Intelligent voucher automatic generation method and system based on multi-modal large model

Abstract

The invention provides an intelligent voucher automatic generation method and system based on a multi-modal large model. The method comprises the following steps: firstly, obtaining a voucher generation demand text containing voucher use scene description, core element requirements and text expression specifications; performing demand semantic anchoring processing on the voucher generation demand text, and performing semantic association matching with the voucher basic element library; then carrying out large model generation rule adaptation processing, and adjusting a text generation trigger condition of the multi-modal large model; feature mapping processing is carried out according to the adaptation scheme, and corresponding association between the core elements and the generated feature library is established; and finally, performing intelligent voucher text generation based on the adaptation scheme and the feature mapping association set. According to the invention, the voucher can be generated intelligently, accurately and flexibly, the generation efficiency and quality are improved, and the human error risk is reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method and system for automatically generating intelligent vouchers based on a multimodal large model. Background Technology

[0002] In the daily operations and various transactions of an enterprise, vouchers serve as important written proof for recording economic transactions and clarifying economic responsibilities. Their accuracy and standardization directly affect the enterprise's financial management, compliance operations, and subsequent auditing work.

[0003] Traditional voucher generation methods primarily rely on manual operation. Staff need to manually collect relevant information based on specific business scenarios and then fill in the voucher content according to predetermined formats and specifications. This method is not only inefficient but also susceptible to human error, leading to errors or omissions of crucial information. For example, in complex business scenarios, staff may mistakenly record core elements such as amounts and dates due to negligence; or, when faced with a large volume of transactions, they may be unable to complete voucher generation in a timely manner, disrupting the normal flow of business.

[0004] With the development of information technology, although some automated voucher generation tools have emerged, most of these tools are single-function and lack a comprehensive understanding and flexible processing ability for voucher generation needs. They often can only generate vouchers according to fixed templates and cannot be dynamically adjusted according to different voucher usage scenarios, core element requirements, and text expression standards, making it difficult to meet the diverse and personalized voucher generation needs of enterprises. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for automatically generating smart credentials based on a multimodal large model, the method comprising: Obtain the preset voucher generation requirement text, which includes a description of the voucher usage scenario, the requirements for the core elements of the voucher, and the voucher text expression specifications, thus obtaining the voucher generation requirement text; The requirement text for voucher generation is subjected to semantic anchoring processing. Various descriptions in the requirement text for voucher generation are semantically matched with the preset voucher basic element library to obtain the requirement semantic anchoring result. The semantic anchoring results of the demand are adapted to the large model generation rules. The text generation triggering conditions of the multimodal large model are adjusted in combination with the expression specifications of the voucher generation demand text to obtain the large model generation rule adaptation scheme. Based on the large model generation rule adaptation scheme, feature mapping processing is performed on the voucher generation requirement text, and the core elements of the voucher generation requirement text are associated with the generation feature library of the multimodal large model to obtain the voucher generation feature mapping association set. Intelligent voucher text generation is performed based on a large model generation rule adaptation scheme and a voucher generation feature mapping association set. The mapped feature information is input into a multimodal large model according to the adapted generation rules to generate text, and the intelligent voucher text generation result is obtained.

[0006] In another aspect, embodiments of the present invention also provide an intelligent voucher automatic generation system based on a multimodal large model, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code. The processor is used to run the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0007] Based on the above, this embodiment of the invention acquires a preset voucher generation requirement text, comprehensively covering key information such as voucher usage scenario descriptions, core voucher element requirements, and voucher text expression specifications. It performs requirement semantic anchoring processing on the voucher generation requirement text, semantically matching it with a preset voucher basic element library. This allows for accurate understanding of various descriptions in the requirement text. The requirement semantic anchoring results are then adapted to the large-scale model generation rules. Combined with expression specifications, the text generation triggering conditions of the multimodal large-scale model are adjusted, enabling the large-scale model to generate compliant voucher text according to specific rules, enhancing the controllability and standardization of the generation process. Feature mapping processing is performed according to the large-scale model generation rule adaptation scheme, establishing corresponding associations between core elements and the large-scale model's generation feature library, further refining the key information for voucher generation and improving the accuracy of the generation results. Finally, intelligent voucher text generation processing is performed based on the adaptation scheme and feature mapping association set. The mapped feature information is input into the multimodal large-scale model according to the adapted rules, generating high-quality intelligent voucher text that meets actual needs. This achieves intelligent, automated, and precise voucher generation, effectively improving the efficiency and quality of voucher generation and reducing the risk of human error. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the intelligent voucher automatic generation method based on a multimodal large model provided in the embodiments of the present invention.

[0009] Figure 2 This is a schematic diagram of the hardware architecture of the intelligent voucher automatic generation system based on a multimodal large model provided in an embodiment of the present invention. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1It is a schematic flowchart of an intelligent voucher automatic generation method based on a multi-modal large model provided by an embodiment of the present invention. The intelligent voucher automatic generation method based on the multi-modal large model will be introduced in detail below.

[0011] Step S110: Obtain a preset voucher generation requirement text, which includes a voucher usage scenario description, voucher core element requirements, and voucher text expression specifications, to obtain the voucher generation requirement text.

[0012] In this embodiment, the automatic generation of enterprise employee business trip expense reimbursement vouchers is used as a unified application scenario. The voucher generation requirement text is obtained through the requirement submission module of the enterprise internal financial system, and this module supports users to enter requirements in the form of structured forms or natural language texts. The voucher usage scenario description is specifically "the reimbursement of business trip expenses incurred by enterprise employees for participating in domestic industry exhibitions", which clarifies the actor (enterprise employees), activity nature (participating in domestic industry exhibitions), expense type (business trip expenses), and core purpose (reimbursement). The voucher core element requirements include the basic information of the reimburser (name, department, employee number), exhibition information (exhibition name, location, time), business trip itinerary information (departure date, return date, departure city, destination city, transportation mode), expense details information (transportation fee amount and receipt number, accommodation fee amount and receipt number, meal subsidy amount, in-city transportation subsidy amount), approval information (name of department head, name of financial reviewer). The voucher text expression specification requires the use of Chinese written语体, avoiding colloquial vocabulary; the amount needs to be marked with both Arabic numerals and Chinese capital letters; the date format is unified as "YYYY year MM month DD day"; the overall text structure needs to include a title, a main body (expounding the reimburser information, itinerary information, expense details, etc. in modules), and a signature section (reimburser signature column, approver signature column, and date filling column). During the acquisition process, a preliminary verification is performed on the submitted requirement text to check for problems such as missing core elements (such as not filling in the exhibition name), contradictory scenario descriptions (such as simultaneously appearing "domestic" and "foreign" locations), or conflicting expression specifications (such as only marking the amount with Arabic numerals). After passing the verification, the requirement text is converted into structured data in JSON format and stored in the requirement database as the input source for subsequent processing.

[0013] Step S120: Perform requirement semantic anchoring processing on the voucher generation requirement text, semantically associate and match various descriptions in the voucher generation requirement text with a preset voucher basic element library, to obtain a requirement semantic anchoring result.

[0014] In this embodiment, the preset voucher basic element library is a structured knowledge base that contains multiple types of voucher templates and their corresponding elements, and is stored in a graph database. Each element node contains attributes such as element name, element type, data type, and constraint conditions. For the business travel expense reimbursement voucher, the first-level elements in the basic element library include "reimbursement subject element", "event background element", "travel itinerary element", "expense element", and "approval element". Each first-level element contains multiple levels of sub-elements. For example, under the "expense element", there are second-level elements such as "transportation expenses", "accommodation expenses", and "subsidy expenses". Under the "transportation expenses", there are further third-level elements such as "expense amount", "invoice type", and "invoice number". The demand semantic anchoring process first preprocesses the demand text generated for the voucher, including Chinese word segmentation (using the jieba word segmentation tool and loading the custom dictionary in the financial field), stop word removal (such as meaningless words like "of" and "is"), and entity recognition (using the BERT pre-trained model to recognize entities such as "name", "department", and "amount" in the text). Subsequently, the semantic similarity between the preprocessed text segments and the element nodes in the voucher basic element library is calculated. The Word2Vec model is used to convert the text segments and element names into word vectors, and the cosine similarity is calculated to determine the degree of association. The similarity threshold is set to 0.75, and the text segments with a similarity higher than this threshold are associated with the element nodes. For example, "participate in domestic industry exhibitions" in the demand text is associated with the node "event background element - activity nature - industry exhibition" in the basic element library, with a similarity of 0.82; "transportation expense amount" is associated with the node "expense element - transportation expenses - expense amount", with a similarity of 0.91. For text segments with ambiguity (such as "location" may correspond to "departure city" or "destination city"), disambiguation is performed in combination with the context. The unique associated element is determined by analyzing adjacent entities (such as "location" after "departure date" is determined as "departure city"). The final demand semantic anchoring result is a list of association relationships, which contains information such as text segments, associated element paths, similarity scores, and data type mappings, and is stored in XML format.

[0015] Step S130: Perform large model generation rule adaptation processing on the demand semantic anchoring result, and adjust the text generation trigger conditions of the multi-modal large model in combination with the expression specifications of the demand text generated for the voucher, so as to obtain a large model generation rule adaptation scheme.

[0016] In this embodiment, the multi-modal large model adopts the GPT-4 architecture, and its text generation trigger conditions include parameters such as feature call threshold, sentence pattern generation probability, and vocabulary selection weight. The purpose of the large model generation rule adaptation process is to adjust these parameters according to the demand semantic anchoring result and expression specification to generate voucher text that meets the requirements. First, extract the importance weights of each associated element from the demand semantic anchoring result, and adjust the feature call priority according to the key degree of the element in voucher generation (for example, the weight of "expense amount" is 0.9, and the weight of "exhibition name" is 0.8). Then, in combination with the requirement of "Chinese written语体" in the voucher text expression specification, reduce the selection weight of colloquial vocabulary (for example, set the weight of words such as "搞定" and "报一下" to 0), and increase the weight of formal vocabulary (for example, set the weight of words such as "兹" and "鉴于" to 0.8). For the "amount annotation" specification, set the trigger condition for dual amount output. When the "expense amount" element is detected, automatically generate two formats: Arabic numerals and Chinese capital letters. The entire adaptation process is implemented through a rule engine, and the adjusted parameters are encapsulated into a configuration file in JSON format as the large model generation rule adaptation plan.

[0017] Step S131: Extract the scenario association features corresponding to the voucher usage scenario description from the demand semantic anchoring result, and perform feature matching between the scenario association features and the preset scenario generation templates in the multi-modal large model to obtain scenario template matching association information.

[0018] In this embodiment, the voucher usage scenario description is "reimbursement of travel expenses incurred by enterprise employees for participating in domestic industry exhibitions". The scenario association features extracted from the demand semantic anchoring result include four core features: "enterprise employees", "domestic industry exhibitions", "travel expenses", and "reimbursement". The preset scenario generation templates in the multi-modal large model are classified according to the four-tuple of "subject - event - behavior - object", including "employee business trip reimbursement template", "customer entertainment reimbursement template", "purchase expenditure reimbursement template", etc. Each template contains attributes such as template ID, scenario feature vector, text structure template, and fixed expression library. When performing feature matching, convert the scenario association features into feature vectors (using One-Hot encoding, with a dimension of 1000, consistent with the dimension of the template feature vector), and calculate the Euclidean distance between the feature vectors of each scenario generation template. The template with the smallest distance is the matching template. For example, the Euclidean distance between the feature vector of the "employee business trip reimbursement template" and the scenario association feature vector is 0.32, and the distance of the "customer entertainment reimbursement template" is 0.78. Therefore, it is determined that the "employee business trip reimbursement template" is the matching template. The scenario template matching association information includes the matching template ID, feature matching degree (0.85), scenario feature correspondence (such as "domestic industry exhibition" corresponding to "business trip reason - business activity" in the template), template structure description, etc., and is stored in the form of a dictionary.

[0019] Step S1311: Traverse all semantic association information of the requirement semantic anchoring result, identify all feature information that has semantic association with the description of the voucher usage scenario in the voucher generation requirement text, and obtain the original information of the scenario association features.

[0020] In this embodiment, the semantic association information of the demand semantic anchoring results is stored in list form, with each piece of information containing the fields of "text fragment", "related element path", and "similarity". During the traversal, records related to the voucher usage scenario description are filtered by keyword matching (such as "scenario", "purpose", "purpose"). For example, the association information corresponding to text fragments containing "attending domestic industry exhibitions" and "travel expense reimbursement" is obtained. After extracting the above information, duplicate features are merged (such as "exhibition" and "industry exhibition" are merged into "industry exhibition"), and low similarity records (similarity < 0.6) are removed to obtain the original scenario association feature information, which includes preliminary features such as "company employee", "attending exhibition", "domestic", "travel", and "reimbursement".

[0021] Step S1312: Extract features from the original information of scene-related features, remove redundant information that is irrelevant to scene generation, and retain the feature content that can represent the core attributes of the voucher usage scenario to obtain purified scene-related features.

[0022] In this embodiment, redundant information includes features not directly related to scene generation (such as "employee gender" and "number of employees in the department") and features expressed repeatedly (such as "travel" and "travel expenses"). Feature extraction uses the TF-IDF algorithm to calculate the weight of each feature, setting a weight threshold of 0.3, and retaining features with weights higher than the threshold. For example, the TF-IDF weight of "attending domestic industry exhibitions" is 0.85, "travel expenses" is 0.72, and "reimbursement" is 0.68, all of which are retained; the weight of "employee age" is 0.21, which is removed. The purified scene-related features are four core features: "company employees," "domestic industry exhibitions," "travel expenses," and "reimbursement," each with a weight and a source text fragment.

[0023] Step S1313: Retrieve all preset scene generation templates in the multimodal large model, extract the template feature identifier of each scene generation template, and form a set of scene generation template feature identifiers. Each template feature identifier uniquely corresponds to a scene generation template.

[0024] In this embodiment, the scene generation templates for the multimodal large model are stored in the model's template library and managed using a MongoDB database. Each scene generation template includes attributes such as template ID (e.g., "TMP-TRAVEL-001"), template name (e.g., "Employee Business Trip Reimbursement Template"), template feature identifier (a list containing 5-8 core features, such as ["Employee", "Business Trip", "Expense", "Reimbursement"]), and applicable scope (e.g., "Domestic Business Trip", "International Business Trip"). After retrieving all templates, the template feature identifier of each template is extracted, and duplicate identifiers are removed (if multiple template feature identifiers are completely identical, the template with the more specific applicable scope is retained), forming a scene generation template feature identifier set. This set is a list, and each element is a dictionary containing a template ID and a list of feature identifiers.

[0025] Step S1314: Perform semantic feature comparison between the purified scene association features and each template feature identifier in the scene generation template feature identifier set, calculate the semantic association degree between the features, and obtain the semantic association degree comparison result.

[0026] In this embodiment, semantic feature comparison uses the Word2Vec model to convert the purified scene association features and template feature identifiers into 300-dimensional word vectors. Then, the cosine similarity between the two sets of vectors is calculated as the degree of semantic association. For example, the cosine similarity between the purified scene association feature vector and the feature identifier vector of "employee business trip expense reimbursement template" is 0.85, and the similarity with "purchase expense reimbursement template" is 0.42. The semantic association degree comparison result is a dictionary list, where each dictionary contains a template ID, a list of feature identifiers, and an association degree value (cosine similarity), arranged in descending order of association degree value.

[0027] Step S1315: Based on the semantic association degree comparison results, select the scene generation template feature identifiers with the highest semantic association degree with the purified scene association features to obtain candidate scene generation template feature identifiers.

[0028] In this embodiment, a correlation threshold of 0.7 is set, and template feature identifiers with correlation values ​​higher than this threshold are selected from the semantic correlation comparison results. If multiple templates meet the criteria (such as "Employee Business Trip Reimbursement Template" and "Business Travel Reimbursement Template"), the quantitative matching degree of the feature identifiers is further compared (e.g., if there are 4 scene-related features, and the template feature identifier contains 3 of them, the quantitative matching degree is 0.75), and the template with the highest quantitative matching degree is selected. Finally, the feature identifier of "Employee Business Trip Reimbursement Template" is determined as the candidate scene generation template feature identifier, with a correlation value of 0.85 and a quantitative matching degree of 1.0.

[0029] Step S1316: Retrieve the scenario generation template corresponding to the candidate scenario generation template feature identifier, and extract the template structure feature, template wording feature, and template element arrangement feature of this scenario generation template to obtain the core features of the candidate scenario generation template.

[0030] In this embodiment, the candidate scenario generation template is a structured file in XML format. The template structure features include document type definition (DTD), tag hierarchy (such as <Reimbursement Form><Reimburser Information><Name>< / Name>< / Reimburser Information>< / Reimbursement Form>), and paragraph division (each tag corresponds to a paragraph or sub-paragraph); the template wording features include fixed phrases (such as "There is an employee of our company", "Total RMB"), sentence structure (such as "[Element Name]: [Element Value]"), and professional terms (such as "Number of Bills", "Amount in Capital Letters"); the template element arrangement features include element order (such as "Reimburser Information" first and then "Trip Information"), logical relationship between elements (such as "Departure Date" needs to be earlier than "Return Date"), and element display priority (such as "Expense Amount" has a high priority and "Remarks" has a low priority). By parsing the XML file, extract the above features and convert them into a dictionary form to obtain the core features of the candidate scenario generation template.

[0031] Step S1317: Perform feature matching on a dimension-by-dimension basis between the purified scenario association features and the core features of the candidate scenario generation template, record the matching results and matching details for each dimension, and obtain a dimensioned matching result record.

[0032] In this embodiment, the dimension-by-dimension matching includes the structure dimension, the wording dimension, and the arrangement dimension. Structure dimension matching checks whether there is a corresponding node in the template structure for the scenario association feature, such as whether "Business Trip Expenses" corresponds to the <Expense Information><Business Trip Expenses> node in the template; wording dimension matching checks whether the terms in the scenario description are consistent with the template wording, such as whether the expression of "Domestic Industry Exhibition" is compatible with "Business Trip Reason - Business Exhibition" in the template; arrangement dimension matching checks whether the logical order of the scenario association elements conforms to the template element arrangement, such as whether "Reimburser" is before "Expense Details". The matching details include the feature points that match successfully (such as the matching of "Business Trip Expenses" with the template structure), the feature points that partially match (such as the semantic relevance of the different terms "Domestic Industry Exhibition" and "Business Exhibition"), and the feature points that do not match (such as the "Urgency" element included in the template but not mentioned in the scenario). The dimensioned matching result record is a two-dimensional list, where the rows represent dimensions, the columns represent matching items, and the values are the matching status (success / partial success / failure) and matching description.

[0033] Step S1318: Record the dimensional matching results, calculate the average matching degree of each dimension's matching results, and use it as the overall matching degree between the purified scene association features and the candidate scene generation template. Record the specific feature points that are successfully matched in each dimension to obtain the overall matching evaluation result.

[0034] In this embodiment, the matching degree for each dimension is calculated as (number of successful feature points + 0.5 × number of partially successful feature points) / total number of feature points. The structure dimension contains 5 feature points, with 4 successful and 1 partially successful, resulting in a matching degree of (4 + 0.5 × 1) / 5 = 0.9; the wording dimension contains 6 feature points, with 3 successful and 2 partially successful, resulting in a matching degree of (3 + 0.5 × 2) / 6 = 0.67; the layout dimension contains 4 feature points, all of which are successful, resulting in a matching degree of 1.0. The overall matching degree is (0.9 + 0.67 + 1.0) / 3 = 0.86. The overall matching evaluation result includes the overall matching degree (0.86), the matching degree for each dimension (structure 0.9, wording 0.67, layout 1.0), and a list of successfully matched feature points (such as "reimbursement recipient information structure" and "travel date order"), stored in JSON format.

[0035] Step S1319: Based on the overall matching evaluation results, mark the feature parts that match the candidate scene generation template and the feature parts that do not match in the scene association features to obtain the matching feature annotation information.

[0036] In this embodiment, the matching feature annotation information uses HTML tags to mark the scene-related feature text, and the matching part uses... <match>Tags (such as) <match> Travel expenses< / match> (Partial matching) <partial>Tags (such as) <partial> Domestic industry exhibitions< / partial> ), not matched <unmatch>Tags (such as) <unmatch> Emergency reimbursement< / unmatch> (If the element exists in the scene but the template does not have a corresponding element). Also, for <partial>Adding a title attribute to the tag to specify the matching differences (e.g., title="The corresponding element in the template is 'Business Exhibition'") will help. <unmatch>Tag addition suggestions (e.g., "Delete this feature or add template elements"). Matching feature annotation information is stored as a string for subsequent template adjustments.

[0037] Step S13110: Integrate the core features of the candidate scene generation template, the dimensional matching result record, the overall matching evaluation result and the matching feature annotation information to form scene template matching association information, which contains all relevant details of scene matching.

[0038] In this embodiment, the scene template matching association information is a nested dictionary. The top-level keys include "template_core_features" (candidate template core features), "dimension_matching_results" (dimension-level matching result records), "overall_matching_evaluation" (overall matching evaluation results), and "feature_annotation" (matching feature annotation information). The values ​​corresponding to each key are the specific data generated in the preceding steps. For example, the value of "template_core_features" is a dictionary containing structural, wording, and layout features, while the value of "overall_matching_evaluation" is a dictionary containing the overall matching degree and the matching degree of each dimension. This association information is used to guide the adjustment of subsequent generation rules.

[0039] Step S132: Extract the core elements of the voucher in the voucher generation requirement text and match them one by one with the basic elements in the multimodal large model's generation feature library to obtain the element feature correspondence information.

[0040] In this embodiment, the core elements requirements of vouchers are extracted from the "core elements" field of the requirement text, including 15 elements such as "reimburser's name", "department", "exhibition name", "departure date", "transportation expense amount", etc. The element features are converted into 768-dimensional feature vectors through a feature extraction function (the feature extraction layer based on the BERT model). The generated feature library of the multimodal large model is a vector database (using FAISS index) that stores the feature vectors of various basic elements of vouchers, including categories such as "personnel elements", "time elements", "amount elements", "event elements", etc. Each basic element feature contains attributes such as feature ID, feature vector, element description, data type, etc. During the one-to-one correspondence process, the cosine similarity is calculated between each element feature vector and the basic element feature vectors of the same category in the generated feature library. The basic element feature with the highest similarity and exceeding the threshold (0.8) is the matching item. For example, the element feature of "transportation expense amount" matches the basic feature of "expense element - transportation expense - amount" in the generated feature library, and the similarity is 0.92. The element feature corresponding association information includes element name, feature ID, matching basic element name, similarity score, data type mapping (such as the data type "string" of the requirement element "departure date" is mapped to the "date type" of the basic element), etc., and is stored in CSV format.

[0041] Step S133: Extract the expression features corresponding to the voucher text expression specification in the voucher generation requirement text, and disassemble the expression features into text sentence pattern features, text diction features, and text length features to obtain the disassembled result of the expression features.

[0042] In this embodiment, the original description of the voucher text expression specification is "Use formal written language and avoid colloquial expressions; the amount should be marked with both Arabic numerals and Chinese capital letters; the date format is YYYY year MM month DD day; the text length is controlled within one page of A4 paper (about 800 - 1200 words); write in chapters, chapter titles are in bold, and the first line of each paragraph is indented by 2 characters". The disassembling of the expression features adopts a rule-based feature extraction method: the text sentence pattern features include sentence pattern type (the proportion of declarative sentences ≥ 90%, imperative sentences ≤ 5%), sentence length distribution (the average sentence length is 15 - 25 words, and the longest sentence does not exceed 40 words), punctuation usage specification (such as adding the word "whole" after the amount); the text diction features include语体风格 (formal degree score ≥ 0.85, based on a pre-trained语体分类模型), vocabulary selection (using written words such as "hereby", "in view of", etc., and disabling colloquial words such as "get it done", "almost", etc.), term accuracy (the accuracy rate of financial terms is 100%); the text length features include total number of words (800 - 1200 words), proportion of chapter word count ("expense details" chapter ≥ 40%), line spacing (1.5 times), and page margins (2.5 cm on the top, 2 cm on the bottom, 3 cm on the left, 2 cm on the right). The disassembled result is represented in a hierarchical JSON structure, and each feature category contains feature items, quantitative indicators, verification rules, etc.

[0043] Step S134: Combine the scene template matching association information and the element feature corresponding association information to determine the feature call priority when the multimodal large model generates voucher text, and obtain the feature call priority ranking result.

[0044] In this embodiment, feature call priority is calculated based on three dimensions: feature importance, scenario relevance, and element completeness. Importance weights are assigned by financial experts (e.g., "Expense Amount" weight 0.9, "Remarks" weight 0.3); scenario relevance is determined by the overall matching evaluation results in the scenario template matching associated information (elements with higher matching degrees have higher relevance scores); element completeness is calculated based on the proportion of non-empty values ​​in the associated information corresponding to the element feature (e.g., "Transportation Expense Amount" and "Invoice Number" are both non-empty, resulting in a completeness score of 1.0). The priority calculation formula is: Priority Score = 0.4 × Importance Weight + 0.3 × Scenario Relevance Score + 0.3 × Element Completeness Score. All element features are sorted in descending order of priority score to obtain the feature call priority ranking result, for example: Reimbursement Employee Name (0.92) > Transportation Expense Amount (0.88) > Departure Date (0.85) > Exhibition Name (0.80) > Department (0.78) > ... (subsequent elements are arranged in order of score).

[0045] Step S135: Based on the text sentence features in the expression feature decomposition results, adjust the sentence construction trigger logic when the multimodal large model generates text to obtain the sentence construction trigger logic adjustment result.

[0046] In this embodiment, the sentence construction module of the multimodal large model includes a sentence library (storing 200+ sentence templates) and triggering rules (logic for selecting sentences based on context). The text sentence feature requires declarative sentences to be the primary form; therefore, the triggering rules are adjusted: the trigger probability of declarative sentences is increased from the default 60% to 85%, imperative sentences are decreased from 20% to 5%, and interrogative and exclamatory sentences are disabled (0% trigger probability). For the "amount labeling" specification, a new sentence triggering condition is added: when a "amount element" is detected, the compound sentence "[Arabic numerals] yuan (uppercase: [Chinese uppercase])" is automatically triggered. The adjusted sentence construction triggering logic results are stored in an XML-formatted rule file, containing parameters such as sentence type, triggering condition (e.g., "element type = amount"), trigger probability, and sentence template ID.

[0047] Step S136: Based on the text wording features in the expression feature decomposition results, adjust the word selection trigger logic when the multimodal large model generates text to obtain the word selection trigger logic adjustment result.

[0048] In this embodiment, the vocabulary selection trigger logic adjustment includes: constructing a formal written vocabulary list (containing 5000+ formal financial terms), replacing colloquial terms in the model's default vocabulary list (e.g., replacing "go" with "go to", and "stay" with "stay"); establishing a terminology mapping table to ensure the consistent use of terms such as "invoice number" and "amount in words"; and setting a list of prohibited words (e.g., "probably", "possibly", "more or less") to filter prohibited words during the vocabulary selection stage. This adjustment is achieved by modifying the model's vocabulary weight matrix, assigning high weights (0.9-1.0) to formal vocabulary and terminology, low weights (0.1-0.3) to colloquial vocabulary, and zero weights to prohibited words. The vocabulary selection trigger logic adjustment results include updated vocabulary weight matrix values, a terminology mapping table, and a list of prohibited words, stored in the model parameter directory in binary file format.

[0049] Step S137: Based on the text length features in the expression feature decomposition results, adjust the content expansion trigger logic when the multimodal large model generates text to obtain the content expansion trigger logic adjustment results.

[0050] In this embodiment, the content expansion triggering logic adjustment is centered on length control: a dynamic monitoring variable for the total word count is set, initially set to 0, and incremented after each sentence is generated. When the word count approaches the lower limit (800 words), content expansion is triggered (e.g., supplementing the expense details); when it approaches the upper limit (1200 words), content simplification is triggered (e.g., deleting repetitive statements). The word count generated by each module is allocated according to the proportion of word count in each chapter, such as allocating 40%-50% of the total word count to the "Expense Details" chapter and 10%-15% to the "Approval Information" chapter. Line spacing and margin parameters are set, and format control is achieved by modifying the CSS style of the generated text. The content expansion triggering logic adjustment result includes length monitoring parameters (upper and lower limit thresholds, chapter proportion), expansion / simplification trigger conditions (e.g., "word count < 800 and chapter proportion < 40%" triggers expansion), and format control parameters (line spacing, margin), stored in JSON format.

[0051] Step S138: Integrate the result of feature call priority sorting, the result of sentence construction triggering logic adjustment, the result of vocabulary selection triggering logic adjustment, and the result of content expansion triggering logic adjustment to form a preliminary generation rule adjustment scheme for a multimodal large model.

[0052] Step S1381: Use the feature call priority ranking result as the core framework of the preliminary rule adjustment scheme, determine the core feature call order when the multimodal large model generates voucher text, and build a preliminary rule framework structure.

[0053] In this embodiment, the core framework adopts a Directed Acyclic Graph (DAG) structure, where nodes represent features and directed edges represent the invocation order. Starting with "Reimbursement Provider Name" in the feature invocation priority ranking results, "Transportation Expense Amount" and "Departure Date" become subsequent nodes, and so on. Each node contains attributes such as feature ID, priority score, and association generation logic (sentence structure, vocabulary, expansion). The initial rule framework structure is generated using a DAG visualization tool to ensure that the feature invocation order conforms to the priority ranking.

[0054] Step S1382: Decompose the sentence construction trigger logic adjustment result into multiple sentence construction sub-logics. According to the feature calling order of the preliminary rule framework structure, embed each sentence construction sub-logic into the corresponding position of the core framework to obtain the rule framework with embedded sentence logic.

[0055] In this embodiment, the sentence structure construction trigger logic adjustment result is broken down into "personnel information sentence structure sub-logic," "amount expression sentence structure sub-logic," and "date expression sentence structure sub-logic," etc. Each sub-logic contains a trigger condition and a sentence template. For example, the trigger condition for the "personnel information sentence structure sub-logic" is "feature type = personnel element," and the sentence template is "[name] is an employee of [department]." According to the feature calling order, the "personnel information sentence structure sub-logic" is embedded into personnel element nodes such as "reimbursement recipient's name" and "department," and the "amount expression sentence structure sub-logic" is embedded into amount element nodes such as "transportation expense amount" and "accommodation expense amount," forming a rule framework for embedded sentence logic.

[0056] Step S1383: Decompose the word selection trigger logic adjustment result into multiple word selection sub-logics. According to the feature call order of the rule framework embedded in the sentence logic, embed each word selection sub-logic into the corresponding position of the framework, forming a logical correspondence with the sentence construction sub-logic, and obtain a rule framework with embedded dual logic.

[0057] In this embodiment, the vocabulary selection trigger logic adjustment result is broken down into "personnel title vocabulary sub-logic," "amount term vocabulary sub-logic," and "date expression vocabulary sub-logic," etc. The "personnel title vocabulary sub-logic" specifies that personnel elements use titles such as "comrade" or "Mr / Ms," embedding them in the "reimbursement applicant's name" node, corresponding to the "personnel information sentence structure sub-logic." The "amount term vocabulary sub-logic" specifies that amount elements use terms such as "yuan" or "capitalized," embedding them in the "transportation expense amount" node, corresponding to the "amount expression sentence structure sub-logic." The vocabulary selection sub-logic and sentence structure construction sub-logic are bound together through a logic chain ID, ensuring that both types of logic are triggered simultaneously when generating the corresponding feature text, forming a rule framework with embedded dual logic.

[0058] Step S1384: Decompose the content expansion trigger logic adjustment result into multiple content expansion sub-logics. According to the feature call order of the rule framework with embedded dual logic, embed each content expansion sub-logic into the corresponding position of the framework, forming a triple logic correspondence with the sentence construction sub-logic and the vocabulary selection sub-logic, thus obtaining a rule framework with embedded triple logic.

[0059] In this embodiment, the content expansion trigger logic adjustment result is broken down into "Expense Details Expansion Sub-Logic," "Itinerary Description Expansion Sub-Logic," and "Exhibition Information Expansion Sub-Logic," etc. The "Expense Details Expansion Sub-Logic" is triggered when the expense element has insufficient characters, supplementing information such as "Ticket Type" and "Expense Composition," and embedding it into the "Transportation Expenses Amount" node, corresponding to the sentence structure and vocabulary sub-logic of that node. The "Itinerary Description Expansion Sub-Logic" is triggered when the itinerary element information is brief, supplementing information such as "Transportation Mode Details" and "Accommodation Location," and embedding it into the "Departure Date" node. These three logics are linked through a logic chain ID, forming a rule framework that embeds the three logics, ensuring that sentence structure, vocabulary, and expansion logic work synergistically during feature generation.

[0060] Step S1385: Extract all logical adjustment information from the rule framework embedded with triple logic, classify and organize it according to feature type, and obtain a set of classified and organized logical adjustment information.

[0061] In this embodiment, the feature types include "personnel features," "time features," "amount features," "event features," and "approval features." All nodes in the rule framework are traversed, and the sentence structure, vocabulary, and extended sub-logic of each node are grouped according to feature type. For example, the logical adjustment information for "reimbursement applicant's name" and "department" is grouped into the "personnel features" category, while the logical adjustment information for "departure date" and "return date" is grouped into the "time features" category. The categorized and organized logical adjustment information set is a dictionary, where the key is the feature type and the value is a list of all logical adjustment information under that type.

[0062] Step S1386: Traverse the sorted and organized logical adjustment information set. For the same feature call position in the preliminary rule framework structure, check whether there are sentence construction sub-logic, vocabulary selection sub-logic, and content expansion sub-logic. If they exist, arrange the above sub-logic at that position in the order of sentence construction, vocabulary selection, and content expansion to obtain a logically coherent adjustment information set.

[0063] In this embodiment, for the same feature call location, such as the "Transportation Cost Amount" node, there are "Amount Expression Sentence Sub-Logic," "Amount Terminology Vocabulary Sub-Logic," and "Fee Details Expansion Sub-Logic." After checking and confirmation, the sub-logics are arranged in the order of "Sentence Structure → Vocabulary → Expansion" to ensure that the sentence structure is determined first during generation, then appropriate vocabulary is selected, and finally the content is expanded as needed. For example, the sentence structure "Transportation Cost Amount: [Arabic Numerals] Yuan (Uppercase: [Chinese Uppercase])" is generated first, then replaced with formal terminology (such as "RMB" replacing "money"), and finally, if the number of characters is insufficient, it is expanded to "(Ticket Type: High-speed Rail Ticket, Ticket Number: JT20231001)". The logically coherent adjustment information set generates an ordered list of sub-logics for each feature call location.

[0064] Step S1387: Add a corresponding feature call identifier to each adjustment logic in the logically coherent adjustment information set, so that each adjustment logic can accurately correspond to a specific feature call position in the core framework, and obtain an adjustment information set with feature identifiers.

[0065] In this embodiment, the feature call identifier adopts the format "feature type-feature ID-logic type", such as "amount feature-F005-sentence structure" to represent the sentence adjustment logic of the element with ID F005 (transportation cost amount) in the amount feature. By embedding this identifier in each adjustment logic, a mapping relationship between the logic and the core framework nodes is established to ensure that the model can call the corresponding adjustment logic according to the currently processed feature when it is generated. Each record in the adjustment information set with feature identifiers contains the feature call identifier, sub-logic content, and triggering condition.

[0066] Step S1388: Match the set of adjustment information with feature identifiers with the feature call priority sorting results again to confirm that the adjustment logic of each feature call position is complete and without omission, and obtain the set of adjustment information after matching confirmation.

[0067] In this embodiment, the matching is achieved by comparing the feature list in the feature call priority ranking result with the feature call identifier in the adjustment information set with feature identifiers. For example, the "Exhibition Name" feature in the priority list must correspond to three adjustment logics: "Event Feature-F012-Sentence Structure", "Event Feature-F012-Vocabulary", and "Event Feature-F012-Extension". If a feature is found to be missing a certain type of logic (such as the "Remarks" feature lacking extension logic), it is marked as "Logic Missing" and a default logic is added (such as "Remarks feature does not extend by default"). The adjustment information set after matching confirmation ensures that the sentence structure, vocabulary, and extension logic of each feature call position are complete.

[0068] Step S1389: Integrate the set of adjustment information after matching confirmation, sort out the triggering conditions and execution requirements of all adjustment logic according to the feature calling order, and obtain the sorted generation rule adjustment information.

[0069] In this embodiment, the integration process extracts the adjustment logic for each feature call location in descending order of feature call priority, clarifying its triggering conditions (e.g., "trigger expansion when amount > 1000 yuan"), execution steps (e.g., "1. Apply sentence template; 2. Replace formal vocabulary; 3. Supplement invoice information"), and output requirements (e.g., "generated text must include capitalized amount"). The results are organized into a linear list, with each element representing the adjustment logic details for a feature call location, including feature identifier, triggering conditions, execution steps, and output requirements fields.

[0070] Step S13810: Standardize and organize the sorted generation rule adjustment information according to the generation rule format of the multimodal large model to form a preliminary generation rule adjustment scheme for the multimodal large model that meets the configuration requirements of the large model.

[0071] In this embodiment, the generation rule format for the multimodal large model is a specific Protocol Buffers (PB) format, containing fields such as "rule_set", "feature_config", and "logic_flow". Normalization maps the refined generation rule adjustment information to the PB format. For example, "execution steps" are converted into a sequence of function calls executable by the model (e.g., "call(sentence_template.apply), call(vocab_replace), call(content_expand)"), and "trigger conditions" are converted into Boolean expressions parsable by the model (e.g., "amount>1000&&ticket_type=='high-speed rail ticket'"). The final generated preliminary generation rule adjustment scheme is a PB file, which can be loaded through the model's configuration interface.

[0072] Step S139: Perform a secondary association between the initial generated rule adjustment scheme and the demand semantic anchoring result, supplement the generated rule details corresponding to the scene association features, and obtain the supplemented generated rule adjustment scheme.

[0073] In this embodiment, the secondary association retrieves missing scenario-specific rules from the initial generation rule adjustment scheme using scenario-related features (such as "domestic industry exhibitions" and "travel expenses") in the demand semantic anchoring results. For example, for the "domestic industry exhibitions" scenario, a generation rule "exhibition information must include the venue and time" is added; for the "travel expenses" scenario, a rule "transportation expenses must distinguish between urban and intercity transportation" is added. During the addition process, scenario-related features are used as keywords to match relevant rules in a preset scenario rule library (which stores special generation rules for different scenarios) and embed them into the corresponding positions in the initial generation rule adjustment scheme (such as the expansion logic of the "exhibition name" feature). The supplemented generation rule adjustment scheme includes basic rules and scenario-specific rules to enhance the scenario adaptability of the generated text.

[0074] Step S1310: Integrate all feature adjustment information of the supplemented generation rule adjustment scheme to form a multimodal large model generation rule adaptation scheme. The large model generation rule adaptation scheme includes all adjusted trigger logic and feature call requirements.

[0075] In this embodiment, the integration process performs deduplication (removing duplicate rule entries), conflict detection (e.g., if the "amount in capital letters" rule conflicts with the "amount in abbreviation" rule, the former is retained), and priority sorting (scenario-specific rules have higher priority than basic rules) on the supplemented generation rule adjustment scheme. The final large model generation rule adaptation scheme is a complete configuration file containing all adjusted triggering logic (sentence structure, vocabulary, extensions), feature call requirements (order, priority, data type), and scenario-specific rules. It is stored in ZIP compressed package and contains PB format rule files, vocabulary, sentence template library, and other resources.

[0076] Step S140: Perform feature mapping processing on the voucher generation requirement text according to the large model generation rule adaptation scheme, establish corresponding associations between the core elements of the voucher generation requirement text and the generation feature library of the multimodal large model, and obtain the voucher generation feature mapping association set.

[0077] Step S141: Extract the feature call priority ranking result from the large model generation rule adaptation scheme, and reorder the core elements of the voucher generation requirement text according to the feature call priority ranking result to obtain the core element priority ranking result.

[0078] In this embodiment, the feature call priority sorting result in the large model generation rule adaptation scheme is [Reimbursement applicant's name, transportation expense amount, departure date, exhibition name, department, ...]. The original order of the core elements in the voucher generation requirement text may be [Department, Reimbursement applicant's name, exhibition name, departure date, ...], which needs to be rearranged according to the priority sorting result to obtain the core element priority sorting result: [Reimbursement applicant's name, transportation expense amount, departure date, exhibition name, department, ...] (consistent with the feature call priority order). The re-sorting is implemented using Python's sorted function, which maps and sorts the elements using the priority sorting result as the key.

[0079] Step S142: Extract all basic features from the generated feature library of the multimodal large model, classify and organize the basic features according to the voucher element type, and obtain the classification and organization results of the basic features.

[0080] In this embodiment, the feature library for generating the multimodal large model contains over 1000 basic features, stored in the "features" table of a PostgreSQL database, including fields such as "feature_id", "feature_name", "element_type", "data_type", and "description". These features are categorized according to voucher element type, specifically based on the value of the "element_type" field, into five categories: "Personnel Information Elements", "Travel Information Elements", "Expense Information Elements", "Event Information Elements", and "Approval Information Elements". For example, "feature_id=1001, feature_name=Reimbursement Person's Name, element_type=Personnel Information Element" or "feature_id=2003, feature_name=Departure Date, element_type=Travel Information Element". The categorization results are stored as a dictionary, with the key being the element type and the value being a list of all basic features under that type. Each basic feature contains information such as feature_id, feature_name, and data_type.

[0081] Step S143: Compare the features of each core element in the priority ranking result with the corresponding voucher element type in the basic feature classification result, and select the basic features that match the core elements to obtain the preliminary screening results of basic features.

[0082] Step S1431: Extract the first core element from the core element priority ranking result, determine the element type and core expression content of the core element, and obtain the basic information of a single core element.

[0083] In this embodiment, the first item in the priority ranking of core elements is "Reimbursement Applicant's Name". By analyzing the description of this element in the requirement text ("Reimbursement Applicant's Name: Zhang San") and its context, its element type is determined to be "Personnel Information Element", and its core content is "the legal name used to identify the subject of travel expense reimbursement". The basic information of a single core element is stored in dictionary form: {"element_name":"Reimbursement Applicant's Name", "element_type":"Personnel Information Element", "core_content":"The legal name used to identify the subject of travel expense reimbursement"}.

[0084] Step S1432: In the results of basic feature classification, find the basic feature classification that matches the feature type of the core element, extract all basic features under the basic feature classification, and obtain a set of basic features of the same type.

[0085] In this embodiment, the core element's element type is "personnel information element". The corresponding basic feature category is found in the basic feature classification results, and all basic features are extracted, such as "name of the person requesting reimbursement", "department", "employee number", and "contact number", forming a set of basic features of the same type. This set is in list form, and each element is a dictionary containing feature_id, feature_name, and data_type.

[0086] Step S1433: Compare the semantic features of the core expression content of a single core element with each basic feature in the same type of basic feature set, identify the basic features that have semantic overlap with the core expression content of the element, and obtain a candidate set of semantically overlapping features.

[0087] In this embodiment, semantic feature comparison uses the BERT model to calculate the semantic similarity between the core content of an element and the description of its basic features. The "legal name used to identify the entity responsible for travel expense reimbursement" and the "description" field of each basic feature in the same type of basic feature set (e.g., "Reimbursement applicant's name: The name of the entity responsible for reimbursement on the travel expense reimbursement form, which must match the ID card) are input into the BERT model, and a similarity score is output. A similarity threshold of 0.7 is set; basic features exceeding this threshold are considered semantically overlapping features. For example, the similarity score of the basic feature "Reimbursement applicant's name" is 0.93, while "Employee ID" is 0.45. Therefore, "Reimbursement applicant's name" is selected into the candidate set of semantically overlapping features.

[0088] Step S1434: Perform feature integrity analysis on each basic feature in the semantically overlapping feature candidate set to confirm whether the basic feature can fully represent the core expression content of the core element and obtain the feature integrity analysis result.

[0089] In this embodiment, feature integrity analysis is performed from three dimensions: data, constraints, and relationships. The data dimension checks whether the data type of the basic feature covers the requirements of the core element (e.g., "reimbursement recipient's name" must be a string with a length of 2-50 characters). The constraints dimension checks whether the constraints of the basic feature meet the integrity requirements of the core element (e.g., non-empty constraints, uniqueness constraints). The relationships dimension checks whether the basic feature has a necessary relationship with other related elements (e.g., "reimbursement recipient's name" must be associated with "employee ID" to ensure uniqueness). The analysis results are presented in a scoring format (0-1.0). For example, if the integrity score of the basic feature "reimbursement recipient's name" is 0.95 (data dimension 0.9, constraint dimension 1.0, relationships dimension 1.0), it indicates that it can fully represent the core element.

[0090] Step S1435: Based on the feature integrity analysis results, select the basic features that can fully represent the core elements and obtain the matching basic features of a single core element.

[0091] In this embodiment, the integrity score threshold is set to 0.85. The score of the basic feature "reimbursement recipient's name" in the semantic overlap feature candidate set is 0.95, which is higher than the threshold. Therefore, it is determined as the matching basic feature of a single core element. If there are multiple basic features that meet the conditions (such as "reimbursement recipient's name" and "applicant's name"), further filtering is performed through precise matching of feature names to select the basic features with completely identical names.

[0092] Step S1436: Associate the basic matching features of a single core element with the core element, annotate the corresponding element type and matching details, and obtain the feature matching record of the single element.

[0093] In this embodiment, the associated record includes the core element name (name of the person seeking reimbursement), the matching basic feature ID (feature_id=1001), the element type (personnel information element), the semantic similarity score (0.93), the completeness score (0.95), and the matching details (such as "the basic feature description is highly consistent with the content expressed by the core element"). The feature matching record of a single element is a dictionary for subsequent integration.

[0094] Step S1437: Extract the next core element from the core element priority ranking result, and repeat the operations of determining the basic information of the element, finding the set of basic features of the same type, semantic feature comparison, feature integrity analysis, and filtering matching basic features and related records to obtain the feature matching record set of all core elements.

[0095] In this embodiment, steps S1431 to S1436 are executed sequentially for the second item "Transportation Cost Amount" and the third item "Departure Date" in the priority ranking result of the core elements. For example, the element type of "Transportation Cost Amount" is "Cost Information Element". The basic feature set of the same type includes "Transportation Cost Amount" and "Accommodation Cost Amount", etc. After semantic comparison, the basic feature of "Transportation Cost Amount" (feature_id=3002) is matched, and the completeness score is 0.90, forming the corresponding feature matching record. Finally, a set containing matching records of all core elements is obtained.

[0096] Step S1438: Compare the core elements in the feature matching record set of all core elements with the core element priority ranking result, and check whether each core element in the core element priority ranking result exists in the feature matching record set to obtain the matching record check result.

[0097] In this embodiment, the core element priority ranking result includes 15 core elements, and the feature matching record set also includes 15 records. By comparing the core element names one by one, it is confirmed that each core element has a corresponding matching record, and the matching record check result is "complete match". If there is a missing element (such as a core element not finding a matching basic feature), the check result is "partially missing", and a manual intervention process is triggered (such as the financial personnel manually specifying the matching basic feature).

[0098] Step S1439: Based on the matching record check results, extract the basic matching features from all feature matching records, arrange them in the order of the core element priority sorting results, and obtain an ordered set of basic matching features.

[0099] In this embodiment, the matching record check result is "complete match". Therefore, the feature_id of the matching basic features is extracted in the order of the results sorted by the priority of the core elements, resulting in an ordered set of matching basic features: [1001, 3002, 2001, 4001, 1002, ...] (corresponding to the basic feature IDs of elements such as the name of the person claiming reimbursement, the amount of transportation expenses, the departure date, the name of the exhibition, and the department to which the person belongs).

[0100] Step S14310: Organize the ordered set of basic matching features as a whole, remove the basic features that appear repeatedly, and retain the unique basic matching feature corresponding to each core element to form the preliminary screening result of basic features.

[0101] In this embodiment, each basic feature ID in the ordered set of basic matching features is unique (because each core element corresponds to a different basic feature), so there is no need to remove duplicates; the ordered set is directly used as the initial screening result of the basic features. If duplicates exist (such as two core elements matching the same basic feature ID), then the matching items with higher scores are retained by comparing semantic similarity scores and completeness scores.

[0102] Step S144: Combine the sentence structure construction trigger logic adjustment results in the large model generation rule adaptation scheme to add sentence structure adaptation tags to the basic features after preliminary screening, and obtain the basic feature set with sentence structure adaptation tags.

[0103] In this embodiment, the sentence structure construction triggering logic adjustment result in the large model generation rule adaptation scheme includes the sentence structure type corresponding to each basic feature (such as "personnel information sentence structure" and "amount expression sentence structure"). Each basic feature ID in the preliminary basic feature screening result is traversed, and the corresponding sentence structure type is found in the sentence structure construction triggering logic adjustment result as a sentence structure adaptation label. For example, basic feature ID=1001 (reimbursement recipient's name) corresponds to the "personnel information sentence structure" label, and ID=3002 (transportation expense amount) corresponds to the "amount expression sentence structure" label. A "syntax_label" field is added to each basic feature to store the label, forming a basic feature set with sentence structure adaptation labels, such as [{"feature_id":1001, "syntax_label":"personnel information sentence structure"}, ...].

[0104] Step S145: Combine the word selection trigger logic adjustment results in the large model generation rule adaptation scheme, add word adaptation labels to each feature in the basic feature set with sentence adaptation labels, and obtain the basic feature set with dual adaptation labels.

[0105] In this embodiment, the vocabulary selection trigger logic adjustment result includes the vocabulary style corresponding to each basic feature (such as "formal title" and "monetary terminology"). Similar to step S144, a "lexical_label" field is added to each basic feature in the basic feature set with sentence structure adaptation labels to store the corresponding wording adaptation label. For example, the basic feature of the claimant's name is labeled with "formal title", and the basic feature of transportation expense amount is labeled with "monetary terminology". The basic feature set with dual adaptation labels includes the feature_id, syntax_label, and lexical_label fields.

[0106] Step S146: Match the set of basic features with dual adaptation tags one-to-one with the core elements of the voucher generation requirement text, record the basic features and adaptation tag information corresponding to each core element, and obtain the element feature correspondence recording results.

[0107] Step S1461: Extract the original list of core elements for generating the required text of the voucher, and rearrange the original list according to the priority of the core elements to obtain an ordered list of core elements.

[0108] In this embodiment, the original list of core elements for the voucher generation requirement text is an unordered list (such as a JSON array) input by the user. By matching the element names with the core element priority sorting result, it is reordered into an ordered list of core elements: ["Reimbursement applicant's name", "Transportation expense amount", "Departure date", "Exhibition name", "Department", ...].

[0109] Step S1462: Extract the first basic feature from the basic feature set with dual adaptation labels, extract the sentence structure adaptation label and vocabulary selection adaptation label corresponding to the basic feature, and obtain the adaptation label information of a single basic feature.

[0110] In this embodiment, the first basic feature in the basic feature set with dual adaptation labels is {"feature_id":1001, "syntax_label":"personnel information sentence", "lexical_label":"formal title"}. Extract its syntax_label and lexical_label values ​​to obtain the adaptation label information of a single basic feature: {"syntax_label":"personnel information sentence", "lexical_label":"formal title"}.

[0111] Step S1463: Search for core elements that match the basic feature in the ordered core element list, determine the position and name of the core element in the ordered list, and obtain the basic information of the matching core element.

[0112] In this embodiment, the basic feature ID (1001) is used to associate with the feature matching record set, and the corresponding core element name "Reimbursement Person's Name" is found, which is located at the first position (index 0) in the ordered core element list. The basic information of the matching core element is {"element_name":"Reimbursement Person's Name", "position":0}.

[0113] Step S1464: Associate the basic information of the matching core element with the basic feature content and adaptation label information of the single basic feature to form a correspondence record between the single core element and the basic feature, and obtain the feature matching record of the single element.

[0114] In this embodiment, the correspondence record includes the core element name (name of the person seeking reimbursement), location index (0), basic feature ID (1001), sentence structure adaptation tag (personnel information sentence structure), and wording adaptation tag (formal title). The feature matching record of a single element is a dictionary for subsequent integration.

[0115] Step S1465: Extract the next basic feature from the basic feature set with dual adaptation labels, and repeat the operations of extracting adaptation label information, finding matching core elements, determining the basic information of core elements, and forming corresponding relationship records to obtain the corresponding relationship record set for all basic features.

[0116] In this embodiment, steps S1462 to S1464 are executed sequentially on the second basic feature (ID=3002, transportation cost amount) and the third basic feature (ID=2001, departure date) in the basic feature set with dual adaptation labels to form a set of corresponding relationship records for all basic features.

[0117] Step S1466: Sort the set of correspondence records corresponding to all basic features, and rearrange the correspondence records according to the order of the ordered core element list to obtain an ordered set of correspondence records.

[0118] In this embodiment, the position index of the ordered core element list is 0 to 14. The record set is sorted according to the "position" field in the corresponding relationship record to ensure that the record order is consistent with the core element order, thus obtaining an ordered corresponding relationship record set.

[0119] Step S1467: For the ordered set of correspondence records, check whether there is a case where one core element corresponds to multiple basic feature records, and whether there is a case where one basic feature corresponds to multiple core elements, and obtain the entry check results.

[0120] In this embodiment, by checking the uniqueness of the "element_name" field and the uniqueness of the "feature_id" field, it is confirmed that each core element corresponds to one basic feature record, and each basic feature corresponds to one core element. There are no one-to-many or many-to-one situations, and the entry check result is "no anomaly". If an anomaly exists (such as "reimbursement recipient's name" corresponding to two basic feature IDs), it is marked as "duplicate match" and manually verified.

[0121] Step S1468: Based on the entry inspection results, correct the ordered set of correspondence records, remove duplicate record entries, adjust the content of mismatched records, and obtain the corrected set of correspondence records.

[0122] In this embodiment, if the entry check result is "no anomaly", the corrected set of correspondence records is consistent with the ordered set of correspondence records. If duplicate entries exist, the record with the highest semantic similarity score is retained; if mismatches exist, the record set is re-associated based on feature matching.

[0123] Step S1469: Extract all record content from the corrected correspondence record set, and organize the core element names, basic feature content, sentence structure adaptation tag content, and vocabulary selection adaptation tag content according to the preset record format to obtain standardized correspondence record content.

[0124] In this embodiment, the preset record format is CSV, which includes columns for "Core Element Name", "Basic Feature ID", "Sentence Structure Adaptation Tag", and "Wording Adaptation Tag". The values ​​of each field in the corrected correspondence record set are filled into the CSV table, for example: "Reimbursement Person's Name, 1001, Personnel Information Sentence Structure, Formal Title" "Transportation Expense Amount, 3002, Amount Expression Sentence Structure, Amount Terminology", etc., to obtain standardized corresponding record content.

[0125] Step S14610: Integrate the standardized corresponding record content as a whole to form the element feature corresponding record result containing the correspondence between all core elements and basic features and the adaptive label information.

[0126] In this embodiment, the integration process converts the standardized CSV format corresponding record content into JSON format for easier subsequent processing. The result of the element feature corresponding record is a list, and each element is a dictionary containing the core element name, basic feature ID, sentence structure adaptation label, and wording adaptation label, such as [{"element_name":"name of the person seeking reimbursement","feature_id":1001,"syntax_label":"personnel information sentence structure","lexical_label":"formal title"},...].

[0127] Step S147: Extract scene association features from the demand semantic anchoring results, add scene association labels to each correspondence in the feature correspondence record results, and obtain feature correspondence with scene association labels.

[0128] In this embodiment, the scene association features in the demand semantic anchoring results are ["company employees", "domestic industry exhibitions", "travel expenses", "reimbursement"]. Based on the relevance between the core elements and the scene, a scene association label is added to each correspondence. For example, the core element "exhibition name" is associated with the scene "domestic industry exhibitions", so a "scene-exhibition" label is added; "transportation expense amount" is associated with the scene "travel expenses", so a "scene-travel" label is added. For element feature correspondences with scene association labels, a "scene_label" field is added to each dictionary in the element feature correspondence record results, such as {"element_name":"exhibition name", ..., "scene_label":"scene-exhibition"}.

[0129] Step S148: Combine the content expansion triggering logic adjustment results in the large model generation rule adaptation scheme to supplement the feature association information corresponding to the feature correspondence with scene association tags, and obtain the supplemented feature correspondence.

[0130] In this embodiment, the content expansion triggering logic adjustment result includes content expansion rules for each scenario. For example, the "Exhibition Name" element under "Scenario - Exhibition" needs to be expanded with "Location" and "Time". Therefore, the associated expanded feature IDs (e.g., "Location" basic feature ID=4002, "Time" ID=4003) and expansion triggering conditions (e.g., "Extension triggered when exhibition name is not empty") are added to the feature correspondence corresponding to "Exhibition Name". The supplemented feature correspondence adds "expand_features" (list of expanded feature IDs) and "expand_condition" (triggering condition) fields to each dictionary.

[0131] Step S149: Perform an item-by-item verification operation on the supplemented element feature correspondence, check whether each core element has been associated with basic features, sentence structure adaptation tags, wording adaptation tags and scene association tags, and check whether the content of the associated adaptation tags is consistent with the corresponding adjustment logic in the large model generation rule adaptation scheme, so as to obtain the sorted element feature correspondence.

[0132] In this embodiment, item-by-item verification is achieved by writing a verification script. The script iterates through the supplemented element feature correspondences, checking for missing fields (such as an empty "syntax_label") or conflicts between the label content and the large model's generation rule adaptation scheme (such as the "amount expression sentence" label not existing in the corresponding rule). For example, it was found that the "scene_label" for the "departure date" element was missing, so the "scene-itinerary" label was added; it was found that the "lexical_label" for "accommodation fee amount" was "colloquial terminology," which conflicted with the rule, so it was corrected to "amount terminology." The sorted element feature correspondences ensure that all labels are complete and consistent.

[0133] Step S1410: Integrate all the information on the correspondence between the elements and features after sorting out the information, and form a set of credential generation feature mapping associations that includes core elements, basic matching features, various adaptation tags and scene association information.

[0134] In this embodiment, the integration process converts the sorted feature correspondences into a comprehensive JSON file containing top-level fields such as "core_elements" (core element list), "feature_mappings" (basic feature mapping details), "label_info" (various label information), "scene_associations" (scene association information), and "expand_rules" (content expansion rules). The credential generation feature mapping association set comprehensively reflects the correspondence between core elements and basic features, as well as various generation-related labels, serving as one of the core inputs for generating credential text in the multimodal large model.

[0135] Step S150: Based on the large model generation rule adaptation scheme and the voucher generation feature mapping association set, perform intelligent voucher text generation processing. Input the mapped feature information into the multimodal large model according to the adapted generation rules to generate text and obtain the intelligent voucher text generation result.

[0136] Step S151: Extract all generation trigger logic adjustment results in the large model generation rule adaptation scheme, input the generation trigger logic adjustment results into the generation decoding layer of the multimodal large model, complete the parameter configuration of the multimodal large model generation logic, and obtain the generation logic parameter configuration result.

[0137] In this embodiment, the generation trigger logic adjustment result includes trigger conditions and parameters (such as sentence trigger probability, vocabulary weight, and expansion threshold) for sentence construction, vocabulary selection, and content expansion. The multimodal large model's generation and decoding layer contains multiple sub-modules (sentence generator, vocabulary selector, and content expander). By calling the model's configuration API, the parameters in the trigger logic adjustment result are loaded into the corresponding sub-modules. For example, the declarative sentence trigger probability of 0.85 is set to the "prob declarative sentence" parameter of the sentence generator, the formal vocabulary weight of 0.9 is loaded into the weight matrix of the vocabulary selector, and the content expansion threshold of 800 words is set to the "min_length" parameter of the content expander. The generation logic parameter configuration result is a configuration status report, which includes the parameter settings and loading status (success / failure) of each sub-module.

[0138] Step S152: Extract all feature correspondences in the feature mapping association set generated by the voucher, and arrange the feature correspondences in the calling order according to the feature calling priority sorting result in the large model generation rule adaptation scheme to obtain the feature calling order arrangement result.

[0139] In this embodiment, the feature mapping relationships are extracted from the "feature_mappings" field of the voucher generation feature mapping association set, which includes the mapping relationships between all core elements and basic features. According to the feature call priority sorting result in the large model generation rule adaptation scheme (reimbursement recipient's name > transportation expense amount > departure date > ...), the above mapping relationships are sorted to obtain the feature call order arrangement result, i.e., a list of feature mapping relationships arranged by priority. This list determines the order in which elements are processed when the large model generates text.

[0140] Step S153: According to the arrangement result of the feature calling order, the corresponding basic features and adaptation label information are sequentially input into the feature encoding layer of the multimodal large model to obtain the feature receiving result of the feature encoding layer.

[0141] In this embodiment, the feature encoding layer is the input processing layer of the multimodal large model, responsible for converting basic features and label information into vector representations that the model can understand. The results are arranged according to the feature call order, sequentially inputting the basic feature values ​​corresponding to the first element feature (e.g., "Reimbursement applicant's name: Zhang San"), sentence structure adaptation labels ("Personnel information sentence structure"), wording adaptation labels ("Formal title"), and scene association labels ("Scene-Personnel"). The feature encoding layer converts text features (e.g., "Zhang San") into 512-dimensional word vectors through a word embedding layer, and converts label information into label vectors through one-hot encoding. Then, the word vectors and label vectors are concatenated into a 768-dimensional comprehensive feature vector. The feature reception result includes a sequence of comprehensive feature vectors arranged in the call order and corresponding label metadata (e.g., label type, triggering conditions).

[0142] Step S154: Based on the adaptation label information in the feature reception result, perform sentence adaptation processing on each basic feature, adjust the text expression form of the feature according to the preset sentence construction logic, and obtain the feature expression information after sentence adaptation.

[0143] Step S1541: Extract the first basic feature and the corresponding adaptation label information from the feature reception result, separate the sentence adaptation label, and obtain the sentence adaptation label of a single basic feature.

[0144] In this embodiment, the first basic feature in the feature reception result is "Reimbursement applicant's name: Zhang San", and the adaptation label information includes "syntax_label=personnel information sentence pattern" and "lexical_label=formal title", etc. The sentence pattern adaptation label is separated as "personnel information sentence pattern".

[0145] Step S1542: Based on the sentence pattern adaptation tag, retrieve the preset sentence pattern construction logic corresponding to the sentence pattern adaptation tag in the multimodal large model, determine the text expression form and sentence structure requirements of the sentence pattern construction logic, and obtain the corresponding sentence pattern construction rules.

[0146] In this embodiment, the pre-defined sentence construction logic in the multimodal large model is stored in a sentence template library and indexed by sentence adaptation tags. The sentence construction logic corresponding to the "Personnel Information Sentence" is "[Element Name]:[Element Value]([Supplementary Information])", and the text expression form is a declarative sentence. The sentence structure requires the element name, element value, and optional supplementary information (such as department, employee number). The sentence construction rules are represented in JSON format and include "template" (template string), "required_fields" (list of required fields), and "optional_fields" (list of optional fields).

[0147] Step S1543: Decompose the feature content of a single basic feature according to the sentence structure requirements of the sentence construction rules, match the core part of the feature content with the corresponding position of the sentence structure, and obtain the matching result of the feature content and the sentence structure.

[0148] In this embodiment, the feature content of a single basic feature is "Reimbursement applicant's name: Zhang San, Department: Marketing Department, Employee ID: EMP001". According to the sentence structure construction rules, the element name is "Reimbursement applicant's name", the element value is "Zhang San", and the supplementary information is "Department: Marketing Department, Employee ID: EMP001". The matching result is to fill the above parts into the corresponding positions in the sentence structure: [Element Name] → "Reimbursement applicant's name", [Element Value] → "Zhang San", [Supplementary Information] → "Department: Marketing Department, Employee ID: EMP001".

[0149] Step S1544: According to the text expression form of the sentence construction rules, the core part of the matched feature content is integrated into the sentence structure to form the preliminary text expression content that meets the sentence requirements, and the preliminary sentence expression information of a single feature is obtained.

[0150] In this embodiment, the matching result is integrated into the sentence structure to obtain the preliminary text expression: "Reimbursement applicant's name: Zhang San (Department: Marketing Department, Employee ID: EMP001)". This content conforms to the declarative sentence form and structural requirements of the "personnel information sentence".

[0151] Step S1545: For the preliminary expression information of the sentence structure of a single feature, check and fill in or replace the prepositions and auxiliary words required to connect the feature content in the sentence structure to obtain the fluent expression information of the sentence structure of the single feature.

[0152] In this embodiment, the grammatical fluency of the initial text is checked. It is found that the supplementary information in "(Department: Marketing Department, Employee Number: EMP001)" lacks conjunctions. A comma is added as a conjunction, and the text is adjusted to "(Department: Marketing Department, Employee Number: EMP001)" (which meets the requirements and requires no further adjustment). If there are missing particles (such as "Zhang San's department" should be changed to "Zhang San's department"), they are added or replaced to obtain a grammatically correct sentence.

[0153] Step S1546: Extract the next basic feature and its corresponding adaptation label information from the feature reception result, and repeatedly perform the operations of separating sentence adaptation labels, calling sentence construction logic, splitting matching feature content, forming preliminary expression information and adjusting text fluency to obtain a set of sentence fluency expression information for all basic features.

[0154] In this embodiment, steps S1541 to S1545 are executed on the second basic feature "Transportation cost: 3500 yuan" (sentence structure matching tag "amount expression sentence") in the feature reception result. The construction logic of "amount expression sentence" is "[element name]: [Arabic numeral] yuan (uppercase: [Chinese uppercase amount])", which forms the preliminary expression information "Transportation cost: 3500 yuan (uppercase: Three thousand five hundred yuan)" after matching. After checking for fluency, no adjustments are made, and it is added to the sentence fluency expression information set. All basic features are processed in sequence to obtain a set containing all feature text expressions.

[0155] Step S1547: Compare each expression information in the set of grammatical fluency of all basic features with the grammatical structure specified by the corresponding grammatical construction rules to check whether the sentence structure of the expression information conforms to the grammatical structure requirements and obtain the grammatical adaptation check result.

[0156] In this embodiment, the comparison is implemented by writing regular expressions. For example, the regular expression for the "amount expression sentence pattern" is "^[^\:]+\:\d+ yuan (in Chinese capital letters: [zero one two three four five six seven eight nine ten hundred thousand million billion]+ yuan in full) $". Using this expression to check "Transportation expenses amount: 3500 yuan (in Chinese capital letters: Three thousand five hundred yuan in full)", the match is successful and the structure meets the requirements. After all the expression information is checked, a sentence pattern adaptation check result is generated, including the check status (passed / failed) of each expression information and the reason for failure (such as "missing Chinese capital amount").

[0157] Step S1548: According to the sentence pattern adaptation check result, correct the expression information with sentence pattern adaptation errors, and readjust the text expression form according to the corresponding sentence pattern construction logic to obtain a corrected set of sentence pattern expression information.

[0158] In this embodiment, if a certain expression information fails the check (for example, "Accommodation expenses amount: 2000 yuan (in Chinese capital letters: Two thousand yuan)" lacks the word "in full"), it is corrected to "Accommodation expenses amount: 2000 yuan (in Chinese capital letters: Two thousand yuan in full)" according to the sentence pattern construction rules. The corrected set of sentence pattern expression information includes all the characteristic text expressions that have passed the check or been corrected.

[0159] Step S1549: Reorder the corrected set of sentence pattern expression information according to the order of the basic features in the feature reception result to obtain an ordered set of sentence pattern expression information; according to the preset voucher text sentence pattern specification, perform punctuation correction and space unification processing on each expression information in the ordered set of sentence pattern expression information to obtain the characteristic expression information after sentence pattern adaptation.

[0160] In this embodiment, the order of the basic features in the feature reception result is the order arranged by priority. Therefore, the corrected set of sentence pattern expression information is reordered according to this order. The preset voucher text sentence pattern specification requires the use of full-width punctuation marks (such as ",", "."), add a space between numbers and Chinese characters (such as "3500 yuan"), and indent the first line of each paragraph by 2 characters. After the correction processing, the characteristic expression information after sentence pattern adaptation is a list of text segments with unified format and correct order.

[0161] Step S155: According to the diction adaptation label information in the feature reception result, perform a diction adaptation process on the characteristic expression information after sentence pattern adaptation, and adjust the word selection form in the characteristic expression to obtain the characteristic expression information after diction adaptation.

[0162] In this embodiment, word choice adaptation tag information, such as "formal title" and "monetary terminology," guides vocabulary adjustment. For example, "Zhang San" is adjusted to "Comrade Zhang San" under the "formal title" tag; "amount" is adjusted to "amount" under the "monetary terminology" tag; and the colloquial phrase "get it done" is replaced with "completed." Vocabulary adaptation is achieved by calling the vocabulary replacement interface of the large model. The interface selects replacement words from the formal vocabulary library based on the tags to ensure that the text style conforms to the expression norms. The feature expression information after vocabulary adaptation is as follows: "Reimbursement applicant's name: Comrade Zhang San (Department: Marketing Department, Employee Number: EMP001)" "Transportation expense amount: 3500 yuan (in words: Three Thousand Five Hundred Yuan Only)".

[0163] Step S156: Based on the scene association label information in the feature reception result, perform scene association processing on the feature expression information after word adaptation, integrate the text expression elements corresponding to the scene, and obtain the scene-associated feature expression information.

[0164] In this embodiment, scene-related tag information, such as "Scene-Exhibition" and "Scene-Travel," guides the integration of scene elements. For example, under the "Scene-Exhibition" tag, "(hereinafter referred to as 'Digital Economy Expo')" is added to "Exhibition Name: 2023 China International Digital Economy Expo" to simplify subsequent reference; under the "Scene-Travel" tag, "(Travel Start Date)" is added to "Departure Date: October 15, 2023" to clarify the scene meaning. Scene-related processing is achieved by inserting scene-related modifiers or supplementary explanations into the feature expression information, enhancing the scene-specificity of the text.

[0165] Step S157: Based on the content expansion trigger logic in the generated logic parameter configuration result, insert preset background description text or connecting statements related to the core element requirements of the voucher into the feature expression information after scene association, and obtain the feature expression information after content expansion.

[0166] In this embodiment, the content expansion trigger logic determines whether to expand based on the current text length and scene tags. For example, when the text length of the "Expense Details" section is less than 40% of the total length, expansion is triggered, inserting a preset background explanatory text: "All travel expenses incurred during the Digital Economy Expo are in accordance with the company's travel expense reimbursement standards." Connecting statements are inserted between different feature expressions, such as "The basic information of the person seeking reimbursement is as follows:" "The itinerary for this trip is as follows:" "The details of each expense are as follows:", making the text transition more natural. The expanded feature expressions include the original feature expressions, scene elements, and expanded content, enhancing the text's completeness and coherence.

[0167] Step S158: Arrange all the feature expression information after content expansion according to the feature calling order and perform text splicing to form a continuous initial voucher text, thus obtaining the intelligent initial voucher text result.

[0168] For example, step S1581: extract the feature call order arrangement result, determine the splicing order and splicing position requirements of all feature expression information, and obtain text splicing rule information.

[0169] In this embodiment, the sequence of feature calls determines the splicing order (from highest to lowest priority), and the splicing position is determined according to the voucher text structure (such as the sections "Reimbursement Person Information", "Itinerary Information", and "Expense Details"). The text splicing rule information includes the section title (such as "I. Basic Information of Reimbursement Person" and "II. Travel Itinerary Information"), the section order, the correspondence between feature expression information and the section (such as "Reimbursement Person Name" belonging to the "Basic Information of Reimbursement Person" section), and the paragraph separator (newline character "\n").

[0170] Step S1582: Extract the first feature expression information from the feature expression information set after content expansion, confirm the splicing position of the feature expression information in the text splicing rule information and the feature expression information connected before and after it, and obtain the splicing positioning information of a single feature.

[0171] In this embodiment, the first feature expression information is "Name of the person requesting reimbursement: Comrade Zhang San (Department: Marketing Department, Employee Number: EMP001)". According to the text splicing rules, its splicing position is the first paragraph under the "I. Basic Information of the Person Requesting Reimbursement" section, with no preceding information and following the "Department" feature expression information. The splicing positioning information of a single feature includes the chapter ID, paragraph index, and preceding and following feature IDs.

[0172] Step S1583: Use the feature expression information as the starting part of the initial voucher text to generate the initial fragment of the initial voucher text, and obtain the basic initial fragment of the text.

[0173] In this embodiment, the initial text fragment is the chapter title plus the first feature expression information: "I. Basic Information of the Reimbursement Person\nReimbursement Person's Name: Comrade Zhang San (Department: Marketing Department, Employee Number: EMP001)".

[0174] Step S1584: Extract the next feature expression information from the feature expression information set after content expansion, confirm the splicing and positioning information of the feature expression information, determine the connection form between the feature expression information and the previous feature expression information, and obtain the connection requirements between features.

[0175] In this embodiment, the next feature expression information is "Department: Marketing Department", the splicing positioning information is the second paragraph under the "I. Basic Information of the Expense Claimant" section, and the connection form with the previous feature expression information is "Supplementary Explanation". The connection requirement is to use a semicolon to separate or start a new sentence.

[0176] Step S1585: According to the requirements for connection between features, the feature expression information is spliced ​​to the subsequent position of the initial basic text segment, and the connecting words between the two expression information are adjusted to make the spliced ​​text sentence coherent, thus obtaining the spliced ​​basic text segment.

[0177] In this embodiment, a new sentence is used for connection, resulting in: "I. Basic Information of the Person Reimbursing Reimbursement\nName of the Person Reimbursing Reimbursement: Comrade Zhang San (Department: Marketing Department, Employee Number: EMP001).\nDepartment: Marketing Department." The conjunction "," is changed to "." to conform to formal written language habits.

[0178] Step S1586: Extract subsequent feature expression information from the feature expression information set after content expansion, repeatedly perform the operations of confirming splicing positioning information, determining connection requirements and splicing adjustment, and splice all feature expression information into the basic text fragment in sequence to obtain the spliced ​​text fragment.

[0179] In this embodiment, feature information such as "employee number," "exhibition name," and "departure date" are sequentially concatenated. Each feature is added to a corresponding section based on its concatenation location information. For example, "exhibition name" is added to the section "II. Travel Itinerary Information," and "transportation cost amount" is added to the section "III. Expense Details." During the concatenation process, connecting words and paragraph structure are dynamically adjusted to ensure smooth text flow.

[0180] Step S1587: Traverse the spliced ​​text fragments, compare the feature expression information contained therein with the feature call order arrangement result, check whether the number and arrangement order of feature expression information are consistent with the feature call order arrangement result, and obtain the splicing integrity check result.

[0181] In this embodiment, the feature call order arrangement result contains 15 features, and the concatenated text fragment also contains 15 feature expression information. The order is consistent, and the concatenation integrity check result is "complete". If there are missing or incorrect order, the check result is "incomplete", and it needs to be re-concatenated.

[0182] Step S1588: Based on the splicing integrity check result, correct the misordered splicing text fragments by adjusting the misordered feature expression information to the correct splicing position to obtain the corrected splicing text fragments.

[0183] In this embodiment, the splicing integrity check result is "complete", no correction is required, and the corrected spliced ​​text fragment is consistent with the spliced ​​text fragment.

[0184] Step S1589: For the corrected spliced ​​text fragment, check and adjust the connecting words between adjacent feature expression information, and unify the referential words for the same core element in the whole text to obtain the optimized spliced ​​text.

[0185] In this embodiment, the conjunction adjustment replaces repeated "." with ";" for concise expression, such as "Employee ID: EMP001. Department: Marketing Department." being adjusted to "Employee ID: EMP001; Department: Marketing Department."; the pronoun is unified so that "Comrade Zhang San" is consistently referred to as "this employee" in the subsequent text, such as "This employee departed for Shijiazhuang to participate in the Digital Economy Expo on October 15, 2023". Step S15810: Perform preliminary formatting of the optimized concatenated text according to the conventional format of voucher text, adjust the line breaks and paragraph breaks of the text to form a continuous initial voucher text that conforms to the basic format of voucher text, and obtain the intelligent initial voucher text result.

[0186] In this embodiment, the preliminary layout includes: bolding chapter titles (using HTML). Tags), paragraph first-line indentation of 2 characters (achieved via CSS style text-indent:2em), and use of bullet points for expense details. Add tags), page headers and footers (e.g., header "Company Travel Expense Reimbursement Form", footer "Page 1 of 1"). The intelligent initial voucher text result is continuous text in HTML format, with a clear structure and standardized format.

[0187] Step S159: Based on the vocabulary selection trigger logic and sentence construction trigger logic in the large model generation rule adaptation scheme, adjust the connecting words and pronouns between different feature expression information in the intelligent initial voucher text result to obtain the optimized voucher text.

[0188] In this embodiment, the conjunction adjustment is triggered by vocabulary selection logic, changing ";", "," etc. to combinations of "." and ";" that are more in line with formal written language. For example, long sentences are broken down into shorter sentences to enhance readability. The pronoun adjustment is based on context logic, changing potentially ambiguous phrases like "the employee" to "the person who reimburses" in a specific context, such as "the person who reimburses 2,000 yuan for accommodation during the exhibition." The adjustments are implemented by calling the text optimization interface of the large model, which outputs the optimized voucher text.

[0189] Step S1510: Output the coherence-optimized voucher text as the intelligent voucher text generation result.

[0190] In this embodiment, the optimized voucher text is in HTML format, which is converted to PDF format by calling a format conversion interface. A company electronic signature (a digital signature based on the PKI system) and an anti-counterfeiting watermark (containing a randomly generated QR code linking to the electronic voucher in the reimbursement system) are added. The final intelligent voucher text is generated as a PDF file, stored in the voucher database, and the claimant and approver are notified via email and system messages.

[0191] In the above embodiment, during the pre-training process of the multimodal large model, a voucher text training dataset is first constructed, which includes historical reimbursement voucher samples, voucher template samples, and annotation information.

[0192] In this embodiment, historical expense reimbursement voucher samples were selected from the company's historical data in the financial system, including 100,000 vouchers of various types such as travel expense reimbursement forms and purchase expense reimbursement forms from the past three years. The formats included PDF, Word, and scanned images (which needed to be converted to text using OCR). Voucher template samples were extracted from financial policy documents, including 200 officially released voucher templates. Annotation information was manually added by financial experts, including element annotations (such as using rectangles to mark the "amount" position), sentence structure annotations (annotating declarative sentences, imperative sentences, etc.), and scenario annotations (annotating scenarios such as "travel" and "purchase"). The training dataset was divided into training and validation sets in an 8:2 ratio, and the format was converted to JSONLines format that the model could read.

[0193] Then, configure the pre-training parameters of the multimodal large model, including learning rate, batch size, number of training epochs, and optimizer type.

[0194] In this embodiment, the pre-training parameters are set as follows: initial learning rate = 5e-5, using a linear learning rate decay strategy; batch size = 32 (adjusted according to GPU memory, set to 32 when a single card has 24GB of VRAM); number of training epochs = 10 (the number of early stopping epochs is determined by the validation set loss; training stops if the validation set loss does not decrease for 3 consecutive epochs); optimizer = AdamW (weight decay = 0.01, β1 = 0.9, β2 = 0.999); dropout rate = 0.1; sequence length = 512 (voucher text is usually short, 512 is sufficient for coverage). The parameters are stored in a YAML configuration file.

[0195] The multimodal large model was pre-trained using a credential text training dataset, employing masked language modeling (MLM) and sentence order prediction (NSP) tasks.

[0196] In this embodiment, pre-training is performed on a large model based on the Transformer architecture (such as BERT-Base). First, pre-trained weights are loaded, and then training continues on the credential text training dataset. The MLM task randomly masks 15% of the tokens in the input text, allowing the model to predict the masked tokens; the NSP task asks the model to predict whether two sentences are consecutive sentence pairs. The training process is accelerated using mixed precision training (FP16) and employs a distributed training strategy (8 GPUs in parallel). After each epoch, the accuracy of MLM and NSP is evaluated on the validation set, and the weights of the best-performing model are saved.

[0197] Step S214: Fine-tune the pre-trained multimodal large model and optimize the model parameters using the travel expense reimbursement voucher dataset.

[0198] In this embodiment, the travel expense reimbursement voucher dataset contains 50,000 more finely annotated travel expense reimbursement forms. The annotation information includes scene features, element relationships, generation rules, etc. During fine-tuning, the parameters of the bottom 9 layers of the model are frozen, and only the top 3 layers and the classification head are fine-tuned. The learning rate is set to 2e-5, batch size=16, and training epochs=5. The fine-tuning task is a text generation task, with the input being the requirement text and the output being the voucher text, optimized using the cross-entropy loss function. The fine-tuned model achieves a BLEU score of 0.85 on the test set, indicating that the generated text is highly similar to the reference text.

[0199] Figure 2 The diagram illustrates the hardware structure of an intelligent credential automatic generation system 100 based on a multimodal large model provided in an embodiment of the present invention, such as... Figure 2 As shown, the intelligent voucher automatic generation system 100 based on a multimodal large model may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0200] Machine-readable storage medium 120 can store data and / or instructions. In some embodiments, machine-readable storage medium 120 can store data acquired from an external terminal. In some embodiments, machine-readable storage medium 120 can store data and / or instructions used by the multimodal large-scale smart credential automatic generation system 100 to execute or use in order to complete the exemplary methods described in this invention. In a specific implementation, one or more processors 110 execute the computer-executable instructions stored in machine-readable storage medium 120, enabling processor 110 to execute the multimodal large-scale smart credential automatic generation method as described in the above method embodiments. Processor 110, machine-readable storage medium 120, and communication unit 140 are connected via bus 130, and processor 110 can be used to control the sending and receiving actions of communication unit 140. The specific implementation process of processor 110 can be found in the various method embodiments executed by the multimodal large-scale smart credential automatic generation system 100 described above, and their implementation principles and technical effects are similar, so they will not be repeated here.

[0201] Furthermore, this embodiment of the invention also provides a readable storage medium containing computer-executable instructions. When the processor executes the computer-executable instructions, the above-mentioned method for automatically generating intelligent credentials based on a multimodal large model is implemented.

[0202] It should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. Similarly, it should be noted that, in order to simplify the description of this invention and thus aid in the understanding of one or more embodiments, the foregoing description of the embodiments of this invention sometimes combines multiple features into a single embodiment, drawing, or description thereof. < / unmatch> < / partial> < / unmatch> < / partial> < / match>

Claims

1. A method for automatically generating intelligent vouchers based on a multimodal large model, characterized in that, The method includes: Obtain the preset voucher generation requirement text, which includes a description of the voucher usage scenario, the requirements for the core elements of the voucher, and the voucher text expression specifications, thus obtaining the voucher generation requirement text; The requirement text for voucher generation is subjected to semantic anchoring processing. Various descriptions in the requirement text for voucher generation are semantically matched with the preset voucher basic element library to obtain the requirement semantic anchoring result. The semantic anchoring results of the demand are adapted to the large model generation rules. The text generation triggering conditions of the multimodal large model are adjusted in combination with the expression specifications of the voucher generation demand text to obtain the large model generation rule adaptation scheme. Based on the large model generation rule adaptation scheme, feature mapping processing is performed on the voucher generation requirement text, and the core elements of the voucher generation requirement text are associated with the generation feature library of the multimodal large model to obtain the voucher generation feature mapping association set. Intelligent voucher text generation is performed based on a large model generation rule adaptation scheme and a voucher generation feature mapping association set. The mapped feature information is input into a multimodal large model according to the adapted generation rules to generate text, and the intelligent voucher text generation result is obtained.

2. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 1, characterized in that, The process of adapting the semantic anchoring results of the demand to the large model generation rules, and adjusting the text generation triggering conditions of the multimodal large model in conjunction with the expression specifications of the voucher generation demand text, yields a large model generation rule adaptation scheme, including: Extract the scene association features corresponding to the voucher usage scenario description from the demand semantic anchoring results, and perform feature matching between the scene association features and the preset scene generation template in the multimodal large model to obtain scene template matching association information; Extract the core elements of the voucher from the voucher generation requirement text, and match the element features with the basic element features in the multimodal large model's generation feature library one by one to obtain the element feature corresponding association information. Extract the expression features corresponding to the voucher text expression specifications in the voucher generation requirement text, and decompose the expression features into text sentence structure features, text wording features, and text length features to obtain the expression feature decomposition results; By combining the scene template matching association information and the element feature corresponding association information, the feature call priority when the multimodal large model generates voucher text is determined, and the feature call priority ranking result is obtained; Based on the text sentence features in the expression feature decomposition results, the sentence construction triggering logic when the multimodal large model generates text is adjusted to obtain the sentence construction triggering logic adjustment results; Based on the text wording features in the expression feature decomposition results, the word selection triggering logic when the multimodal large model generates text is adjusted to obtain the word selection triggering logic adjustment results; Based on the text length features in the expression feature decomposition results, the content expansion triggering logic when the multimodal large model generates text is adjusted to obtain the content expansion triggering logic adjustment results; By integrating the results of feature call priority sorting, sentence structure construction triggering logic adjustment, vocabulary selection triggering logic adjustment, and content expansion triggering logic adjustment, a preliminary generation rule adjustment scheme for a multimodal large model is formed. The initial generated rule adjustment scheme is then correlated with the semantic anchoring result of the requirement, and the generated rule details corresponding to the scene association features are supplemented to obtain the supplemented generated rule adjustment scheme. Integrate all feature adjustment information of the supplemented generation rule adjustment scheme to form a multimodal large model generation rule adaptation scheme. The large model generation rule adaptation scheme includes all adjusted trigger logic and feature call requirements.

3. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 1, characterized in that, The feature mapping process performed on the voucher generation requirement text according to the large model generation rule adaptation scheme establishes a corresponding association between the core elements of the voucher generation requirement text and the generation feature library of the multimodal large model, resulting in a voucher generation feature mapping association set, including: Extract the feature call priority ranking result from the large model generation rule adaptation scheme, and reorder the core elements of the voucher generation requirement text according to the feature call priority ranking result to obtain the core element priority ranking result. Extract all basic features from the generated feature library of the multimodal large model, and classify and organize the basic features according to the voucher element type to obtain the classification and organization results of the basic features; Each core element in the priority ranking result of core elements is compared with the corresponding voucher element type in the basic feature classification result to filter out the basic features that match the core elements and obtain the preliminary screening result of basic features. By combining the sentence structure construction trigger logic adjustment results in the large model generation rule adaptation scheme, sentence structure adaptation tags are added to the basic features after preliminary screening, resulting in a set of basic features with sentence structure adaptation tags. By combining the word selection triggering logic adjustment results in the large model generation rule adaptation scheme, word choice adaptation labels are added to each feature in the basic feature set with sentence structure adaptation labels, resulting in a basic feature set with dual adaptation labels. The basic feature set with dual adaptation labels is matched one-to-one with the core elements of the voucher generation requirement text. The basic features and adaptation label information corresponding to each core element are recorded to obtain the element feature correspondence recording results. Extract scene association features from the semantic anchoring results of demand, add scene association labels to each correspondence in the feature correspondence record results, and obtain feature correspondence with scene association labels; By combining the content expansion triggering logic adjustment results in the large model generation rule adaptation scheme, the feature association information corresponding to the feature correspondence with scene association tags is supplemented with the content expansion, resulting in the supplemented feature correspondence. The supplemented element feature correspondence is checked item by item to see if each core element has been associated with basic features, sentence structure adaptation tags, wording adaptation tags and scene association tags, and to see if the content of the associated adaptation tags is consistent with the corresponding adjustment logic in the large model generation rule adaptation scheme, so as to obtain the sorted element feature correspondence. All information on the correspondence between the integrated and sorted elements and features is used to form a set of credential generation feature mapping associations that includes core elements, basic matching features, various adaptation tags, and scenario-related information.

4. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 1, characterized in that, The intelligent voucher text generation process, based on the large model generation rule adaptation scheme and the voucher generation feature mapping association set, involves inputting the mapped feature information into a multimodal large model according to the adapted generation rules for text generation, resulting in the intelligent voucher text generation result, including: Extract all generation trigger logic adjustment results from the large model generation rule adaptation scheme, input the generation trigger logic adjustment results into the generation decoding layer of the multimodal large model, complete the parameter configuration of the multimodal large model generation logic, and obtain the generation logic parameter configuration result; Extract all feature correspondences of elements in the feature mapping association set generated from the voucher, and arrange the feature correspondences in the calling order according to the feature calling priority sorting result in the large model generation rule adaptation scheme to obtain the feature calling order arrangement result; Based on the arrangement of the feature call order, the corresponding basic features and adaptation label information are sequentially input into the feature encoding layer of the multimodal large model to obtain the feature reception result of the feature encoding layer. Based on the adaptation label information in the feature reception results, sentence adaptation processing is performed on each basic feature. The text expression form of the feature is adjusted according to the preset sentence construction logic to obtain the sentence-adapted feature expression information. Based on the word matching tag information in the feature reception results, the word matching processing is performed on the feature expression information after sentence matching, and the word selection form in the feature expression is adjusted to obtain the feature expression information after word matching. Based on the scene association label information in the feature reception result, the feature expression information after word adaptation is processed by scene association, and the text expression elements corresponding to the scene are integrated to obtain the scene-associated feature expression information. Based on the content expansion trigger logic in the configuration result of the generated logic parameters, a preset background description text or connecting statement related to the core element requirements of the voucher is inserted into the feature expression information after the scenario association, so as to obtain the feature expression information after content expansion. The results of arranging all the feature expression information after content expansion according to the feature calling order are processed by text splicing to form a continuous initial voucher text, resulting in the intelligent initial voucher text result. Based on the vocabulary selection triggering logic and sentence construction triggering logic in the large model generation rule adaptation scheme, the connecting words and pronouns between different feature expression information in the intelligent initial voucher text result are adjusted to obtain the optimized voucher text; Output the coherence-optimized voucher text as the intelligent voucher text generation result.

5. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 2, characterized in that, The extraction of scene-related features corresponding to the voucher usage scenario description from the semantic anchoring results of the demand is performed. These scene-related features are then matched with preset scene generation templates in a multimodal large model to obtain scene template matching association information, including: Traverse all semantic association information of the requirement semantic anchoring result, identify all feature information that has semantic association with the voucher usage scenario description in the voucher generation requirement text, and obtain the original information of scenario association features; Feature extraction is performed on the original information of scene-related features, and redundant information that is irrelevant to scene generation is removed. The feature content that can represent the core attributes of the voucher usage scenario is retained to obtain purified scene-related features. Retrieve all preset scene generation templates in the multimodal large model, extract the template feature identifier of each scene generation template, and form a set of scene generation template feature identifiers. Each template feature identifier uniquely corresponds to a scene generation template. The purified scene association features are semantically compared with each template feature identifier in the scene generation template feature identifier set. The degree of semantic association between the features is calculated to obtain the semantic association degree comparison result. Based on the semantic association degree comparison results, the scene generation template feature identifiers with the highest semantic association degree with the purified scene association features are selected to obtain candidate scene generation template feature identifiers. Retrieve the scene generation template corresponding to the feature identifier of the candidate scene generation template, and extract the template structure features, template wording features and template element arrangement features of the scene generation template to obtain the core features of the candidate scene generation template; The purified scene association features are matched with the core features of the candidate scene generation template in a dimensional manner, and the matching results and matching details of each dimension are recorded to obtain a dimensional matching result record. For the dimensional matching results recorded, the average matching degree of each dimension matching results is calculated as the overall matching degree between the purified scene association features and the candidate scene generation template, and the specific feature points that are successfully matched in each dimension are recorded to obtain the overall matching evaluation result. Based on the overall matching evaluation results, the feature parts that match the candidate scene generation template and the feature parts that do not match in the scene association features are marked to obtain the matching feature annotation information; The core features of the candidate scene generation template, the dimensional matching result records, the overall matching evaluation results, and the matching feature annotation information are integrated to form scene template matching association information, which contains all relevant details of scene matching.

6. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 2, characterized in that, The results of the fusion feature call priority ranking, sentence structure construction trigger logic adjustment, vocabulary selection trigger logic adjustment, and content expansion trigger logic adjustment form a preliminary generation rule adjustment scheme for the multimodal large model, including: The feature call priority ranking result is used as the core framework of the preliminary rule adjustment scheme to determine the core feature call order when the multimodal large model generates voucher text and to build a preliminary rule framework structure. The result of the sentence construction trigger logic adjustment is decomposed into multiple sentence construction sub-logics. According to the feature calling order of the preliminary rule framework structure, each sentence construction sub-logic is embedded into the corresponding position of the core framework to obtain the rule framework with embedded sentence logic. The result of adjusting the word selection trigger logic is decomposed into multiple word selection sub-logics. According to the feature calling order of the rule framework embedded in the sentence logic, each word selection sub-logic is embedded into the corresponding position of the framework, forming a logical correspondence with the sentence construction sub-logic, thus obtaining a rule framework with embedded dual logic. The content expansion trigger logic adjustment result is decomposed into multiple content expansion sub-logics. According to the feature call order of the rule framework with embedded dual logic, each content expansion sub-logic is embedded into the corresponding position of the framework, forming a triple logic correspondence with the sentence construction sub-logic and the vocabulary selection sub-logic, thus obtaining a rule framework with embedded triple logic. Extract all logical adjustment information from the rule framework embedded with triple logic, classify and organize it according to feature type, and obtain a set of classified and organized logical adjustment information; After traversing the categorized and organized set of logical adjustment information, for the same feature call position in the preliminary rule framework structure, check whether there are sentence construction sub-logic, vocabulary selection sub-logic, and content expansion sub-logic; if so, arrange the above sub-logic at that position in the order of sentence construction, vocabulary selection, and content expansion to obtain a logically coherent set of adjustment information. Add a corresponding feature call identifier to each adjustment logic in the logically coherent adjustment information set, so that each adjustment logic can be accurately mapped to a specific feature call position in the core framework, resulting in an adjustment information set with feature identifiers; The set of adjustment information with feature identifiers is matched again with the feature call priority sorting results to confirm that the adjustment logic of each feature call position is complete and without omission, and the set of adjustment information after matching confirmation is obtained. The set of adjustment information after matching and confirmation is integrated as a whole. The triggering conditions and execution requirements of all adjustment logics are sorted out according to the feature calling order to obtain the sorted generation rule adjustment information. The sorted generation rule adjustment information is standardized and organized according to the generation rule format of multimodal large models to form a preliminary generation rule adjustment scheme for multimodal large models that meets the configuration requirements of large models.

7. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 3, characterized in that, The step involves comparing each core element in the priority ranking result with the corresponding voucher element type in the basic feature classification result to filter out the basic features that match the core elements, thus obtaining the preliminary basic feature filtering result, including: Extract the first core element from the priority ranking results, determine the element type and core content of the element, and obtain the basic information of a single core element. In the results of basic feature classification, find the basic feature classification that matches the feature type of the core element, extract all basic features under the basic feature classification, and obtain a set of basic features of the same type. The core expression content of a single core element is compared with each basic feature in the same type of basic feature set to identify basic features that have semantic overlap with the core expression content of the element, thus obtaining a candidate set of semantically overlapping features. Perform feature integrity analysis on each basic feature in the semantically overlapping feature candidate set to confirm whether the basic feature can fully represent the core expression content of the core element and obtain the feature integrity analysis result. Based on the feature integrity analysis results, the basic features that can fully represent the core elements are selected to obtain the matching basic features of a single core element. The basic matching features of a single core element are associated with that core element, and the corresponding element type and matching details are labeled to obtain the feature matching record of a single element. Extract the next core element from the core element priority ranking result, and repeat the operations of determining the basic information of the element, finding the set of basic features of the same type, semantic feature comparison, feature integrity analysis, and filtering and matching basic features and related records to obtain the feature matching record set of all core elements. The core elements in the feature matching record set of all core elements are compared with the core element priority ranking result. It is checked whether each core element in the core element priority ranking result exists in the feature matching record set, and the matching record check result is obtained. Based on the results of the matching record check, the basic matching features of all feature matching records are extracted and arranged in the order of the priority sorting results of the core elements to obtain an ordered set of basic matching features. The ordered set of basic matching features is organized as a whole, and the duplicate basic features are removed. The unique basic matching feature corresponding to each core element is retained to form the preliminary screening results of basic features.

8. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 3, characterized in that, The process involves mapping the set of basic features with dual adaptation tags to the core elements of the voucher generation requirement text, recording the basic features and adaptation tag information corresponding to each core element, and obtaining the element feature correspondence recording results, including: Extract the original list of core elements from the voucher to generate the required text, and rearrange the original list according to the priority of the core elements to obtain an ordered list of core elements. Extract the first basic feature from the basic feature set with dual adaptation labels, and extract the sentence structure adaptation label and vocabulary selection adaptation label corresponding to the basic feature to obtain the adaptation label information of a single basic feature. Search for core elements that match the basic feature in the ordered list of core elements, determine the position and name of the core element in the ordered list, and obtain the basic information of the matching core element. The basic information of the matching core elements is associated with the basic feature content and matching tag information of a single basic feature to form a correspondence record between a single core element and a basic feature, thus obtaining a single correspondence record. Extract the next basic feature from the set of basic features with dual adaptation labels, and repeat the operations of extracting adaptation label information, finding matching core elements, determining the basic information of core elements and forming corresponding relationship records to obtain the set of corresponding relationship records for all basic features. Sort the set of correspondence records corresponding to all basic features, and rearrange the correspondence records according to the order of the ordered core element list to obtain an ordered set of correspondence records; For an ordered set of correspondence records, check whether there is a case where one core element corresponds to multiple basic feature records, and whether there is a case where one basic feature corresponds to multiple core elements, and obtain the item check results; Based on the item inspection results, the ordered set of correspondence records is corrected by removing duplicate record items and adjusting the content of mismatched records to obtain the corrected set of correspondence records. Extract all record content from the corrected correspondence record set, and organize the core element names, basic feature content, sentence structure adaptation tag content, and vocabulary selection adaptation tag content according to the preset record format to obtain standardized correspondence record content. The standardized corresponding records are integrated as a whole to form the element feature correspondence record results, which include the correspondence between all core elements and basic features, as well as the adaptive label information.

9. The method for automatically generating intelligent vouchers based on a multimodal large model according to claim 4, characterized in that, The step involves performing sentence adaptation processing on each basic feature based on the adaptation tag information in the feature reception result, adjusting the textual expression form of the feature according to a preset sentence construction logic, and obtaining the sentence-adapted feature expression information, including: Extract the first basic feature and the corresponding adaptation label information from the feature reception result, separate the sentence adaptation label, and obtain the sentence adaptation label of a single basic feature. Based on the sentence pattern adaptation tag, the pre-set sentence pattern construction logic corresponding to the sentence pattern adaptation tag in the multimodal large model is retrieved to determine the text expression form and sentence structure requirements of the sentence pattern construction logic, and obtain the corresponding sentence pattern construction rules. The feature content of a single basic feature is split according to the sentence structure requirements of the sentence construction rules. The core part of the feature content is matched with the corresponding position of the sentence structure to obtain the matching result of the feature content and the sentence structure. According to the text expression form of the sentence structure construction rules, the core part of the matched feature content is integrated into the sentence structure to form the preliminary text expression content that meets the sentence requirements, and the preliminary sentence expression information of a single feature is obtained. The sentence structure of a single feature is initially expressed. The prepositions and auxiliary words required to connect the feature content in the sentence structure are checked, filled, or replaced to obtain the fluent sentence expression information of the single feature. Extract the next basic feature and its corresponding adaptation label information from the feature reception result, and repeatedly perform the operations of separating sentence adaptation labels, calling sentence construction logic, splitting matching feature content, forming preliminary expression information and adjusting text fluency to obtain a set of sentence fluency expression information for all basic features; Each expression information in the set of sentence fluency expression information of all the basic features is compared with the sentence structure specified by the corresponding sentence construction rule to check whether the sentence structure of the expression information conforms to the sentence structure requirements, and the sentence adaptation check result is obtained. Based on the sentence matching check results, the expression information with incorrect sentence matching is corrected, and the text expression form is readjusted according to the corresponding sentence construction logic to obtain the corrected sentence expression information set. According to the basic feature order in the feature reception results, the corrected sentence expression information set is reordered to obtain an ordered sentence expression information set; according to the preset voucher text sentence format standard, punctuation correction and space unification processing are performed on each expression information in the ordered sentence expression information set to obtain the sentence-adapted feature expression information.

10. An intelligent voucher automatic generation system based on a multimodal large model, characterized in that, The intelligent voucher automatic generation system based on a multimodal large model includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to run the programs, instructions or code in the memory to implement the intelligent voucher automatic generation method based on a multimodal large model as described in any one of claims 1-9.

Images