Multi-dimensional intelligent processing method and system for auditing text

Abstract

The application discloses a kind of multi-dimension intelligent processing method and system of auditing text, it is related to data processing technical field, and the text feature vector and problem feature vector are constructed by the automatic word segmentation, part-of-speech tagging and vectorization coding of the to-be-processed auditing text and problem description text in the present method, eliminate the subjective judgment difference of different auditors, combine static knowledge document full extraction and dynamic case base semantic retrieval, realize the accurate recall of historical cases by cosine similarity calculation and reordering model screening, and the structured input sequence is input into the private deployment language model to obtain the structured text sequence, which is converted into the target format document by format analysis for direct use, while the newly added case vector of artificial confirmation is written into dynamic case base to trigger incremental update, and when the cumulative new case exceeds the iteration threshold, the dynamic case base is executed full reconstruction, forming the closed-loop iteration mechanism of continuously improving case retrieval accuracy.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a multi-dimensional intelligent processing method and system for audit text. Background Technology

[0002] In the internal audit work of an enterprise, the audit department puts forward rectification requirements to the department responsible for rectification based on the problems found in the audit. The department responsible for rectification then drafts the rectification measures and submits them to the audit department for review.

[0003] The rectification measures document typically includes a problem response plan, division of responsibilities, and timeline. Auditors must review the rectification measures document item by item and provide feedback for revision.

[0004] This review process relies on auditors' reading comprehension and experience in interpreting the problem descriptions and corrective measures, involving multi-dimensional analysis and repeated comparisons of the text's semantics.

[0005] In practice, auditors need to refer to multiple types of knowledge documents, including problem classification guidelines, evaluation rules, common problems summary documents, and historical audit case documents. These documents are stored in the office system in electronic format.

[0006] During the review process, auditors need to manually search historical case documents, locate historical cases similar to the current problem description as references, and compare the evaluation points in each document with the rectification measures to be audited item by item. After completing the writing and formatting of the audit opinions, they will provide feedback to the department responsible for rectification.

[0007] In existing technologies, the aforementioned text processing and retrieval processes rely on manual operation. There is a lack of automated means for feature extraction of problem description texts and rectification measure texts, and different auditors have different semantic understandings of the same text. At the same time, the retrieval of historical case documents relies on manual browsing or keyword matching, without vectorization encoding and semantic similarity calculation of case texts, making it difficult to accurately locate semantically related content from a large number of historical cases. In addition, the updating of knowledge documents and the accumulation of cases lack a closed-loop iterative mechanism, and the historical case database cannot automatically update the retrieval index as new cases are confirmed, resulting in limited coverage and accuracy of case retrieval. Summary of the Invention

[0008] The main purpose of this application is to provide a multi-dimensional intelligent processing method and system for audit texts, in order to solve the problems in the existing technology where auditors rely on manual operation to complete the text processing and retrieval process, lack automated means for feature extraction of problem description texts and rectification measure texts, and different auditors have different semantic understandings of the same text, which can easily cause audit errors and reduce work efficiency.

[0009] To achieve the above objectives, this application provides the following technical solution: A multi-dimensional intelligent processing method for audit text, the multi-dimensional intelligent processing method comprising: Step S1: Obtain the audit text to be processed and the problem description text; perform word segmentation and part-of-speech tagging on the audit text to be processed; extract noun words and verb words to construct a text feature vector; and perform the same operation on the problem description text to construct a problem feature vector. Step S2: Perform full text extraction on the pre-configured static knowledge document set to obtain a static knowledge text sequence, and match the problem feature vector with the pre-configured dynamic case library to obtain a search case text sequence; Step S3: The text feature vector, the question feature vector, the static knowledge text sequence, and the retrieval case text sequence are concatenated into a structured input sequence according to a preset template. The structured input sequence is then input into the privately deployed language model, and the structured text sequence output by the privately deployed language model is obtained. Step S4: Extract the heading level tags and table structure tags of the structured text sequence, and convert the structured text sequence into a target format document according to the heading level tags and table structure tags; Step S5: In response to the manual confirmation mark of the target format document, the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark are taken as new cases, the new cases are vectorized and encoded and written into the dynamic case library, triggering the incremental update of the dynamic case library; Step S6: Count the number of newly added cases in the dynamic case library within the preset time window. When the number of newly added cases exceeds the preset iteration threshold, perform a full reconstruction of the dynamic case library and use the fully reconstructed dynamic case library for the next round of matching retrieval.

[0010] Beneficial effects of steps S1 to S6: This method addresses issues such as reliance on manual operation in audit text processing, insufficient accuracy in historical case retrieval, and lack of closed-loop iteration in knowledge base updates. It constructs text feature vectors and issue feature vectors by automatically segmenting, tagging, and vectorizing the audit text and issue description text, eliminating subjective judgment differences among different auditors. It combines full extraction of static knowledge documents with semantic retrieval from a dynamic case library, using cosine similarity calculation and a re-ranking model to accurately recall historical cases. The method inputs structured input sequences into a privately deployed language model to obtain structured text sequences, which are then converted into target format documents for direct use. Simultaneously, manually confirmed new cases are vectorized and written into the dynamic case library to trigger incremental updates. When the accumulated new cases exceed an iteration threshold, a full reconstruction of the dynamic case library is performed, forming a closed-loop iteration mechanism that continuously improves case retrieval accuracy.

[0011] Specifically, step S1 extracts noun and verb words through word segmentation and part-of-speech tagging to construct feature vectors, thereby achieving automated and structured representation of audit text; step S2 combines full extraction of static knowledge with dynamic case library matching and retrieval, balancing the integrity of the evaluation framework with accurate reuse of historical experience; step S3 achieves multi-source knowledge fusion processing and structured output through preset template splicing and private language model reasoning; step S4 achieves automatic formatting of the output document through the extraction and conversion of title level tags and table structure tags; step S5 achieves dynamic expansion of the case library by triggering vectorized writing and incremental index updates for newly added cases through manual confirmation tags; and step S6 achieves periodic optimization of the case vector index and continuous improvement of retrieval accuracy through statistical analysis of the number of newly added cases and full reconstruction.

[0012] As a further improvement to this application, step S1 involves obtaining the audit text to be processed and the problem description text, performing word segmentation and part-of-speech tagging on the audit text to be processed, extracting noun words and verb words to construct a text feature vector, and performing the same operation on the problem description text to construct a problem feature vector, including: Step S1.1: Obtain the audit text to be processed and the problem description text, and perform word segmentation on the audit text to be processed to obtain the word segmentation result; Step S1.2: Perform part-of-speech tagging on the word segmentation results, filter noun words and verb words to obtain a candidate word set; Step S1.3: Vectorize each word in the candidate word set to obtain a word vector sequence; Step S1.4: Perform mean pooling on the word vector sequence to obtain the text feature vector of the audit text to be processed; Step S1.5: Perform the same word segmentation, part-of-speech tagging, vectorization encoding, and mean pooling operations on the problem description text as on the audit text to be processed, to obtain the problem feature vector.

[0013] Beneficial effects of steps S1.1 to S1.5: This series of steps addresses the lack of automated feature extraction methods for audit texts. By using word segmentation, part-of-speech tagging, and vectorized encoding to construct text feature vectors and question feature vectors, the audit text is automatically converted from unstructured natural language to structured numerical representation, providing standardized input for subsequent matching retrieval and model inference.

[0014] In this process, step S1.1 divides the continuous text into independent word units through word segmentation to obtain the word segmentation results; step S1.2 filters noun and verb words through part-of-speech tagging and removes functional word noise to obtain a candidate word set; step S1.3 maps discrete words to continuous numerical vectors by vectorizing each word to obtain a word vector sequence; step S1.4 compresses the variable-length sequence into a fixed-dimensional vector by performing mean pooling on the word vector sequence to obtain the text feature vector; and step S1.5 performs the same processing flow on the question description text to obtain the question feature vector, providing a retrieval basis for the matching retrieval in the subsequent step S2.

[0015] As a further improvement to this application, step S2 involves performing full text extraction on a pre-configured set of static knowledge documents to obtain a static knowledge text sequence, and then matching the problem feature vector with a pre-configured dynamic case library to obtain a search case text sequence, including: Step S2.1: Traverse each document in the pre-configured static knowledge document set, perform text extraction operation on each document, and obtain the plain text content corresponding to each document; Step S2.2: Concatenate the plain text content corresponding to each document in document order to obtain a static knowledge text sequence; Step S2.3: Vectorize and encode each case text in the pre-configured dynamic case library to obtain a case vector set; Step S2.4: Calculate the cosine similarity between the problem feature vector and each case vector in the case vector set, sort them in descending order of similarity, and select the top K cases as the candidate case set; Step S2.5: Calculate the relevance score of each candidate case using the re-ranking model for the candidate case set, and filter out target cases whose relevance scores exceed a preset threshold. Step S2.6: Concatenate the case texts of the target case in descending order of relevance score to obtain the search case text sequence.

[0016] Beneficial effects of steps S2.1 to S2.6: This series of steps addresses the issues of insufficient accuracy in historical case retrieval due to reliance on manual review and keyword matching. By employing a hybrid strategy of extracting the full text of a static knowledge document set and semantic matching retrieval from a dynamic case library, it achieves a balance between the completeness of knowledge acquisition and the accuracy of retrieval, providing comprehensive and highly relevant reference texts for subsequent model reasoning.

[0017] In this process, step S2.1 involves extracting text from each document to convert knowledge documents of different formats into plain text content. Step S2.2 involves concatenating the plain text content of each document in document order to obtain a continuous static knowledge text sequence. Step S2.3 involves vectorizing each case text to map the semantic information of the case into numerical vectors, resulting in a case vector set. Step S2.4 involves calculating the cosine similarity between the question feature vector and each case vector and selecting the top K in descending order to achieve preliminary semantic recall of the cases. Step S2.5 involves calculating the relevance score using a re-ranking model and filtering out target cases that exceed a preset threshold, thus filtering out low-relevance cases to improve retrieval accuracy. Step S2.6 involves concatenating the target case texts in descending order of relevance score to obtain a retrieval case text sequence, providing highly relevant reference content for the prompt word assembly in the subsequent step S3.

[0018] As a further improvement to this application, step S3 involves concatenating the text feature vector, the question feature vector, the static knowledge text sequence, and the retrieval case text sequence into a structured input sequence according to a preset template. This structured input sequence is then input into a privately deployed language model, and the structured text sequence output by the privately deployed language model is obtained, including: Step S3.1: Concatenate the text feature vector and the question feature vector to form a question representation vector; Step S3.2: Add paragraph start markers to the static knowledge text sequence and the retrieval case text sequence respectively to obtain the marked knowledge text and the marked case text; Step S3.3: Concatenate the problem representation vector, the labeled knowledge text, and the labeled case text into a structured input sequence according to a preset template; Step S3.4: Input the structured input sequence into the privately deployed language model to obtain the structured text sequence generated by the privately deployed language model performing autoregressive decoding on the structured input sequence.

[0019] Beneficial effects of steps S3.1 to S3.4: This series of steps addresses the lack of automated fusion methods between multi-source knowledge text and the text to be processed. By concatenating text feature vectors, question feature vectors, static knowledge text sequences, and retrieval case text sequences according to a preset template into a structured input sequence and inputting it into a privately deployed language model, the standardized fusion of multi-source information and the automated execution of model inference are achieved, providing structured text output for subsequent format conversion.

[0020] Specifically, step S3.1 concatenates the text feature vector and the question feature vector to form a question representation vector, thereby achieving a unified representation of the text to be processed and the question description information; step S3.2 adds paragraph header markers to the static knowledge text sequence and the retrieval case text sequence respectively to achieve boundary identification of texts from different sources; step S3.3 concatenates the question representation vector with the marked knowledge text and case text according to a preset template to obtain a structured input sequence with a complete structure; step S3.4 inputs the structured input sequence into a privately deployed language model to perform autoregressive decoding, thereby achieving text generation under the condition of multi-source knowledge fusion and obtaining a structured text sequence.

[0021] As a further improvement to this application, step S4, extracting the heading level tags and table structure tags of the structured text sequence, and converting the structured text sequence into a target format document according to the heading level tags and table structure tags, includes: Step S4.1: Perform a marker scan on the structured text sequence to identify title markers and table markers; Step S4.2: Determine the document heading level corresponding to each heading based on the hierarchical nesting relationship of the heading markers; Step S4.3: Parse the number of rows, columns, and cell contents of the table based on the table identifiers to construct the table structure data; Step S4.4: Write the structured text sequence into the target format template according to the document title level and the table structure data to generate the target format document.

[0022] Beneficial effects of steps S4.1 to S4.4: This series of steps addresses the issue of manually formatting the output text of models. By extracting and converting the heading level markers and table structure markers of structured text sequences, it achieves automated formatting from model output to deliverable documents, eliminating the need for manual formatting adjustments.

[0023] Specifically, step S4.1 involves scanning the structured text sequence to identify heading and table markers, providing structural markers for subsequent format mapping; step S4.2 determines the document heading level based on the nested relationship of heading markers, achieving automatic heading level recognition; step S4.3 parses the number of rows, columns, and cell contents of the table based on table markers to construct table structure data, achieving automatic table structure restoration; and step S4.4 writes the structured text sequence into the target format template according to the document heading level and table structure data, generating the target format document and achieving automatic formatting of the output document.

[0024] As a further improvement to this application, step S5, in response to the manual confirmation mark of the target format document, takes the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark as a new case, performs vectorized encoding on the new case and writes it into the dynamic case library, triggering an incremental update of the dynamic case library, including: Step S5.1: Detect whether the target format document carries a manual confirmation mark. If it does, extract the audit text to be processed and the problem description text corresponding to the target format document. Step S5.2: Assemble the audit text to be processed and the problem description text according to a preset case format to obtain a new case; Step S5.3: Perform vectorization encoding on the newly added case to obtain the newly added case vector; Step S5.4: Write the newly added case vector into the dynamic case library, and perform an incremental index update on the dynamic case library.

[0025] Beneficial effects of steps S5.1 to S5.4: This series of steps addresses the lack of a closed-loop iterative mechanism for knowledge base updates. By responding to manual confirmation markers, newly added cases are vectorized and written into a dynamic case library, triggering incremental index updates, thereby achieving dynamic expansion of the case library and continuous improvement of retrieval coverage.

[0026] Specifically, step S5.1 detects whether the target format document carries a manual confirmation mark and extracts the corresponding audit text to be processed and the problem description text to achieve automatic identification of confirmation cases and recovery of original text; step S5.2 assembles the audit text to be processed and the problem description text according to a preset case format to obtain new cases, achieving standardized encapsulation of case data; step S5.3 performs vectorization encoding on the new cases to obtain new case vectors, providing vector data for index updates; step S5.4 writes the new case vectors into the dynamic case library and performs incremental index updates to achieve real-time expansion of the case library and dynamic expansion of the retrieval scope.

[0027] As a further improvement to this application, step S6 involves counting the number of newly added cases in the dynamic case library within a preset time window. When the number of newly added cases exceeds a preset iteration threshold, a full reconstruction of the dynamic case library is performed, and the fully reconstructed dynamic case library is used for the next round of matching retrieval, including: Step S6.1: Count the number of new cases added to the dynamic case library within the preset time window to obtain the new case count; Step S6.2: Compare the new case count with a preset iteration threshold to determine whether the new case count exceeds the preset iteration threshold; Step S6.3: If the number of newly added cases exceeds the preset iteration threshold, then all cases in the dynamic case library are re-vectorized to obtain a set of case vectors after full reconstruction. Step S6.4: Replace the existing case vectors in the dynamic case library with the fully reconstructed case vector set for the next round of matching retrieval.

[0028] Beneficial effects of steps S6.1 to S6.4: This series of steps addresses the issue of decreased retrieval accuracy caused by the cumulative deviation of the case vector index during incremental updates. By statistically analyzing the number of new cases added within a preset time window and performing a full reconstruction of the dynamic case library when the number exceeds an iteration threshold, the case vector index can be periodically optimized and retrieval accuracy can be continuously maintained.

[0029] Specifically, step S6.1 involves counting the number of new cases added to the dynamic case library within a preset time window to obtain a count of new cases, providing a quantitative basis for iteration determination; step S6.2 involves comparing the count of new cases with a preset iteration threshold to determine whether a full reconstruction is triggered, avoiding frequent or delayed reconstruction; step S6.3 involves re-encoding all cases in the dynamic case library using vectorization to obtain a set of case vectors after full reconstruction, eliminating the index bias accumulated from incremental updates; and step S6.4 involves replacing the existing case vectors with the set of case vectors after full reconstruction for use in the next round of matching retrieval, achieving periodic recovery and improvement of retrieval accuracy.

[0030] To achieve the above objectives, this application also provides the following technical solutions: A multi-dimensional intelligent processing system for audit text, wherein the multi-dimensional intelligent processing system is applied to the multi-dimensional intelligent processing method described above, and the multi-dimensional intelligent processing system includes: The audit text vector construction module is used to obtain the audit text to be processed and the problem description text, perform word segmentation and part-of-speech tagging on the audit text to be processed, extract noun words and verb words to construct text feature vectors, and perform the same operation on the problem description text to construct problem feature vectors; The retrieval case text sequence acquisition module is used to perform full text extraction on a pre-configured static knowledge document set to obtain a static knowledge text sequence, and to match and retrieve the problem feature vector with a pre-configured dynamic case library to obtain a retrieval case text sequence; The structured text sequence acquisition module is used to concatenate the text feature vector, the question feature vector, the static knowledge text sequence, and the retrieval case text sequence into a structured input sequence according to a preset template, input the structured input sequence into a privately deployed language model, and obtain the structured text sequence output by the privately deployed language model. The target format document conversion module is used to extract the heading level marks and table structure marks of the structured text sequence, and convert the structured text sequence into a target format document according to the heading level marks and table structure marks; The dynamic case library incremental update module is used to respond to the manual confirmation mark of the target format document, take the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark as a new case, perform vectorization encoding on the new case and write it into the dynamic case library, and trigger the incremental update of the dynamic case library. The dynamic case library full reconstruction module is used to count the number of new cases in the dynamic case library within a preset time window. When the number of new cases exceeds a preset iteration threshold, the dynamic case library is fully reconstructed and the reconstructed dynamic case library is used for the next round of matching retrieval.

[0031] To achieve the above objectives, this application also provides the following technical solutions: An electronic device includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the multi-dimensional intelligent processing method for audit text as described above.

[0032] To achieve the above objectives, this application also provides the following technical solutions: A computer-readable storage medium storing program instructions, which, when executed by a processor, enable the implementation of the multi-dimensional intelligent processing method for audit text as described above. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the steps of an embodiment of the multi-dimensional intelligent processing method for audit text according to this application. Figure 2 This is a schematic diagram of the functional modules of an embodiment of the multi-dimensional intelligent processing system for audit text according to this application; Figure 3 This is a schematic diagram of the structure of an embodiment of the electronic device of this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the storage medium of this application. Detailed Implementation

[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0035] The terms "first," "second," and "third" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (e.g., as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive having. For example, a process, method, system, product, or device having a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0036] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be present in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0037] It should be noted that, due to the limited types and number of symbols or letters that can represent specific meanings, for embodiments with many formulas or codes, there may be situations where symbols or letters cannot meet the usage requirements. Therefore, the interpretation of formula symbols in the steps or sub-steps of the embodiments is only valid for the current step or sub-step.

[0038] If the same symbol has different interpretations in different steps or sub-steps, the interpretation in the current step or sub-step shall prevail; if the same symbol appears in different steps or sub-steps, but no interpretation is given in subsequent steps or sub-steps after its first appearance, the interpretation in the first step or sub-step shall be used.

[0039] For example, "i=1,2,...,n" is a writing convention and a well-known meaning. If "i=1,2,...,n" has already appeared once in an embodiment and needs to be used again in formulas with different scenarios and meanings in the future, then the meaning of "i=1,2,...,n" should be interpreted according to the corresponding formulas. If "i=1,2,...,n" is not used for the first time and is changed to a symbol or letter with a non-well-known meaning, such as "p=1,2,...,q", to avoid repetition, it is more likely to cause confusion, ambiguity, and unclear problems.

[0040] like Figure 1 As shown, this embodiment provides an example of a multi-dimensional intelligent processing method for audit text. In this embodiment, the multi-dimensional intelligent processing method includes the following steps: Step S1: Obtain the audit text to be processed and the problem description text. Perform word segmentation and part-of-speech tagging on the audit text to be processed, extract noun words and verb words to construct text feature vectors, and perform the same operation on the problem description text to construct problem feature vectors.

[0041] Furthermore, step S1 specifically includes the following steps: Step S1.1: Obtain the audit text to be processed and the problem description text, perform word segmentation on the audit text to be processed, and obtain the word segmentation result.

[0042] Preferably, the core of this step is to complete the word segmentation of the audit text to be processed. By dividing the continuous natural language text into independent lexical units, it provides discretized input for subsequent part-of-speech tagging and vectorized encoding.

[0043] Among them, pending audit documents refer to the rectification measures documents submitted by the responsible departments to the audit department. The documents describe the problem response plan, division of responsibilities, and timeline in natural language. Pending audit documents are denoted as The text length is denoted as , in units of character count.

[0044] The problem description text refers to the problem description information recorded by auditors in their audit findings. The problem description text is denoted as... The text length is denoted as .

[0045] For word segmentation, the Jieba word segmentation tool is used to load a custom audit domain dictionary. ,right Perform word segmentation, and record the segmentation result as a word sequence. ,in for The total number of words after word segmentation For the first One word.

[0046] Preferably, a custom audit domain dictionary The total number of proprietary terms in areas such as audit terminology, regulation names, rectification categories, and responsible entities is [number missing]. The word segmentation mode uses the precise mode, and stop word filtering is enabled. .

[0047] For example, audit text to be processed The content, "Regarding the issue of irregular reimbursement of meeting expenses, the finance department has completed the revision of the system and organized training for all employees," is segmented into words. Regarding the issue of irregularities in meeting expense reimbursement, the finance department has completed the revision of the relevant regulations and organized training for all staff. Total number of words The corresponding record identifier is Audit-YYYY-MM-DD-HH:MM:SS-001.

[0048] Step S1.2: Perform part-of-speech tagging on the word segmentation results, filter out noun words and verb words, and obtain a candidate word set.

[0049] Preferably, the core of this step is to complete the part-of-speech tagging and vocabulary selection of the word segmentation results. By retaining noun words and verb words and removing functional word noise, a semantically carrying vocabulary set is provided for subsequent vectorized encoding.

[0050] Among them, for part-of-speech tagging, the word segmentation results Chinese vocabulary Perform part-of-speech tagging, and record the tagging results as follows: ,in For vocabulary Part-of-speech tags.

[0051] For vocabulary filtering, the filtering rule is to retain words whose part-of-speech tags belong to the noun vocabulary set. n,nr, ns, nt, nz, vn, j With verbal vocabulary v, vd, vn, vyou The vocabulary was filtered to obtain a candidate vocabulary set. ,in To determine the number of words after filtering. .

[0052] Preferably, a set of noun words Includes common nouns (n), personal names (nr), place names (ns), organization names (nt), other proper nouns (nz), verbal nouns (vn) and abbreviations (j), a collection of verbal vocabulary. It includes the common verb v, the effusive verb vd, the noun verb vn, and the verb vyou.

[0053] For example, word segmentation results (Regarding, p), (Meeting expenses, n), (Reimbursement, v), (Irregularities, a), (Problems, n), (Finance Department, nt), (Already, d), (Completed, v), (System, n), (Revised, vn), (And, c), (Organization, v), (All staff, n), (Training, vn) The set of candidate words after filtering Meeting expenses, reimbursement, issues, finance department, completion, system, revision, organization, all staff, training , .

[0054] Step S1.3: Vectorize each word in the candidate word set to obtain a word vector sequence.

[0055] Preferably, the core of this step is to complete the vectorization encoding of candidate words, which maps discrete words into continuous numerical vectors to provide vector-level input for subsequent mean pooling.

[0056] For vectorized encoding, a pre-trained word vector model is used. For candidate word set Chinese vocabulary Perform vectorization encoding to obtain a sequence of word vectors. ,in , The dimension is the word vector.

[0057] Among them, for pre-trained word vector models This paper employs a pre-trained Chinese language model based on the Transformer architecture to extract word-level vectors. The model input is a sequence of characters for a single word, and the output is the corresponding word's... The word vector is a dense vector, and for words that do not exist in the vocabulary of the word vector model, the vector corresponding to the UNK special tag is used as a fallback encoding.

[0058] Preferably, word vector dimension The UNK vector is initialized with a mean of 1. variance is The random vectors and word vector model are deployed on a local inference service, with the inference interface address denoted as [address to be filled in]. .

[0059] For example, candidate word set The Middle The term "meeting expenses" was vectorized and encoded to obtain... ,in , , Vocabulary vector sequence Sequence length The corresponding record identifier is Embed-YYYY-MM-DD-HH:MM:SS-001.

[0060] Step S1.4: Perform mean pooling on the word vector sequence to obtain the text feature vector of the audit text to be processed.

[0061] Preferably, the core of this step is to perform mean pooling of the word vector sequence, which compresses the variable-length vector sequence into a fixed-dimensional vector to obtain the overall semantic representation of the audit text to be processed.

[0062] For mean pooling, the word vector sequence The calculation formula is as follows: (The formula is missing from the provided text.) .

[0063] in, The text feature vector of the audit text to be processed. The dimensional components are .

[0064] Preferably, the vector dimension after mean pooling is... The pooling strategy is to directly take the arithmetic average without introducing weights or attention mechanisms.

[0065] For example, word vector sequences , After mean pooling, the result is obtained ,in , The text feature vector record is identified as Vec-Audit-YYYY-MM-DD-HH:MM:SS-001.

[0066] Step S1.5: Perform the same word segmentation, part-of-speech tagging, vectorization encoding, and mean pooling operations on the problem description text as on the audit text to be processed, to obtain the problem feature vector.

[0067] Preferably, the core of this step is to extract features from the problem description text and construct a problem feature vector through the same processing procedure as the audit text to be processed, so as to provide a retrieval basis for the matching retrieval in the subsequent step S2.

[0068] Specifically, the problem description text processing follows the same procedure from steps S1.1 to S1.4. The word segmentation process is performed sequentially to obtain the results. Part-of-speech tagging and filtering were obtained Vectorized encoding Mean pooling yields the feature vector of the problem. The calculation formula is: .

[0069] in, for Number of candidate words after filtering. for No. The word vectors of each candidate word. .

[0070] Preferably, the problem description text processing parameters are completely consistent with the audit text to be processed, including the word segmentation dictionary. Part-of-speech tag filtering set and Word vector model Word vector dimension .

[0071] For example, the problem description text The content is "Does the meeting expense reimbursement process comply with financial management regulations?", and the candidate keywords are selected after word segmentation and part-of-speech tagging. Meeting expenses, reimbursement process, complies with financial management regulations. , The problem feature vector is obtained after vectorization encoding and mean pooling. ,in , The problem feature vector record is identified as Vec-Question-YYYY-MM-DD-HH:MM:SS-001.

[0072] Beneficial effects of steps S1.1 to S1.5: This series of steps addresses the lack of automated feature extraction methods for audit texts. By using word segmentation, part-of-speech tagging, and vectorized encoding to construct text feature vectors and question feature vectors, the audit text is automatically converted from unstructured natural language to structured numerical representation, providing standardized input for subsequent matching retrieval and model inference.

[0073] In this process, step S1.1 divides the continuous text into independent word units through word segmentation to obtain the word segmentation results; step S1.2 filters noun and verb words through part-of-speech tagging and removes functional word noise to obtain a candidate word set; step S1.3 maps discrete words to continuous numerical vectors by vectorizing each word to obtain a word vector sequence; step S1.4 compresses the variable-length sequence into a fixed-dimensional vector by performing mean pooling on the word vector sequence to obtain the text feature vector; and step S1.5 performs the same processing flow on the question description text to obtain the question feature vector, providing a retrieval basis for the matching retrieval in the subsequent step S2.

[0074] Step S2: Perform full text extraction on the pre-configured static knowledge document set to obtain a static knowledge text sequence, and match the problem feature vector with the pre-configured dynamic case library to obtain a retrieval case text sequence.

[0075] Furthermore, step S2 specifically includes the following steps: Step S2.1: Traverse each document in the pre-configured static knowledge document set, perform text extraction operation on each document, and obtain the plain text content corresponding to each document.

[0076] Preferably, the core of this step is to traverse the static knowledge document set and extract the text. By uniformly converting knowledge documents of different formats into plain text content, it provides uniformly formatted text data for subsequent splicing.

[0077] Among them, the static knowledge document set refers to the pre-configured audit domain knowledge document set. ,in The total number of documents is [number], and the document formats include DOCX, PDF, and TXT. Each document contains guidelines for problem categorization, evaluation rules, and a summary of frequently asked questions. The storage path for this document collection is denoted as [path]. .

[0078] For text extraction, the document format type is considered. Use the corresponding text extraction method: when The text content is extracted paragraph by paragraph using the python-docx library; when The text content is extracted page by page using the pdfplumber library; when At that time, the full text is read directly, and the plain text content of each document is extracted and recorded as follows: .

[0079] Preferably, a static knowledge document collection Storage path This is a local knowledge base directory. After extracting the text, newline characters and extra spaces are removed, and the text is unified into a single continuous line.

[0080] For example, static knowledge document collections Include The document, number One document The document is titled "Guidelines for Classifying Audit Issues.docx". The format is... After extracting the text paragraph by paragraph using python-docx, the plain text content is obtained. Text length The character corresponds to the record identifier Static-DOC-001.

[0081] Step S2.2: Concatenate the plain text content corresponding to each document in document order to obtain a static knowledge text sequence.

[0082] Preferably, the core of this step is to sequentially concatenate the plain text content of each document. By concatenating the documents in order into a continuous text sequence, a complete static knowledge background is provided for the subsequent assembly of the structured input sequence.

[0083] For sequential concatenation, the plain text content of each document is concatenated according to the document index order, with separator marks inserted between documents. The splicing formula is: .

[0084] in, For text concatenation operators For the concatenated static knowledge text sequence, delimiters are used. This is a pre-defined special string used to identify document boundaries.

[0085] Preferably, document separator mark The resulting static knowledge text sequence Maximum length limit is Characters; if the text exceeds the specified limit, the last part of the text will be truncated in document order.

[0086] For example, The plain text content of each document is concatenated sequentially, the first... One document splicing the first One document Insert between Finally, a static knowledge text sequence is obtained. Sequence length Characters, not exceeding The upper limit corresponds to the record identifier Seq-Static-YYYY-MM-DD-HH:MM:SS-001.

[0087] Step S2.3: Vectorize and encode each case text in the pre-configured dynamic case library to obtain a case vector set.

[0088] Preferably, the core of this step is to complete the vectorization encoding of each case text in the dynamic case library. By mapping the semantic information of the cases into numerical vectors, a vector-level retrieval basis is provided for subsequent cosine similarity calculation.

[0089] The dynamic case library refers to a vector retrieval library that stores historical audit cases. Each case in the case library contains a case identifier. Case text Case Vector The total number of cases in the current case library is denoted as .

[0090] Specifically, for vectorized encoding, the text of each case in the dynamic case library... The same word vector model as in step S1.3 is used. Generate case vectors with mean pooling operation The encoding method is consistent with the audit text to be processed, ensuring the consistency of the vector space and obtaining a set of case vectors. .

[0091] Preferably, the case vector dimension Consistent with the dimension of text feature vectors, the initial number of cases in the dynamic case library The case vector index is built using the Faiss library, and the index type is plane exact search IndexFlatIP.

[0092] For example, dynamic case libraries The Middle Case Case text The content is "non-compliant invoices discovered during travel expense reimbursement review," which is then vectorized to obtain the case vector. ,in , Case vector set Total of Each case vector corresponds to a record identifier of Case-Vec-YYYY-MM-DD-HH:MM:SS-001.

[0093] Step S2.4: Calculate the cosine similarity between the problem feature vector and each case vector in the case vector set, sort them in descending order of similarity, and select the top K cases as the candidate case set.

[0094] Preferably, the core of this step is to calculate and initially screen the similarity between the problem feature vector and each case vector, and select the top K cases by ranking them by cosine similarity to achieve preliminary semantic recall of the cases.

[0095] Among them, for cosine similarity calculation, the problem feature vector Case Vector The formula for calculating cosine similarity is: .

[0096] in, For vector dot product, for Norm, .

[0097] For the initial screening and ranking, the cosine similarity of each case is used. Sort in descending order and select the top... The cases constitute the candidate case set subscript This indicates the similarity score after descending order. Case study of position.

[0098] Preferably, the initial screening quantity The similarity calculation uses the batch cosine similarity interface of the Faiss library, and the calculation precision is double-precision floating-point numbers.

[0099] For example, problem feature vectors and Calculate the cosine similarity of each case vector, and sort them in descending order. The similarity of the cases is as follows: , , , , Candidate Case Set Include Case number 1 Case The corresponding record identifier is Cand-YYYY-MM-DD-HH:MM:SS-001.

[0100] Step S2.5: Call the reordering model to calculate the relevance score of each candidate case in the candidate case set, and filter the target cases whose relevance scores exceed the preset threshold.

[0101] Preferably, the core of this step is to refine the candidate case set. By using a re-ranking model to calculate the relevance score of each candidate case and filtering out target cases that exceed the threshold, low-relevance cases remaining from the initial screening stage are removed, thereby improving retrieval accuracy.

[0102] For the reordering model, a pre-trained cross-encoder model is used. The model input is the problem description text. With candidate case text The concatenated text pairs are output as relevance scores by the model. The calculation formula is: .

[0103] in, For the Sigmoid function, For cross encoder models, For the first The relevance score of each candidate case.

[0104] For threshold screening, the relevance score of each candidate case is used. With preset threshold Compare and filter The cases constitute the target case set .

[0105] Preferably, the reordering model Using the bce-reranker-base model, the maximum input sequence length is... Each token, relevance score threshold .

[0106] For example, candidate case set middle The candidate cases were re-ranked using a relevance score calculation model, among which... , , , , ,filter Target cases Article, No. Case Filtered out, target case set The corresponding record identifier is Target-YYYY-MM-DD-HH:MM:SS-001.

[0107] Step S2.6: Concatenate the case texts of the target case in descending order of relevance score to obtain the search case text sequence.

[0108] Preferably, the core of this step is to sequentially splice the target case text, and then splice it into a continuous text sequence by descending order of relevance score, so as to provide highly relevant reference content for the structured input sequence assembly in the subsequent step S3.

[0109] For sequential concatenation, the target case set is... The texts of each case study are ranked by relevance score. Arrange the cases in descending order and then concatenate them sequentially, inserting separators between them. The splicing formula is: .

[0110] in, The first one is sorted in descending order of relevance score. Case text for each target case, The total number of target cases To retrieve the case text sequence, Use case separators.

[0111] Preferably, case separator marker The concatenated search case text sequence Maximum length limit is Characters, if exceeding the limit, the last part of the content is truncated in descending order of relevance score.

[0112] For example, the target case set common The target cases are concatenated in descending order of relevance score, starting with the [number]th case. Case The content is "Rectification Measures for Non-compliant Receipts in Travel Expense Reimbursement Audits," No. Case The content is a "Rectification Plan for Optimizing the Expense Reimbursement Approval Process," with an insert in the middle. Finally, the retrieved case text sequence was obtained. Sequence length Characters, not exceeding The upper limit corresponds to the record identifier Seq-Retrieval-YYYY-MM-DD-HH:MM:SS-001.

[0113] Beneficial effects of steps S2.1 to S2.6: This series of steps addresses the issues of insufficient accuracy in historical case retrieval due to reliance on manual review and keyword matching. By employing a hybrid strategy of extracting the full text of a static knowledge document set and semantic matching retrieval from a dynamic case library, it achieves a balance between the completeness of knowledge acquisition and the accuracy of retrieval, providing comprehensive and highly relevant reference texts for subsequent model reasoning.

[0114] In this process, step S2.1 involves extracting text from each document to convert knowledge documents of different formats into plain text content. Step S2.2 involves concatenating the plain text content of each document in document order to obtain a continuous static knowledge text sequence. Step S2.3 involves vectorizing each case text to map the semantic information of the case into numerical vectors, resulting in a case vector set. Step S2.4 involves calculating the cosine similarity between the question feature vector and each case vector and selecting the top K in descending order to achieve preliminary semantic recall of the cases. Step S2.5 involves calculating the relevance score using a re-ranking model and filtering out target cases that exceed a preset threshold, thus filtering out low-relevance cases to improve retrieval accuracy. Step S2.6 involves concatenating the target case texts in descending order of relevance score to obtain a retrieval case text sequence, providing highly relevant reference content for the prompt word assembly in the subsequent step S3.

[0115] Step S3: Concatenate the text feature vector, question feature vector, static knowledge text sequence, and retrieval case text sequence into a structured input sequence according to a preset template. Input the structured input sequence into the privately deployed language model and obtain the structured text sequence output by the privately deployed language model.

[0116] Furthermore, step S3 specifically includes the following steps: Step S3.1: Concatenate the text feature vector and the question feature vector to form the question representation vector.

[0117] Preferably, the core of this step is to concatenate the text feature vector and the problem feature vector. By merging the semantics of the audit text to be processed and the semantics of the problem description into a unified representation, the problem side vector input is provided for the subsequent assembly of the structured input sequence.

[0118] Specifically, for vector concatenation, the text feature vectors are... With the problem feature vector By concatenating the last two digits, we obtain the problem representation vector. The splicing formula is: .

[0119] in, This is the vector concatenation operator. The former Weiwei ,back Weiwei .

[0120] Preferably, the dimension of the problem representation vector The concatenation order is text feature vector first, followed by question feature vector, without introducing an additional projection layer.

[0121] For example, text feature vectors With the problem feature vector The concatenation yields the problem representation vector. ,in , The corresponding record identifier is Query-Vec-YYYY-MM-DD-HH:MM:SS-001.

[0122] Step S3.2: Add paragraph start markers to the static knowledge text sequence and the retrieval case text sequence respectively to obtain the marked knowledge text and the marked case text.

[0123] Preferably, the core of this step is to mark the beginning of the knowledge text and the case text. By adding distinguishing markers to texts from different sources, the boundary recognition of different text paragraphs can be achieved during subsequent model inference.

[0124] Among them, for the paragraph beginning marker, in the static knowledge text sequence Add a knowledge section start marker at the beginning Obtain the labeled knowledge text ; in retrieving case text sequences Add a case paragraph start marker at the beginning The marked case text is obtained. .

[0125] Among them, regarding the role of the marker, the paragraph start marker is a pre-defined special string that serves as a paragraph start signal during the subsequent private deployment of the language model inference process, helping the model distinguish the source boundary between knowledge text paragraphs and case text paragraphs.

[0126] Preferably, the knowledge segment start marker Case paragraph start marker The marker and the text are separated by a newline character.

[0127] For example, static knowledge text sequences Adding paragraph start markers results in the marked knowledge text. Retrieve case text sequences Adding paragraph start markers results in the marked example text. The corresponding record identifier is Tag-Text-YYYY-MM-DD-HH:MM:SS-001.

[0128] Step S3.3: Concatenate the problem representation vector, the labeled knowledge text, and the labeled case text into a structured input sequence according to the preset template.

[0129] Preferably, the core of this step is to complete the templated assembly of the structured input sequence. By splicing the problem representation vector with the labeled knowledge text and case text according to the preset template, a complete sequence is formed, providing structured input for subsequent model reasoning.

[0130] For the preset template, the template structure consists of four fields arranged in a fixed order: system instruction field Problem characterization fields Knowledge text field With case text field Structured input sequence The assembly formula is: .

[0131] in, This is the preset system command text. This indicates that the input content will be filled into the corresponding template field position.

[0132] Among them, for system instruction text It includes role settings, task descriptions, and output format requirements, instructing the model to generate structured text, including title level markers and table structure markers, based on the input problem representation vector, knowledge text, and case text.

[0133] Preferably, the system instruction field in the preset template has a fixed length. Characters, structured input sequence Maximum length limit is When the number of characters exceeds the limit, the content at the end will be truncated in the order of knowledge text field first, followed by case text field.

[0134] For example, system instruction text The content is: "You are an audit rectification evaluation expert. Based on the following problem description and references, please generate a rectification evaluation document including heading levels and table structure." Problem representation vector. Enter the problem representation field, followed by the tagged knowledge text. Enter the knowledge text field, followed by the marked case text. Enter the example text field and assemble the structured input sequence. Total length The character corresponds to the record identifier Seq-Input-YYYY-MM-DD-HH:MM:SS-001.

[0135] Step S3.4: Input the structured input sequence into the privately deployed language model and obtain the structured text sequence generated by the privately deployed language model performing autoregressive decoding on the structured input sequence.

[0136] Preferably, the core of this step is to complete the model inference of the structured input sequence, and to perform autoregressive decoding by privately deploying the language model to achieve structured text generation under the condition of multi-source knowledge fusion, thereby obtaining a structured text sequence.

[0137] Among them, the privately deployed language model refers to the large language model deployed on a local inference server. The model uses an autoregressive decoding method to generate output text token by token, given a structured input sequence. The model in each decoding step Based on the historical output token sequence Calculate the probability distribution of the next token using the input sequence: .

[0138] in, For the first The token generated step by step, The softmax function is used to sample each token according to a probability distribution or to generate each token with the highest probability, until the end token is generated. All generated tokens are concatenated into a structured text sequence. .

[0139] For private deployments, the model Deployed on a local GPU inference server, the inference interface address is denoted as The deployment method is containerized, the model weights are stored locally, and the inference process does not rely on external API calls, ensuring data security and inference controllability.

[0140] Preferably, the language model is deployed privately. Parameters Maximum input sequence length One token, maximum generation length A token, decoded temperature top-p sampling parameters Repeat penalty coefficient .

[0141] For example, structured input sequences Input to private deployment language model The model generates output token by token in an autoregressive manner. During the generation process, the first... step token The corresponding word is "#" (first-level heading marker), the first... step token The corresponding term is "rectification". step token The corresponding word "evaluation" is used until an end marker is generated. All tokens are concatenated into a structured text sequence. Sequence length The character corresponds to the record identifier Seq-Output-YYYY-MM-DD-HH:MM:SS-001.

[0142] Beneficial effects of steps S3.1 to S3.4: This series of steps addresses the lack of automated fusion methods between multi-source knowledge text and the text to be processed. By concatenating text feature vectors, question feature vectors, static knowledge text sequences, and retrieval case text sequences according to a preset template into a structured input sequence and inputting it into a privately deployed language model, the standardized fusion of multi-source information and the automated execution of model inference are achieved, providing structured text output for subsequent format conversion.

[0143] Specifically, step S3.1 concatenates the text feature vector and the question feature vector to form a question representation vector, thereby achieving a unified representation of the text to be processed and the question description information; step S3.2 adds paragraph header markers to the static knowledge text sequence and the retrieval case text sequence respectively to achieve boundary identification of texts from different sources; step S3.3 concatenates the question representation vector with the marked knowledge text and case text according to a preset template to obtain a structured input sequence with a complete structure; step S3.4 inputs the structured input sequence into a privately deployed language model to perform autoregressive decoding, thereby achieving text generation under the condition of multi-source knowledge fusion and obtaining a structured text sequence.

[0144] Step S4: Extract the heading level tags and table structure tags from the structured text sequence, and convert the structured text sequence into a target format document according to the heading level tags and table structure tags.

[0145] Furthermore, step S4 specifically includes the following steps: Step S4.1: Perform a marker scan on the structured text sequence to identify title markers and table markers.

[0146] Preferably, the core of this step is to scan and recognize the markers of the structured text sequence, and to provide structural marker information for subsequent format mapping by distinguishing between title markers and table markers.

[0147] Specifically, for the marker scanning, the structured text sequence is traversed character by character from front to back, scanning the set of title markers and the set of table markers. The number of hash symbols in the heading markers corresponds to the heading level, and the number of hash symbols corresponds to the first level heading. The table marker "|" is used to identify table cell separators.

[0148] Among these, for the recognition rules, each marker is recorded during the scanning process. Character position in With tag type For title markers, also record the number of hash symbols. As hierarchical information, the scan results are denoted as the marker sequence. ,in The total number of identified tags, corresponding to the table tags. .

[0149] Preferably, the title level supports This refers to headings from level one to level four. Table markers are used to identify rows, and consecutive rows starting with "|" are considered table rows.

[0150] For example, structured text sequences The scan identified the title marker "#rectification evaluation" at the specified character position. hierarchical The title marker "##I. Problem Analysis" is located at the character position. hierarchical The table marker "|Serial Number|Problem Description|Corrective Measures|" is located at the character position. hierarchical , tag sequence Common identification Each marker corresponds to a record identifier of Mark-Scan-YYYY-MM-DD-HH:MM:SS-001.

[0151] Step S4.2: Determine the document heading level corresponding to each heading based on the hierarchical nesting relationship of the heading markers.

[0152] Preferably, the core of this step is to determine the mapping of heading levels. By mapping the number of hash symbols to document heading styles, the heading levels can be automatically identified.

[0153] For hierarchical nesting relationships, based on the tag sequence Hierarchy of heading tags The corresponding document title style is determined, and the mapping relationship is as follows: Corresponding to the first-level heading style (Centered, Bold, Font Size) pound), Corresponding second-level heading style (Left-aligned, bold, font size) pound), Corresponding to third-level heading style (Left-aligned, bold, font size) pound), Corresponding to level 4 heading style (Left alignment, regular, font size) pound).

[0154] For the nesting check, after determining the level of each heading, a nesting rationality check is performed on adjacent headings, and the criterion is that the difference between the levels of adjacent headings does not exceed a certain threshold. ,Right now If the level exceeds the limit, it is considered a level jump exception and a warning log is logged. The processing flow is not interrupted.

[0155] Preferably, the title style mapping is fixed as follows: Level, document title style to The font sizes are as follows: pound, pound, pound, pound.

[0156] For example, the first in the labeled sequence Heading markup levels Mapped to The title text is "Rectification Evaluation"; the first... Heading markup levels Mapped to The title text is "I. Problem Analysis"; Heading markup levels Mapped to The title text is "II. Rectification Measures", and the hierarchical difference between adjacent titles is... The verification passed, and the corresponding record is identified as Heading-Map-YYYY-MM-DD-HH:MM:SS-001.

[0157] Step S4.3: Parse the number of rows, columns, and cell contents of the table based on the table tags to construct the table structure data.

[0158] Preferably, the core of this step is to parse the table markers and reconstruct the table structure. By parsing the number of rows, columns, and cell contents, the table structure can be automatically restored.

[0159] For table parsing, from the tag sequence Extract the line text corresponding to all table markers, group them by consecutive lines, and form a table for each group. ,in For each table, use the table number as the reference. The table is split into cells using the "|" separator. The first row is the header row, and subsequent rows are the data rows, indicating the number of rows. This represents the total number of rows and columns in the table. This represents the maximum number of cells in the table.

[0160] For table-structured data, each table Constructed as a two-dimensional data structure ,in For the first Line 1 The text content of each column cell is aligned with the number of columns during construction, and rows with fewer than the required number of columns are padded with empty strings.

[0161] Preferably, leading and trailing spaces are removed from cell text, and table-separating rows (rows containing only "---" or ":---") are skipped during parsing and not counted in the row count. .

[0162] For example, the first in the labeled sequence To the The markers form consecutive table rows, constituting the first... Table The first row, "|Serial Number|Problem Description|Rectification Measures|", is the header row. It is split using the "|" to obtain the table. The column contains cells with the content ["Serial Number", "Problem Description", "Rectification Measures"]. (Follow-up...) Behavioral data rows, number of table rows , No. Line 1 List Table structure data The corresponding record identifier is Table-Parse-YYYY-MM-DD-HH:MM:SS-001.

[0163] Step S4.4: Write the structured text sequence into the target format template according to the document title level and table structure data to generate the target format document.

[0164] Preferably, the core of this step is to write the structured text sequence into the target format document. By writing the target format template according to the heading level style and table structure data, the automatic formatting of the output document is achieved.

[0165] Specifically, the target format template is a DOCX format template. The template includes default body text styles. (SimSun font, size) Points, line spacing (times), written according to structured text sequence The characters are traversed in order, and when a title marker is encountered, it is matched according to the mapped title style. When writing heading paragraphs, data should be written according to table structure when table markers are encountered. Write the text into the table object, and style the rest of the text as body text. Write it into the main text paragraph.

[0166] Specifically, for writing table objects, each table Write as a DOCX table object, with the header row bold and centered, data rows left-aligned, and cell margins set. Centimeters, table border is a single solid line pound.

[0167] Preferably, the target format template Preset top and bottom margins centimeters, around centimeters, text style For Song typeface pt, heading style to The font sizes are as follows: pound, pound, pound, points, table border Solid line on the weighbridge slip.

[0168] For example, structured text sequences Write to the target format template in character order. "#Rectification Evaluation" should be written as a first-level heading paragraph (centered, bold). (pt), "##I. Problem Analysis" should be written as a second-level heading paragraph (left-aligned, bold). (pounds), table Write as OK List DOCX table objects, write the remaining text as body paragraphs, and finally generate the target format document. File size The corresponding record identifier is Doc-Output-YYYY-MM-DD-HH:MM:SS-001.

[0169] Beneficial effects of steps S4.1 to S4.4: This series of steps addresses the issue of manually formatting the output text of models. By extracting and converting the heading level markers and table structure markers of structured text sequences, it achieves automated formatting from model output to deliverable documents, eliminating the need for manual formatting adjustments.

[0170] Specifically, step S4.1 involves scanning the structured text sequence to identify heading and table markers, providing structural markers for subsequent format mapping; step S4.2 determines the document heading level based on the nested relationship of heading markers, achieving automatic heading level recognition; step S4.3 parses the number of rows, columns, and cell contents of the table based on table markers to construct table structure data, achieving automatic table structure restoration; and step S4.4 writes the structured text sequence into the target format template according to the document heading level and table structure data, generating the target format document and achieving automatic formatting of the output document.

[0171] Step S5: In response to the manual confirmation mark on the target format document, the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark are added as new cases. The new cases are vectorized and encoded and written into the dynamic case library, triggering the incremental update of the dynamic case library.

[0172] Furthermore, step S5 specifically includes the following steps: Step S5.1: Detect whether the target format document carries a manual confirmation mark. If it does, extract the audit text to be processed and the problem description text corresponding to the target format document.

[0173] Preferably, the core of this step is to detect the confirmation mark of the target format document and collect the original text. By detecting the manual confirmation mark, the reviewed documents are screened and the corresponding original input text is extracted, providing a data source for the assembly of new cases in the future.

[0174] Among them, the manual confirmation mark refers to the mark made by the reviewer on the target format document. Added confirmation flag field ,in This indicates that the message has not been confirmed. This indicates that the confirmation has been received. The confirmation flag is stored in the document metadata and can be obtained by reading the document metadata fields.

[0175] Among them, for the recycling of original text, when At that time, according to the target format document Document identifier From the cache of audit text to be processed Search for the corresponding audit text to be processed. With the problem description text The cache stores data in key-value pairs. .

[0176] Preferably, the confirmation flag field The default value is After the reviewer clicks the confirmation button on the document review interface, the mark is set to [value]. Pending audit text cache Capacity limit If the number of records exceeds the limit, the earliest record will be eliminated using a first-in, first-out (FIFO) strategy.

[0177] For example, target format documents Document identifier Read metadata fields to confirm the flag From cache The corresponding search results were found in the middle. and , The content states, "Regarding the issue of irregularities in meeting expense reimbursement, the finance department has completed the revision of the policy and organized training for all employees." The content is "Does the meeting expense reimbursement process comply with financial management regulations?", and the corresponding record identifier is Confirm-YYYY-MM-DD-HH:MM:SS-001.

[0178] Step S5.2: Assemble the audit text to be processed and the problem description text according to the preset case format to obtain a new case.

[0179] Preferably, the core of this step is to encapsulate the original text into cases. By assembling the audit text to be processed and the problem description text into standardized cases according to a preset case format, a uniformly formatted case data is provided for subsequent vectorized encoding.

[0180] For the preset case format, the case format template includes a case identifier field, a problem description field, and a rectification measure field. Adding a new case... The assembly formula is: .

[0181] in, Format the function for the case. To add a case identifier, The generation rules are as follows ,in For the current date, For the serial number of the day Initially increasing.

[0182] Among them, for the case formatting function , describe the problem text Enter the problem description field and specify the audit text to be processed. Enter the rectification measures field and the case identifier field. Separate the three fields with markers. The link is for a single case text.

[0183] Preferably, field separator mark Case identifiers are generated using a date plus serial number rule, with the serial number having a width of [number missing]. Bit.

[0184] For example, audit text to be processed With the problem description text Assemble according to the preset case format and add case identifiers. Case text The content is "Case-20260818-0001 [FIELD_SEP] Does the meeting expense reimbursement process comply with financial management regulations? [FIELD_SEP] Regarding the issue of irregular meeting expense reimbursement, the finance department has completed the revision of the system and organized training for all employees." The corresponding record is identified as New-Case-YYYY-MM-DD-HH:MM:SS-001.

[0185] Step S5.3: Perform vectorization encoding on the newly added cases to obtain the newly added case vector.

[0186] Preferably, the core of this step is to complete the vectorization encoding of the new cases, generating vectors for the new cases using the same encoding method as the existing cases, thus providing vector data for writing into the dynamic case library.

[0187] For vectorized encoding, for newly added cases The same processing procedure as steps S1.3 to S1.4 is adopted, that is, for Perform word segmentation, part-of-speech tagging and filtering, and word vector modeling. Vectorized encoding and mean pooling are used to obtain the vector of new cases. The calculation formula is: .

[0188] in, for Number of candidate words after filtering. For the first Word vectors of candidate words.

[0189] Preferably, the vectorized encoding parameters of the newly added cases are completely consistent with those of the existing cases, including the word segmentation dictionary. Part-of-speech tag filtering set and Word vector model Word vector dimension .

[0190] For example, new cases Number of candidate words after word segmentation and part-of-speech tagging The newly added case vector is obtained after vectorization encoding and mean pooling. ,in , The corresponding record identifier is New-Vec-YYYY-MM-DD-HH:MM:SS-001.

[0191] Step S5.4: Write the newly added case vectors into the dynamic case library and perform an incremental index update on the dynamic case library.

[0192] Preferably, the core of this step is to write the newly added case vectors into the database and update the index. By writing the newly added case vectors into the dynamic case database and triggering incremental index reconstruction, the case database can be expanded in real time and the retrieval scope can be dynamically expanded.

[0193] Specifically, for vector writing, a new case vector will be added. Along with case label With case text Write to dynamic case library The write operation includes: Add to the set of case vectors ,Will and Add to the case metadata table to update the total number of cases in the case library. .

[0194] Specifically, for incremental index updates, an incremental update of the Faiss index is triggered after the vector is written. The incremental update method is to... Adding to the existing IndexFlatIP index, since IndexFlatIP is a planar index and does not require rebuilding the training step, the incremental addition operation directly performs vector appending. After the update, the number of vectors in the index is the same as the new index. Consistent.

[0195] Preferably, the upper limit of cases in the dynamic case library When the upper limit is reached, the oldest case vector is removed using the earliest case elimination strategy. Index updates adopt a synchronous write method, and the changes take effect immediately after the write is completed and can be used in the next round of retrieval.

[0196] For example, adding case vectors Write to dynamic case library Case Identifier Total number of cases before writing After writing Faiss index incremental update append The number of index vectors is then updated synchronously. The next round of matching and retrieval will then find the newly added case, with the corresponding record identifier being DB-Update-YYYY-MM-DD-HH:MM:SS-001.

[0197] Beneficial effects of steps S5.1 to S5.4: This series of steps addresses the lack of a closed-loop iterative mechanism for knowledge base updates. By responding to manual confirmation markers, newly added cases are vectorized and written into a dynamic case library, triggering incremental index updates, thereby achieving dynamic expansion of the case library and continuous improvement of retrieval coverage.

[0198] Specifically, step S5.1 detects whether the target format document carries a manual confirmation mark and extracts the corresponding audit text to be processed and the problem description text to achieve automatic identification of confirmation cases and recovery of original text; step S5.2 assembles the audit text to be processed and the problem description text according to a preset case format to obtain new cases, achieving standardized encapsulation of case data; step S5.3 performs vectorization encoding on the new cases to obtain new case vectors, providing vector data for index updates; step S5.4 writes the new case vectors into the dynamic case library and performs incremental index updates to achieve real-time expansion of the case library and dynamic expansion of the retrieval scope.

[0199] Step S6: Count the number of new cases added to the dynamic case library within the preset time window. When the number of new cases exceeds the preset iteration threshold, perform a full reconstruction of the dynamic case library and use the fully reconstructed dynamic case library for the next round of matching retrieval.

[0200] Furthermore, step S6 specifically includes the following steps: Step S6.1: Count the number of new cases added to the dynamic case library within the preset time window to obtain the new case count.

[0201] Preferably, the core of this step is to complete the time window statistics of the number of new cases. By counting the number of new cases written into the dynamic case library within the preset time window, a quantitative basis is provided for subsequent iteration judgment.

[0202] Among them, the time window refers to the current moment. As the end point, with The time interval starting from ,in Add case count to the time window width. The calculation formula is: .

[0203] in, For the first Each case is added to the dynamic case library. timestamp, This is an indicator function that takes a value when a condition is true. Otherwise, take the value , This represents the total number of new cases added within the time window.

[0204] Specifically, for the case write timestamp record, the write timestamp is recorded when each case is written to the dynamic case library. The write time is stored in the write time field of the case metadata table. During statistics, the write time field of all cases in the metadata table is traversed to count.

[0205] Preferably, the time window width The day, that is, the most recent statistics The number of new cases written within a day is counted, and the statistical operation is automatically triggered after each execution of step S5.4 vector write.

[0206] For example, the current moment Time window Coverage of the latest Heaven, traversing the dynamic case library Metadata table, where timestamps are written Falling The cases within Item, new case count The corresponding record identifier is Count-New-YYYY-MM-DD-HH:MM:SS-001.

[0207] Step S6.2: Compare the number of new cases with the preset iteration threshold to determine whether the number of new cases exceeds the preset iteration threshold.

[0208] Preferably, the core of this step is to compare and determine the number of new cases with the iteration threshold. By judging whether the number of new cases exceeds the threshold, it is determined whether to trigger a full reconstruction, thus avoiding frequent reconstruction or delayed reconstruction.

[0209] Among these, the comparison and judgment will include an additional case count. With preset iteration threshold The comparison and judgment logic is as follows: .

[0210] in, To rebuild the trigger flag, This indicates that a full rebuild has been triggered. This indicates that it will not be triggered.

[0211] Among them, the threshold setting basis is the preset iteration threshold. Based on the initial case size of the dynamic case library The system is configured to trigger a full rebuild when the number of new cases reaches a preset proportion of the initial case size, in order to eliminate the index deviation accumulated during the incremental update process.

[0212] Preferably, a preset iteration threshold is used. That is, when New cases reached within a day A full rebuild is triggered when the condition is met; the time window counter is reset after the rebuild is triggered. .

[0213] For example, the number of new cases Preset iteration threshold Comparison results ,determination If a full reconstruction is not triggered, the corresponding record identifier is Compare-YYYY-MM-DD-HH:MM:SS-001.

[0214] Step S6.3: If the number of newly added cases exceeds the preset iteration threshold, then all cases in the dynamic case library are re-vectorized to obtain a set of case vectors after full reconstruction.

[0215] Preferably, the core of this step is to re-encode the entire dynamic case library. By re-executing vectorized encoding on all cases, the index bias accumulated from incremental updates is eliminated, resulting in a fully reconstructed set of case vectors.

[0216] Among them, for full re-encoding, when At that time, traverse the dynamic case library. All Each case, for each case text Repeat the same vectorization encoding process as step S5.3 to obtain the fully reconstructed case vectors. The set of case vectors after full reconstruction is as follows: .

[0217] in, The total number of cases during full reconstruction is [number], and the full recoding uses the same word vector model as the initial encoding. The mean pooling operation ensures the consistency of the vector space before and after reconstruction.

[0218] Regarding the difference between full reconstruction and incremental update, incremental update only adds new case vectors to the existing index without modifying existing case vectors; full reconstruction recodes all cases, which may cause changes in existing case vectors due to word vector model version updates or case text revisions, thereby eliminating the index bias accumulated by long-term incremental updates.

[0219] Preferably, the full reconstruction is performed in batch processing, with a batch size of [missing information]. That is, each processing For each case text, the dynamic case library retrieval service is suspended during the full reconstruction process and will resume after reconstruction is complete.

[0220] For example, A full rebuild is triggered at any time, and the current dynamic case library is currently... Case study, batch size ,point Batch processing, all cases are re-vectorized and encoded to obtain common cases vector, where (Compared to before reconstruction) There are minor differences (caused by floating-point errors in word vector model inference), and the corresponding record is identified as Rebuild-Vec-YYYY-MM-DD-HH:MM:SS-001.

[0221] Step S6.4: Replace the existing case vectors in the dynamic case library with the fully reconstructed case vector set for the next round of matching retrieval.

[0222] Preferably, the core of this step is to replace and deploy the case vectors after full reconstruction. By replacing the existing vectors with the reconstructed case vector set and rebuilding the retrieval index, the retrieval accuracy is periodically restored and improved.

[0223] For vector replacement, a dynamic case library will be used. The existing set of case vectors Replace with the set of case vectors after full reconstruction The replacement operation includes: clearing all vectors in the existing Faiss index, and... middle Each vector is written in batches to a newly created Faiss index, while the index type remains unchanged as IndexFlatIP.

[0224] Specifically, for the status update after replacement, the index version number of the dynamic case library is updated after the replacement is completed. Record reconstruction timestamp Reset the time window counter The reconstructed dynamic case library immediately resumes retrieval services and is used for matching retrieval in the next round of step S2.

[0225] Preferably, the index version number The initial value is Each full reconstruction increments Reconstruct timestamp The precision is at the millisecond level, and the replacement operation is an atomic operation, ensuring that there is no index gap during the replacement process.

[0226] For example, the set of case vectors after full reconstruction. common Replace the existing Faiss index with a vector. The old vector, the index version number after replacement (No. (Full reconstruction), reconstruction timestamp Counter reset The dynamic case library resumes retrieval services, and the next round of matching retrieval uses the rebuilt index, with the corresponding record identifier being Replace-Index-YYYY-MM-DD-HH:MM:SS-001.

[0227] Beneficial effects of steps S6.1 to S6.4: This series of steps addresses the issue of decreased retrieval accuracy caused by the cumulative deviation of the case vector index during incremental updates. By statistically analyzing the number of new cases added within a preset time window and performing a full reconstruction of the dynamic case library when the number exceeds an iteration threshold, the case vector index can be periodically optimized and retrieval accuracy can be continuously maintained.

[0228] Specifically, step S6.1 involves counting the number of new cases added to the dynamic case library within a preset time window to obtain a count of new cases, providing a quantitative basis for iteration determination; step S6.2 involves comparing the count of new cases with a preset iteration threshold to determine whether a full reconstruction is triggered, avoiding frequent or delayed reconstruction; step S6.3 involves re-encoding all cases in the dynamic case library using vectorization to obtain a set of case vectors after full reconstruction, eliminating the index bias accumulated from incremental updates; and step S6.4 involves replacing the existing case vectors with the set of case vectors after full reconstruction for use in the next round of matching retrieval, achieving periodic recovery and improvement of retrieval accuracy.

[0229] In summary, the overall beneficial effects of steps S1 to S6 in this embodiment are as follows: This method addresses issues such as reliance on manual operation in audit text processing, insufficient accuracy in historical case retrieval, and lack of closed-loop iteration in knowledge base updates. It constructs text feature vectors and issue feature vectors by automatically segmenting, tagging, and vectorizing the audit text and issue description text, eliminating subjective judgment differences among different auditors. It combines full extraction of static knowledge documents with semantic retrieval from a dynamic case library, using cosine similarity calculation and a re-ranking model to accurately recall historical cases. The method inputs structured input sequences into a privately deployed language model to obtain structured text sequences, which are then converted into target format documents for direct use. Simultaneously, manually confirmed new cases are vectorized and written into the dynamic case library to trigger incremental updates. When the accumulated new cases exceed an iteration threshold, a full reconstruction of the dynamic case library is performed, forming a closed-loop iteration mechanism that continuously improves case retrieval accuracy.

[0230] Specifically, step S1 extracts noun and verb words through word segmentation and part-of-speech tagging to construct feature vectors, thereby achieving automated and structured representation of audit text; step S2 combines full extraction of static knowledge with dynamic case library matching and retrieval, balancing the integrity of the evaluation framework with accurate reuse of historical experience; step S3 achieves multi-source knowledge fusion processing and structured output through preset template splicing and private language model reasoning; step S4 achieves automatic formatting of the output document through the extraction and conversion of title level tags and table structure tags; step S5 achieves dynamic expansion of the case library by triggering vectorized writing and incremental index updates for newly added cases through manual confirmation tags; and step S6 achieves periodic optimization of the case vector index and continuous improvement of retrieval accuracy through statistical analysis of the number of newly added cases and full reconstruction.

[0231] like Figure 2 As shown, this embodiment provides a multi-dimensional intelligent processing system for audit text. In this embodiment, the multi-dimensional intelligent processing system is applied to the multi-dimensional intelligent processing method as described in the above embodiment.

[0232] Specifically, the multi-dimensional intelligent processing system includes an audit text vector construction module 1, which is electrically connected and then communicatively connected in sequence; a retrieval case text sequence acquisition module 2; a structured text sequence acquisition module 3; a target format document conversion module 4; a dynamic case library incremental update module 5; and a dynamic case library full reconstruction module 6.

[0233] The system comprises several modules: Module 1 (Audit Text Vector Construction Module 1) acquires the audit text to be processed and the problem description text, performs word segmentation and part-of-speech tagging on the audit text, extracts noun and verb words to construct text feature vectors, and performs the same operation on the problem description text to construct problem feature vectors; Module 2 (Sequence Retrieval Case Text Sequence Acquisition Module 2) performs full text extraction on a pre-configured static knowledge document set to obtain a static knowledge text sequence, and matches the problem feature vectors with a pre-configured dynamic case library to obtain a retrieval case text sequence; Module 3 (Structured Text Sequence Acquisition Module 3) concatenates the text feature vectors, problem feature vectors, static knowledge text sequences, and retrieval case text sequences according to a preset template into a structured input sequence, inputs the structured input sequence into a privately deployed language model, and acquires the privately deployed language model. The output is a structured text sequence; the target format document conversion module 4 is used to extract the heading level markers and table structure markers of the structured text sequence, and convert the structured text sequence into a target format document according to the heading level markers and table structure markers; the dynamic case library incremental update module 5 is used to respond to the manual confirmation markers of the target format document, and take the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation marker as a new case, vectorize the new case and write it into the dynamic case library, triggering the incremental update of the dynamic case library; the dynamic case library full reconstruction module 6 is used to count the number of new cases in the dynamic case library within a preset time window. When the number of new cases exceeds the preset iteration threshold, the dynamic case library is fully reconstructed and the fully reconstructed dynamic case library is used for the next round of matching retrieval.

[0234] like Figure 3 As shown, the electronic device 7 includes a processor 71 and a memory 72 coupled to the processor 71.

[0235] The memory 72 stores program instructions for implementing the multi-dimensional intelligent processing method for audit text in any of the above embodiments.

[0236] The processor 71 is used to execute program instructions stored in the memory 72 to perform multi-dimensional intelligent processing of audit text.

[0237] The processor 71 can also be referred to as a CPU (Central Processing Unit). The processor 71 may be an integrated circuit chip with signal processing capabilities. The processor 71 can also be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor can be a microprocessor or any conventional processor.

[0238] Furthermore, Figure 4 This is a schematic diagram of the structure of a storage medium according to an embodiment of this application. See also: Figure 4 The storage medium 8 in this embodiment stores program instructions 81 capable of implementing all the above methods. These program instructions 81 can be stored in the storage medium as a software product, including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods in each embodiment of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, or terminal devices such as computers, servers, mobile phones, and tablets.

[0239] In the several embodiments provided in this application, it should be understood that the disclosed systems, methods, and approaches can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For instance, the multi-dimensional intelligent processing of audit text for a unit is only one logical function of multi-dimensional intelligent processing of audit text. In actual implementation, there may be other multi-dimensional intelligent processing methods for audit text. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, signal, or other forms.

[0240] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-dimensional intelligent processing method for audit text, characterized in that, The multi-dimensional intelligent processing method includes: Step S1: Obtain the audit text to be processed and the problem description text; perform word segmentation and part-of-speech tagging on the audit text to be processed; extract noun words and verb words to construct a text feature vector; and perform the same operation on the problem description text to construct a problem feature vector. Step S2: Perform full text extraction on the pre-configured static knowledge document set to obtain a static knowledge text sequence, and match the problem feature vector with the pre-configured dynamic case library to obtain a search case text sequence; Step S3: The text feature vector, the question feature vector, the static knowledge text sequence, and the retrieval case text sequence are concatenated into a structured input sequence according to a preset template. The structured input sequence is then input into the privately deployed language model, and the structured text sequence output by the privately deployed language model is obtained. Step S4: Extract the heading level tags and table structure tags of the structured text sequence, and convert the structured text sequence into a target format document according to the heading level tags and table structure tags; Step S5: In response to the manual confirmation mark of the target format document, the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark are taken as new cases, the new cases are vectorized and encoded and written into the dynamic case library, triggering the incremental update of the dynamic case library; Step S6: Count the number of newly added cases in the dynamic case library within the preset time window. When the number of newly added cases exceeds the preset iteration threshold, perform a full reconstruction of the dynamic case library and use the fully reconstructed dynamic case library for the next round of matching retrieval.

2. The multi-dimensional intelligent processing method according to claim 1, characterized in that, Step S1: Obtain the audit text to be processed and the problem description text. Perform word segmentation and part-of-speech tagging on the audit text to be processed, extract noun words and verb words to construct a text feature vector, and perform the same operation on the problem description text to construct a problem feature vector, including: Step S1.1: Obtain the audit text to be processed and the problem description text, and perform word segmentation on the audit text to be processed to obtain the word segmentation result; Step S1.2: Perform part-of-speech tagging on the word segmentation results, filter noun words and verb words to obtain a candidate word set; Step S1.3: Vectorize each word in the candidate word set to obtain a word vector sequence; Step S1.4: Perform mean pooling on the word vector sequence to obtain the text feature vector of the audit text to be processed; Step S1.5: Perform the same word segmentation, part-of-speech tagging, vectorization encoding, and mean pooling operations on the problem description text as on the audit text to be processed, to obtain the problem feature vector.

3. The multi-dimensional intelligent processing method according to claim 1, characterized in that, Step S2 involves performing full text extraction on the pre-configured static knowledge document set to obtain a static knowledge text sequence, and then matching the problem feature vector with a pre-configured dynamic case library to obtain a retrieval case text sequence, including: Step S2.1: Traverse each document in the pre-configured static knowledge document set, perform text extraction operation on each document, and obtain the plain text content corresponding to each document; Step S2.2: Concatenate the plain text content corresponding to each document in document order to obtain a static knowledge text sequence; Step S2.3: Vectorize and encode each case text in the pre-configured dynamic case library to obtain a case vector set; Step S2.4: Calculate the cosine similarity between the problem feature vector and each case vector in the case vector set, sort them in descending order of similarity, and select the top K cases as the candidate case set; Step S2.5: Calculate the relevance score of each candidate case using the re-ranking model for the candidate case set, and filter out target cases whose relevance scores exceed a preset threshold. Step S2.6: Concatenate the case texts of the target case in descending order of relevance score to obtain the search case text sequence.

4. The multi-dimensional intelligent processing method according to claim 1, characterized in that, Step S3: Concatenate the text feature vector, the question feature vector, the static knowledge text sequence, and the retrieval case text sequence according to a preset template into a structured input sequence. Input the structured input sequence into the privately deployed language model and obtain the structured text sequence output by the privately deployed language model, including: Step S3.1: Concatenate the text feature vector and the question feature vector to form a question representation vector; Step S3.2: Add paragraph start markers to the static knowledge text sequence and the retrieval case text sequence respectively to obtain the marked knowledge text and the marked case text; Step S3.3: Concatenate the problem representation vector, the labeled knowledge text, and the labeled case text into a structured input sequence according to a preset template; Step S3.4: Input the structured input sequence into the privately deployed language model to obtain the structured text sequence generated by the privately deployed language model performing autoregressive decoding on the structured input sequence.

5. The multi-dimensional intelligent processing method according to claim 1, characterized in that, Step S4: Extract the heading level tags and table structure tags from the structured text sequence, and convert the structured text sequence into a target format document according to the heading level tags and table structure tags, including: Step S4.1: Perform a marker scan on the structured text sequence to identify title markers and table markers; Step S4.2: Determine the document heading level corresponding to each heading based on the hierarchical nesting relationship of the heading markers; Step S4.3: Parse the number of rows, columns, and cell contents of the table based on the table identifiers to construct the table structure data; Step S4.4: Write the structured text sequence into the target format template according to the document title level and the table structure data to generate the target format document.

6. The multi-dimensional intelligent processing method according to claim 1, characterized in that, Step S5, in response to the manual confirmation mark on the target format document, the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark are added as new cases. The new cases are vectorized and encoded and written into the dynamic case library, triggering an incremental update of the dynamic case library, including: Step S5.1: Detect whether the target format document carries a manual confirmation mark. If it does, extract the audit text to be processed and the problem description text corresponding to the target format document. Step S5.2: Assemble the audit text to be processed and the problem description text according to a preset case format to obtain a new case; Step S5.3: Perform vectorization encoding on the newly added case to obtain the newly added case vector; Step S5.4: Write the newly added case vector into the dynamic case library, and perform an incremental index update on the dynamic case library.

7. The multi-dimensional intelligent processing method according to claim 1, characterized in that, Step S6: Count the number of newly added cases in the dynamic case library within a preset time window. When the number of newly added cases exceeds a preset iteration threshold, perform a full reconstruction of the dynamic case library. Use the fully reconstructed dynamic case library for the next round of matching retrieval, including: Step S6.1: Count the number of new cases added to the dynamic case library within the preset time window to obtain the new case count; Step S6.2: Compare the new case count with a preset iteration threshold to determine whether the new case count exceeds the preset iteration threshold; Step S6.3: If the number of newly added cases exceeds the preset iteration threshold, then all cases in the dynamic case library are re-vectorized to obtain a set of case vectors after full reconstruction. Step S6.4: Replace the existing case vectors in the dynamic case library with the fully reconstructed case vector set for the next round of matching retrieval.

8. A multi-dimensional intelligent processing system for audit text, wherein the multi-dimensional intelligent processing system is applied to the multi-dimensional intelligent processing method as described in any one of claims 1 to 7, characterized in that, The multi-dimensional intelligent processing system includes: The audit text vector construction module is used to obtain the audit text to be processed and the problem description text, perform word segmentation and part-of-speech tagging on the audit text to be processed, extract noun words and verb words to construct text feature vectors, and perform the same operation on the problem description text to construct problem feature vectors; The retrieval case text sequence acquisition module is used to perform full text extraction on a pre-configured static knowledge document set to obtain a static knowledge text sequence, and to match and retrieve the problem feature vector with a pre-configured dynamic case library to obtain a retrieval case text sequence; The structured text sequence acquisition module is used to concatenate the text feature vector, the question feature vector, the static knowledge text sequence, and the retrieval case text sequence into a structured input sequence according to a preset template, input the structured input sequence into a privately deployed language model, and obtain the structured text sequence output by the privately deployed language model. The target format document conversion module is used to extract the heading level marks and table structure marks of the structured text sequence, and convert the structured text sequence into a target format document according to the heading level marks and table structure marks; The dynamic case library incremental update module is used to respond to the manual confirmation mark of the target format document, take the audit text to be processed and the problem description text corresponding to the target format document with the manual confirmation mark as a new case, perform vectorization encoding on the new case and write it into the dynamic case library, and trigger the incremental update of the dynamic case library. The dynamic case library full reconstruction module is used to count the number of new cases in the dynamic case library within a preset time window. When the number of new cases exceeds a preset iteration threshold, the dynamic case library is fully reconstructed and the reconstructed dynamic case library is used for the next round of matching retrieval.

9. An electronic device, characterized in that, The method includes a processor and a memory coupled to the processor, the memory storing program instructions executable by the processor; when the processor executes the program instructions stored in the memory, it implements the multi-dimensional intelligent processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions that, when executed by a processor, enable the implementation of the multi-dimensional intelligent processing method as described in any one of claims 1 to 7.

Images