. A stable number is assigned to each slot. The number consists of three parts: the clause node number, the obligation action sequence number, and the slot sequence number. Stable numbers facilitate cross-system referencing and auditing. Each obligation slot records the following fields: slot number, clause node number, obligation action, object category set, fee category set, voucher type main item, acceptable alternative list, region identifier, time range, scope restriction, exception restriction, source location index, and submission time reminder.

[0045] Event records are retrieved from the procurement system, contract system, expense reimbursement system, project management system, and document system. Collected fields include event identifier, object category, activity stage, expense category, amount, occurrence date, project number, and department information. Each source field is transcribed into standard terminology using a terminology mapping index, ensuring that key fields such as object category, expense category, and voucher type are identical in name and level to policy-related terminology. Image and text records are collected for each event record. Examples of image records include invoice images, scanned copies of contracts, and photos of acceptance forms. Examples of text records include searchable versions of test reports and exported versions of experimental records. Image and text records are bundled by event identifier; if a document involves multiple events, the document reference is copied, and the original location and fingerprint information are retained on each reference to prevent duplicate statistics. For each image and text record, common elements such as the recipient, issuance date, signature status, amount, quantity, material name, specifications, contract number, and counterparty name are extracted. The extraction process combines rules and key phrase location methods. For example, the amount field is usually located as a numerical value to the right of the word "amount," and the issuance date is usually located as a date string to the right of the word "date." After extraction, an integrity check is performed. If the document type requires both the header unit and the issuance date, the document is marked as incomplete if either is missing.

[0046] The event record and its associated materials are merged into a single evidence unit. Each evidence unit contains event-dimensional and material-dimensional fields, along with source information and a location index for traceability. A stable number is assigned to each evidence unit. This number consists of an event identifier segment and a sequence number. If an event generates multiple materials, each material corresponds to an evidence sub-unit; a parent-child relationship is established between evidence sub-units and evidence units. This parent-child relationship allows for pairing at the material level and also provides a comprehensive overview at the event level. When a material covers multiple expense categories, such as an invoice containing material costs and testing fees, it is split into two evidence sub-units. During splitting, the same image location and issuance date are used, but the material cost and testing fee are recorded separately in the expense category field. This splitting ensures a one-to-one correspondence between each document and avoids confusion in aggregation due to an invoice covering multiple categories. Availability labels are calculated for each evidence unit. Availability marking is based on three hard conditions: first, the voucher type belongs to a standard item or acceptable alternative in the terminology mapping index; second, the occurrence date falls within the relevant time range of the clause or has a clear explanation of the time span; and third, the material completeness check passes. If any condition is not met, the evidence unit is marked as unusable for pairing and retained in the redundant evidence list for manual review.

[0047] A stable matching process requires both parties to have a preference order for each other. This embodiment uses deterministic sorting rules to ensure consistent results in every run. Each obligation slot generates an independent evidence preference list. The sorting rules compare items one by one in the following deterministic order until a difference is found: Step 1, compare whether the voucher type is consistent with the main item. Those consistent with the main item take priority; if both are inconsistent, those consistent with the acceptable alternative take priority. Step 2, compare whether the expense category is completely consistent. Those completely consistent take priority; if both are consistent, proceed to the next step; if only one side is consistent, that side takes priority. Step 3, compare whether the object category is included in the object category set of the obligation slot. Those included take priority. Step 4, compare whether the occurrence date is within a narrower time range. Those within a narrower time range take priority. For example, if the clause allows an annual range, evidence within the corresponding month of the project's monthly plan takes priority. Step 5, compare the completeness of the materials. Those with complete and essential materials take priority. Step 6, compare the chronological relationship between the issuance date and the declaration node. Those with an issuance date earlier than the application deadline and closer to the application deadline will be given priority. Step 7: If there is still a tie, the evidence number with the earlier lexicographical order will be given priority.

[0048] Each evidence unit generates a slot preference list. The sorting rules are determined by comparing items in the following order: Step 1, compare the document type with the main or substitute item. Slots consistent with the main item take priority. Step 2, compare the expense category. Slots completely consistent take priority. Step 3, compare the object category or whether it is consistent after being rolled up. Consistent slots take priority. Step 4, compare the regional identifier or whether it is consistent after being rolled up. Consistent slots take priority. Step 5, compare the time range. Slots completely within the time range take priority. Step 6, if still tied, the earlier one in the lexicographical order of the clause node number takes priority; if still tied, the earlier one in the lexicographical order of the slot number takes priority. This evidence preference list prevents evidence from being drawn to obligation slots unrelated to its original design purpose, reducing the burden of subsequent review.

[0049] The default convention is one slot corresponding to one evidence unit, and one evidence unit corresponding to one slot. When the terms explicitly allow one piece of evidence to be used for multiple purposes or one piece of evidence to correspond to multiple items, this convention is implemented as multiple parallel slots during slot expansion, with each slot still paired one evidence unit in a one-to-one manner. By expanding at the structural level, the matching process maintains a one-to-one form, ensuring the convergence and interpretability of the algorithm.

[0050] This embodiment employs a delayed acceptance stable matching process, initiated by the obligatory slot. The reason for choosing the obligatory slot as the initiator is that, with the business process aiming for compliance, proactively seeking evidence is more suitable for compliance verification scenarios than passively waiting for evidence to be attached. All obligatory slots and all evidence units are marked as unmatched. Each obligatory slot holds a list of evidence that has not yet been submitted, initially the entirety of its preference list. Each evidence unit holds an empty current reservation record. In the first round, all unmatched obligatory slots simultaneously submit their evidence to the first evidence unit in their respective preference lists that has not yet been submitted. Each submitted evidence unit checks its own current reservation record. If the reservation is empty, the submission is recorded as a temporary reservation. If the reservation already exists in another slot, the two slots are compared according to the evidence-side preference order rule. The one with the higher preference becomes the new temporary reservation, and the other slot is reverted to the unmatched state. An evidence unit retains only one submission at any given time to ensure the exclusive agreement of one evidence unit per slot. Slots that are pushed back to an unmatched state have their previously proposed evidence units removed from their preference list, and in the next round, they are proposed to the new, highest-ranking evidence unit that has not yet been proposed. If a slot's preference list has been completely filled, the slot will not propose any more evidence units and will enter a pending state, later being added to the gap list. The above proposal and retention process is executed iteratively round by round. When no new proposals are made in a round, the process terminates. The temporary retention of all evidence units is upgraded to final pairing upon termination. This process ensures that each slot proposes to each piece of evidence at most once, thus limiting the total number of rounds. Since the comparison order is fixed, the results of any two runs are consistent. During the proposal and retention process, if a conflict is found between the geographic identifier or time range of an evidence unit and a slot, the evidence unit immediately rejects the proposal, and the slot unconditionally pushes back and proposes the next object in the next round. The benefit of immediate rejection is that unusable combinations are removed as early as possible, reducing subsequent meaningless comparisons.

[0051] After matching is complete, the system needs to output the results in a structured format for easy archiving, submission, and auditing. The final matched items are divided into several batches using the clause node number as the first grouping key, the activity stage as the second grouping key, and the expense category as the third grouping key. This grouping is based on the consideration that audits at the clause level typically examine the clauses first, then the stages and expenses, allowing for the verification of similar materials within a single batch. All unmatched obligation slots are added to a gap list. For each slot, the system lists the main voucher type, a list of acceptable alternatives, a set of object categories, a set of expense categories, a geographic identifier, and a time range and scope restrictions. If an alternative at the same level as the main voucher exists in the terminology mapping index, it is suggested in the list. If a submission time suggestion exists, a suggested completion date is provided, such as a suggestion to complete the completion 7 days before the submission deadline. All unselected evidence units are added to a redundant evidence list. For each evidence unit, the system lists its closest slot candidate and the reason for rejection, such as the voucher type not being in the main or alternative items, the occurrence date being outside the time range, or the material completeness check failing. Redundancy lists help business personnel handle materials that are mistakenly uploaded or miscategorized. A link record is generated for each final pairing. The link record includes slot number, clause node number, obligation action, expense category, object category, evidence number, material list, location index, and comparison process summary. The comparison process summary describes the key decision points in the preference order, facilitating explanation during audits of why this evidence was chosen over others. Link records are written to audit storage and can be replayed at any time.

[0052] In one alternative implementation, when the date of issuance of evidence and the date of occurrence are in adjacent years and the terms allow for the submission of materials during the tax return filing, a cross-period description can be added to the time range field. During preference comparison, evidence with a cross-period description and whose occurrence date falls within the time range is considered to be within the time range, ensuring that the evidence can still be used as the filing deadline approaches. When a document contains multiple content blocks, such as a test report containing multiple test sub-conclusions, the text record can be divided into content blocks during the evidence unit generation stage, with each content block forming an evidence sub-unit. During matching, each evidence sub-unit participates in the matching process independently, avoiding the entire report being used as a whole.

[0053] Step 4: Generate an action list and process checkpoint rules based on the pairing results, and solidify the generated terms-evidence-action chain that includes the embedded box and hyperbolic embedded coordinate snapshots.

[0054] Specifically, the goal of step 4 is to transcribe the final pairings and gap list into a set of actionable actions for the business system. The action list uses an atomic action design, where each action corresponds to only one specific task, facilitating verification and replay. Atomic actions are used because requirements often appear in parallel; if multiple tasks are mixed in the same action, it becomes impossible to make a clear judgment when some tasks are partially completed.

[0055] There are six fixed action types. The first type is data collection, used to guide the completion of materials in the gap list. The second type is data collection and accounting, used to classify paired evidence units into the corresponding expense categories and collection batches. The third type is caliber verification, used to confirm the consistency of names, levels, and enumerated values ​​by comparing them with the terminology mapping index. The fourth type is consistency check, used to verify the consistency of fields for the same event in different systems, such as the amount field and the project number. The fifth type is declaration preparation, used to generate catalog files, packaged images, and compile catalog numbers before the declaration node. The sixth type is archiving and evidence preservation, used to write the completed materials and link records into the audit storage. For each final paired record, action entries are generated in the following order: Step 1, generate the data collection and accounting action. The action content includes the clause node number, the expense category in the obligation slot, the evidence number, and the collection batch number. The reason for generating data collection and accounting first is that this action puts the evidence into a clear position in the accounting caliber, and subsequent verification and declaration are based on this. Step 2, generate the caliber verification action. This action compares the object category, expense category, voucher type, and terminology mapping index of the evidence unit with the baseline name. If aliases appear, the correspondence between the aliases and baseline names must be written into the equivalence table upon completion of the action, and a change record must be retained in the link record. Step 3: Generate a consistency verification action. This action compares the key fields of the same evidence in different systems. If the difference comes from an interpretable format difference, such as a different date format, it is marked as a format difference in the action record, and a normalized copy is generated upon completion. Step 4: Generate a declaration preparation action. The action content includes the image list, text list, directory structure, naming rules, and clause node numbers that need to be presented on the directory page to be included in the package. Step 5: Generate an archiving and evidence preservation action. This action stores the completion results of the actions generated by this pairing into the database and adds the completion time and the person who completed the action to the link record. For each slot in the gap list, a data collection action is generated. The action item lists the main voucher type and acceptable alternatives, the object category set, expense category set, regional identifier, and time range, and provides a submission time prompt. The data collection process explicitly requires the provision of materials with identifiable location indexes to avoid uploading fragmented documents that cannot be linked to the event. The action list is sorted by clause node number from smallest to largest, and within the same clause, by activity stage. Within the same stage, data collection is performed first, followed by caliber verification and consistency checks, then data aggregation and accounting, and finally, application preparation and archiving. This sequence ensures that materials are completed first, calibers are standardized first, data aggregation is based on evidence, and applications are packaged last, aligning with the actual processing path and providing clear deliverables upon completion of each step.

[0056] Each action includes an action identifier, action type, clause node number, obligation slot number or gap slot number, evidence number or target material description, collection batch number, object category set, expense category set, geographic identifier, time range, scope limitation, exception limitation, source location index, completion criteria, failure message, and submission time prompt. Completion criteria use directly verifiable conditions, such as the existence of an image file, the existence of a readable issuance date, and consistency with the terminology mapping index base name. Prompts use clear, short natural language sentences, such as "Please upload an image of the contract containing the signature," to avoid ambiguity. Process checkpoint rules are used to block submissions that do not meet clause requirements at key nodes or to perform consistency checks during node transitions. Checkpoint rules must be directly executable by the business system; therefore, rule expressions should be structured, and the decision path should be clear and replayable.

[0057] like Figure 3 The diagram illustrates the distribution of the three-part clause structure (object-condition-obligation) in semantic space. This diagram, presented in three dimensions, shows how the clause nodes generated after step one's parsing form clusters in semantic space based on their object scope, condition scope, and obligation text. Figure 3 A three-dimensional coordinate system was established, with the X-axis representing the object dimension, the Y-axis representing the condition dimension, and the Z-axis representing the obligation dimension. Typical object categories such as R&D projects, technological upgrades, and pilot projects are marked on the X-axis; time condition types such as year-end, specific stage, and inter-period conditions are marked on the Y-axis; and different obligation action types are represented by the Z-axis. In this three-dimensional space, a single clause node is represented by a solid black circle. Multiple clause nodes naturally cluster together to form clause clusters based on the similarity of their objects, conditions, and obligations, with the boundaries of each clause cluster marked by dashed ellipses. The figure shows four main clause clusters: the first cluster is related to material costs, containing 5 nodes distributed around coordinates (280, 430); the second cluster is related to testing costs, containing 6 nodes distributed around coordinates (425, 285); the third cluster is related to labor costs, containing 4 nodes distributed around coordinates (535, 385); and the fourth cluster is a cross-category clause cluster, containing 3 nodes located around coordinates (380, 200). The diagram specifically highlights three dashed lines emanating from a given clause node, pointing to the scope of the object, the scope of the conditions, and the text of the obligation, respectively, illustrating that each clause node contains complete information about these three structures. Furthermore, the diagram uses solid and dashed lines to mark the semantic distances d1 and d2 between clause clusters, indicating the proximity of different clause clusters in the semantic space. Irregular curves also delineate high-similarity regions, where clause nodes exhibit a high degree of overlap in the three dimensions of object, condition, and obligation.

[0058] The rules are fixed into seven categories. Category 1 is the pre-submission verification rule, triggered at the document submission interface. Category 2 is the inter-node transfer verification rule, triggered during the transition between project initiation, execution, and completion. Category 3 is the pre-archiving verification rule, triggered before document entry into the database. Category 4 is the cross-period acceptance verification rule, triggered during the check before the application node. Category 5 is the regional applicability verification rule, triggered during project establishment and cost collection. Category 6 is the exception exclusion verification rule, triggered during cost collection. Category 7 is the duplicate occupancy verification rule, triggered when the action is completed and entered into the database. For each final paired record and each missing slot, the rules are generated according to the following steps.

[0059] The pre-submission validation rule generation method includes: Step 1, establishing mandatory validations at the pre-submission validation location based on the main document type and acceptable alternatives in the obligation slot. Submission is only allowed if any type of material exists and passes the material integrity check. Mandatory validations are used to ensure the minimum material set is in place and avoid empty submissions. Step 2, adding a consistency check between object category and expense category to the pre-submission validation. Submission is only allowed if the object category is included in the slot's object category set and the expense category matches the slot's expense category. This prevents incorrect materials from occupying incorrect slots. Step 3, adding a time range check. When the occurrence date is outside the time range, a cross-period explanation is required. When a cross-period explanation occurs, the pre-submission validation rule binds this material to the raw material and submits it as a complete package. The benefit of this binding is that subsequent reviews can check the supporting evidence together.

[0060] The generation method for inter-node transfer verification rules includes: Step 1, when transitioning from project initiation to execution, checking whether the object category is consistent with the project initiation record. If inconsistent, blocking the transfer and prompting to correct the object category using the base name of the terminology mapping index. Step 2, when transitioning from execution to project completion, checking whether each collection batch has completed the collection and accounting entry actions and the scope verification actions. If not completed, listing the incomplete action identifiers and blocking the transfer. This check ensures that the materials have been completely included in the accounting scope by the time of project completion.

[0061] The pre-archiving verification rule generation method includes: Step 1, checking whether the completion result of each action has a location index, which can be located in the source text or image. Materials lacking location indexes are not allowed to be archived. Step 2, checking whether each link record has an evidentiary fingerprint. Links without evidentiary fingerprints must not be archived. Evidence fingerprints are used to verify that materials have not been replaced in subsequent audits.

[0062] The method for generating cross-period processing verification rules includes: Step 1, when there is a time requirement for the application node in the clause node, a check is triggered 7 days before the application node. If there are any gaps or incomplete slots, an emergency data collection action is generated and the reminder frequency is increased. The 7-day advance selection gives business personnel an actionable window to fill in the gaps, avoiding blockages on the day of the node.

[0063] The method for generating regional applicability verification rules includes: Step 1, verifying the regional identifier during cost collection. If the clause is effective at the provincial level but the project's regional identifier is at the prefecture-level city level, then uploading to the corresponding provincial level is allowed; if the clause is effective at the prefecture-level city level but the project's identifier is at the provincial level, downloading is not allowed, the operation is blocked, and a prompt is displayed to supplement the materials according to the prefecture-level city level. This process ensures that the scope of application is not expanded.

[0064] The exception exclusion verification rule generation method includes: Step 1, checking exception restrictions during cost aggregation. If the cost category of the evidence unit matches the exception restriction, such as depreciation of non-R&D dedicated equipment, the aggregation is directly blocked and a message is displayed indicating that the category is excluded.

[0065] The duplicate occupancy check rule generation method includes: Step 1, checking whether the same evidence number has been occupied by other slots before being put into storage. If occupied, checking whether it is a slot under the same clause. If so, allowing sharing and recording the reason; if not, blocking and prompting that each piece of evidence can only be used by one slot. Occupancy checks can prevent materials from being misused in irrelevant clauses. To facilitate execution by various business systems, each rule also generates a field mapping list. The field mapping list lists the correspondence between field names that can be directly read in the system and fields on the clause side. For example, expense category corresponds to the expense category field in the expense system, occurrence date corresponds to the expense occurrence date field, region identifier corresponds to the project region field, and location index corresponds to the page offset field in the file system. By clearly defining the mapping relationship, business systems can execute the rules without understanding the clause text. Each rule is accompanied by human-readable prompts and correction paths. Prompts use clear, directive sentences, such as "Please upload an image of the contract containing the signature" or "Please select the R&D project in the object category." Correction paths clearly point to the system location and field name that needs to be operated on, reducing communication costs.

[0066] Link solidification involves storing each complete path from the clause node to the evidence unit and then to the action completion result in auditable storage, along with a reproducible context snapshot. The context snapshot includes an embedding box and hyperbolic embedding coordinate snapshot, enabling accurate reconstruction of the original alignment basis during subsequent changes in scope or cross-period reviews. Each link record contains the following elements: Item 1 is clause-side information, including clause node number, object scope text and location index, condition scope text and location index, obligation text and location index, geographical identifier, time range, scope restrictions, and exception restrictions. Item 2 is evidence-side information, including evidence number, material list, material location index, evidence fingerprint, object category, expense category, occurrence date, and issuance date. Item 3 is action-side information, including action identifier, action type, completion criteria, completion time, person responsible, number of retries, and final status. Item 4 is the alignment snapshot, including an embedding box and hyperbolic embedding coordinate snapshot. Item 5 is a summary of the generation process, including the comparison order during matching, the number of rejected candidates, and a summary of the main rejection reasons.

[0067] The embedded box snapshot records the value set of eight axes. Each axis stores a list of enumeration numbers and their equivalent names at that time. The enumeration number list is recorded because it does not drift with text changes and can be stably aligned across versions. For entries excluded by exceptions, an exclusion list is retained in the snapshot, showing the exclusion scope at that time. The hyperbolic embedded coordinate snapshot records the radial and angular indices of related terms in the four hierarchical trees. The radial indices equal the level depth, and the angular indices equal the intra-level order. The snapshot also retains common superordinate terms and level depth. The advantage of saving these discrete coordinates is that if new leaves are introduced into the hierarchical trees or nodes at the same level are rearranged, the original nearest neighbor relationships can still be restored using the original angular indices. Link fingerprints are generated simultaneously when the link record is written to audit storage. The link fingerprint is calculated by concatenating fields from the clause-side information, evidence-side information, alignment snapshot, and action-side information. Any change in any field will change the fingerprint. Auditors can use this to verify whether the link has been modified. The link record employs an append-only writing strategy; new revisions generate new link version numbers, while older versions are retained for review. In auditing or review scenarios, the pairing and judgment process can be replayed starting from the link record. During playback, the embedded box snapshot and hyperbolic embedded coordinate snapshot are read sequentially, and the intersection retrieval and nearest neighbor filtering steps are reenacted according to the determined comparison order. Then, each completion criterion in the action list is checked against the actual action. Since all information is embedded within the link, playback does not depend on external runtime conditions, ensuring repeatability.

[0068] like Figure 3 The figure shows a comparison of the accuracy of different term alignment methods. This figure demonstrates the significant advantages of the method of this invention compared to existing technologies through experimental data comparison. Figure 3In the figure, the horizontal axis represents the number of term pairs, in units of 100, with the scale gradually increasing from 100 to 700; the vertical axis represents the alignment accuracy, in percentage (%), with the scale ranging from 10% to over 90%. The figure shows three curves, each representing a different alignment method. The first curve, a solid line with a solid dot, represents the complete method of this invention, which uses a combined alignment method of embedded boxes, hyperbolic embedded coordinates, and Sheaf consistency checks. The second curve, a long dashed line with a hollow dot, represents a simplified method that uses only embedded boxes without hyperbolic coordinates and consistency checks. The third curve, a short dashed line with square dots, represents the alignment method based on traditional string similarity. The curve trends show that the accuracy curve of this invention is consistently at the top, reaching 82% accuracy with 100 term pairs, and stabilizing above 91% as the number of term pairs increases to 700. In comparison, the accuracy of methods using only embedded boxes is between 70% and 75%, while the accuracy of traditional string similarity methods is only between 50% and 55%. The figure specifically highlights two performance improvements: with 400 term pairs, the method of this invention improves by 9 percentage points compared to the method using only embedded boxes; with 650 term pairs, it improves by 29 percentage points compared to the string similarity method. These experimental results fully demonstrate that this invention, by introducing hyperbolic embedded coordinates to represent hierarchical relationships and combining Sheaf consistency diagnostics to repair semantic differences across contexts, can significantly improve the accuracy of term alignment, especially when dealing with large-scale term sets.

[0069] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A tax policy semantic parsing and automatic adaptation engine for scientific research operations, characterized in that: The engine performs the following steps: Step 1: Standardize and parse the tax policy text to form a policy clause structure centered on clause nodes; Step 2: Align policy terms with scientific research business terms to generate a term mapping index. The alignment process includes: constructing embedding boxes to represent the scope of application of terms, constructing hyperbolic embedding coordinates to represent the hierarchical relationship of terms, performing intersection retrieval on the embedding boxes and nearest neighbor filtering on the hyperbolic embedding coordinates to jointly confirm mapping pairs, and repairing semantic differences across contexts through Sheaf consistency diagnosis. Step 3: Break down the obligations into a set of obligation slots, transform business events into a set of evidence units, and complete the pairing by executing a stable matching process based on the order of mutual preferences. Step 4: Generate an action list and process checkpoint rules based on the pairing results, and solidify the generated terms-evidence-action chain that includes the embedded box and hyperbolic embedded coordinate snapshots.

2. The method according to claim 1, characterized in that, The standardization and parsing steps for policy clauses specifically include: establishing a domain terminology and a list of node types; extracting the three-part structure of object, condition, and obligation based on sentence segmentation and domain template identification, while also extracting exceptions, time, and scope, and marking cross-sentence references within the same clause to generate a set of clause fragments; and mapping the set of clause fragments into a policy clause structure representation centered on clause nodes and connecting object, condition, obligation, exception, time, and scope nodes.

3. The method according to claim 1, characterized in that, The construction of the embedded box specifically includes: performing discrete value statistics along a fixed axis system composed of object category, activity stage, expense category, voucher type, regional identifier, time limit marker, range marker and reference relationship, and determining the coverage start point and coverage end point on each axis to form the embedded box; for axes with exception markers, the coverage set of the axis is differentially processed by the exception markers.

4. The method according to claim 1, characterized in that, The construction of hyperbolic embedding coordinates specifically includes: generating a term hierarchy tree and a relational adjacency table based on the domain hierarchy and the hierarchical relationship of terms; assigning hyperbolic embedding coordinates to each term with the level depth as the radial index and the same-level lexicographical order as the angular index.

5. The method according to claim 1, characterized in that, The intersection retrieval and nearest neighbor selection specifically include: calculating the intersection state of each axis for each pair of terms and generating three types of intersection judgment labels: fully compatible, partially compatible, and incompatible; for candidate pairs that are fully compatible and partially compatible, searching the hyperbolic nearest neighbor table, and if the reference set has items with the same superordinate terms on both sides, then the candidate pair is marked as structural adjacency matching; the intersection judgment label and the structural adjacency matching label are combined as the basis for alignment confirmation.

6. The method according to claim 1, characterized in that, Sheaf consistency diagnosis specifically includes: establishing a base space graph composed of clause nodes, reference nodes, and version nodes; establishing local fibers containing embedded boxes and hyperbolic embedded coordinates on the coverage sets such as clause contexts and interpretive attachments; defining constraint mappings for semantic equivalence merging and attribute projection; pushing fiber content along each loop of the base space graph according to the constraint mappings and comparing them to detect and record consistency differences; and repairing consistency difference records sequentially through semantic equivalence merging, embedded box expansion, or segmentation.

7. The method according to claim 1, characterized in that, The refinement of the obligation slot set and the transformation of the evidence unit set specifically includes: traversing the clause nodes in the clause list, refining the obligation nodes, and generating an obligation slot set containing information on the required set of voucher types and the set of allowed fee categories; accessing event records, image records and text records from the business event flow and evidence warehouse, converting business-side terms into policy-side terms based on the terminology mapping index, and binding records associated with the same event to form an evidence unit set.

8. The method according to claim 1, characterized in that, The process of executing a stable matchmaking process specifically includes: establishing an evidence preference order for obligation slots and establishing an obligation preference order for evidence units; adopting a delayed acceptance process, whereby unmatched obligation slots initiate proposals according to their own preference order until no new proposals are made, thus forming a collection of matchmaking lists.

9. The method according to claim 1, characterized in that, The specific steps for generating action lists and process checkpoint rules include: summarizing the paired items in the collection and matching list to generate a list of required data, and generating a gap list for unmatched slots; and generating three types of process checkpoint rules for each collection batch based on the condition nodes and scope nodes in the clause list, namely, pre-submission verification, inter-node transfer verification, and pre-archiving verification.

10. The method according to claim 1, characterized in that, The solidified clause-evidence-action link specifically includes: generating a link record for each pair containing a clause identifier, evidence number, and action list item number; attaching an embedding box and hyperbolic embedding coordinate snapshot to each link record; and writing the link record to audit storage.

Images