Text recommendation system and method for interpreting audit reports based on large models

Abstract

This invention discloses a text recommendation system and method for interpreting audit reports based on a large-scale model. The invention utilizes a specially trained large-scale model to deeply interpret audit reports and construct an audit knowledge graph. Based on this knowledge graph, key descriptive information from the reports is extracted. During the text recommendation process, corresponding recommendation suggestions are generated for each user's input query. The large-scale model then responds to these suggestions by providing several audit reports with the highest matching degree as references. This improves the accuracy and efficiency of text recommendation, achieving automated and intelligent interpretation of audit reports and enhancing the overall efficiency and quality of audit work.

Description

Technical Field

[0001] The present invention relates to the field of digital technology, and in particular to a text recommendation system and method for interpreting audit reports based on a large model. Background Art

[0002] The text of an audit report is standardized, logical, objective, and detailed; currently, the interpretation and use of audit reports is a complex and time-consuming process. Relying on manual reading and analysis of audit reports is inefficient and prone to missing key information. When using traditional methods such as keyword matching to recommend audit reports, the following problems and shortcomings may be encountered: the frequent repetition of professional terms and industry-specific vocabulary in audit reports may result in low accuracy of keyword matching, inability to accurately capture the core content of the report, and inability to effectively identify the contextual relationships and logical structures in the report, resulting in low relevance of recommendations.

[0003] Although some studies have begun to use natural language processing technology based on knowledge graphs to assist in text analysis, these technologies are often unable to deeply understand the professional terms and complex structures in audit reports, resulting in limitations in entity and relationship extraction and knowledge graph construction for large-scale texts in the audit field, and their accuracy and efficiency are low. Summary of the invention

[0004] In view of the above defects, the technical problem solved by the present invention is: based on a specially trained large model, the audit report is deeply interpreted, key entities and relationships are extracted, and an audit knowledge graph is constructed; based on the audit knowledge graph, the key description information of the report is extracted from each audit report template that can be recommended; in the text recommendation process, the user inputs a query request, and the entities and relationships that match the above knowledge graph are extracted from the query request, and the corresponding recommendation prompt words are generated using the prompt word engineering; the large model responds to the recommendation prompt word, evaluates the degree of match between the key description information of each audit report and the entities and relationships of the query request, and provides several audit reports with the highest degree of match as reference, thereby improving the accuracy and efficiency of text recommendation. Through this technical solution, it is possible to achieve automated and intelligent interpretation of audit reports, provide users with high-quality text recommendations, and improve the overall efficiency and quality of audit work.

[0005] The first aspect of the present invention provides a text recommendation system for interpreting audit reports based on a large model, comprising:

[0006] The audit knowledge graph construction large model unit is used to perform semantic analysis on the text content of the audit report, identify and extract the key entities and relationships therein, and construct a structured audit knowledge graph based on the key entities and their relationships; wherein the key entities identified and extracted by the audit knowledge graph construction large model unit include: audit object, audit objective, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; the relationships describe the mutual relationships between the above key entities;

[0007] The report key description information extraction unit is used to refine the information of each audit report template that can be recommended, identify and extract the report key description information with a correlation value higher than the threshold with the key entity in the audit knowledge graph;

[0008] The audit text recommendation big model unit is used to receive the audit description content input by the user, combine it with the audit knowledge graph, identify the key entities and their relationships in the audit description content, automatically generate recommended prompt words through prompt word engineering, and guide the audit text recommendation big model to respond to the recommended prompt words, score the applicability of the key description information of the report, and recommend the audit report template with the highest score.

[0009] Among them, the audit knowledge graph constructs a large model in the following manner, and the audit knowledge graph constructs a large model for targeted training and fine-tuning, including: preparing a data set, collecting and organizing a data set containing diverse and representative audit reports; selecting a pre-trained model, and selecting a pre-trained large model based on the Transformer architecture as a basis; inputting the audit report text into the Transformer pre-trained large model, extracting the feature representation of the audit report text, and adding a new hierarchy on the basis of the Transformer pre-trained large model, which is specifically used to identify and classify key entities and their relationships in the audit report; in the fine-tuning process, for entity and relationship recognition tasks, the cross entropy loss function is used to calculate the difference between the predicted label and the true label of the entity type and entity relationship, and the gradient descent algorithm is used to update the model weights; the fine-tuned pre-trained large model is evaluated.

[0010] The report key description information extraction unit takes each sentence in the audit report template as a unit, and calculates the correlation value between the unit and the entity in the audit knowledge graph as follows:

[0011]

[0012] in Represents a sentence unit in an audit report template and audit entities in the knowledge graph The correlation value of and are the vector representations of the sentence unit and entity in the word space, and The modulus lengths of the vectors representing sentence units and entities respectively; for sentence units in the audit report template whose correlation values ​​with key entities in the audit knowledge graph are higher than a preset threshold, they are extracted as key description information of the report.

[0013] The audit text recommendation model unit pre-processes the audit description content input by the user; identifies the key entities in the audit description content through named entity recognition technology, and extracts the relationship between the key entities based on the audit knowledge graph; the entities identified in the audit description content are represented as entity sets ,in Indicates entity and The relationship between ; Based on the extracted key entities and their relationships, the recommended prompt words containing the recommended task description and the recommended boundary restrictions are automatically written to provide clear guidance information for the audit report template recommendation; Among them, for the recommended task description, the identified key entities and relationships are mapped to the template of the audit task description. This mapping process uses the following formula to represent the generation of the task description: ,in Represents the recommendation task description, function It means that a suitable description template is selected according to the entity set and the relationship set, and the entity name is filled in to form a recommended task description; for the recommended boundary restrictions, the recommended boundary restrictions are generated according to the context of the entity and the relationship, combined with the constraint condition set containing specific rules of the audit field; the key description information in each audit report template is compared with the recommended prompt words. The audit text recommendation model compares the similarity between the recommended prompt words and the key description information, and outputs the applicability score; the calculation method of the applicability score is:

[0014]

[0015] in represents the suitability score, is the number of entities and relations in the recommended prompt words, is the weight value of the i-th entity, Represents the i-th entity evaluated by the large model With key description information The similarity.

[0016] A second aspect of the present invention provides a text recommendation method for interpreting an audit report based on a large model, comprising:

[0017] Audit knowledge graph construction steps: construct a large model through the audit knowledge graph, perform semantic analysis on the text content of the audit report, identify and extract the key entities and relationships therein, and construct a structured audit knowledge graph based on the key entities and their relationships; wherein the key entities identified and extracted by the audit knowledge graph construction large model include: audit object, audit objective, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; the relationships describe the mutual relationships between the above key entities;

[0018] Extraction step of key descriptive information of the report: for each audit report template that can be recommended, perform refined information extraction, identify and extract key descriptive information of the report whose correlation value with key entities in the audit knowledge graph is higher than the threshold;

[0019] Audit text recommendation step: receive the audit description content input by the user, combine it with the audit knowledge graph, identify the key entities and their relationships in the audit description content, automatically generate recommended prompt words through prompt word engineering, and guide the audit text recommendation model to respond to the recommended prompt words, score the applicability of the key description information of the report, and recommend the audit report template with the highest score.

[0020] Among them, the audit knowledge graph construction step also specifically includes: targeted training and fine-tuning of the audit knowledge graph to construct a large model, including: data set preparation, collecting and organizing a data set containing diverse and representative audit reports; pre-training model selection, selecting a pre-training large model based on the Transformer architecture as a basis; inputting the audit report text into the Transformer pre-training large model, extracting the feature representation of the audit report text, and adding a new hierarchy on the basis of the Transformer pre-training large model, which is specifically used to identify and classify key entities and their relationships in the audit report; in the fine-tuning process, for entity and relationship recognition tasks, the cross entropy loss function is used to calculate the difference between the predicted label and the true label of the entity type and entity relationship, and the gradient descent algorithm is used to update the model weight; the fine-tuned pre-training large model is evaluated.

[0021] Among them, the audit knowledge graph construction steps also specifically include: performing entity normalization, normalizing the extracted entities; performing relationship modeling, establishing a relationship model based on the relationship between entities analyzed by the large model; performing graph construction, using a graph database or a graph computing framework to integrate entities and their relationships into the knowledge graph, with each entity as a node in the graph and each relationship as an edge connecting the nodes.

[0022] The step of extracting key description information of the report specifically includes: taking each sentence in the audit report template as a unit, calculating the correlation value between the unit and the entity in the audit knowledge graph as follows:

[0023]

[0024] in Represents a sentence unit in an audit report template and audit entities in the knowledge graph The correlation value of and are the vector representations of the sentence unit and entity in the word space, and The modulus lengths of the vectors representing sentence units and entities respectively; for sentence units in the audit report template whose correlation values ​​with key entities in the audit knowledge graph are higher than a preset threshold, they are extracted as key description information of the report.

[0025] The audit text recommendation step specifically includes: preprocessing the audit description content input by the user; identifying the key entities in the audit description content through named entity recognition technology, and extracting the relationship between the key entities based on the audit knowledge graph; for the entities identified in the audit description content, representing them as entity sets ,in Indicates entity and The relationship between ; Based on the extracted key entities and their relationships, the recommended prompt words containing the recommended task description and the recommended boundary restrictions are automatically written to provide clear guidance information for the audit report template recommendation; Among them, for the recommended task description, the identified key entities and relationships are mapped to the template of the audit task description. This mapping process uses the following formula to represent the generation of the task description: ,in Represents the recommendation task description, function It means selecting a suitable description template based on the entity set and relationship set, and filling in the entity name to form a recommended task description; for the recommended boundary restrictions, the recommended boundary restrictions will be generated based on the context of the entities and relationships, combined with a set of constraints that contain specific rules in the audit field.

[0026] The audit text recommendation step specifically includes: comparing the key description information in each audit report template with the recommended prompt words, and the audit text recommendation model compares the similarity between the recommended prompt words and the key description information, thereby outputting a suitability score; the calculation method of the suitability score is:

[0027]

[0028] in represents the suitability score, is the number of entities and relations in the recommended prompt words, is the weight value of the i-th entity, Represents the i-th entity evaluated by the large model With key description information The similarity.

[0029] Therefore, the present invention is based on a specially trained big model and utilizes the semantic understanding and generation capabilities of the big model to realize the process of in-depth interpretation of audit reports, construction of audit knowledge graphs, extraction of key descriptive information in reports, and audit text matching and recommendation, thereby realizing automated and intelligent interpretation of audit reports, providing users with high-quality text recommendations, and improving the overall efficiency and quality of audit work. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0031] Figure 1 A framework diagram of a text recommendation system for interpreting audit reports based on a large model disclosed in an embodiment of the present invention;

[0032] Figure 2 This is a flow chart of a text recommendation method for interpreting audit reports based on a large model disclosed in an embodiment of the present invention.

[0033] The above drawings show clear embodiments of the present disclosure, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present disclosure in any way, but to illustrate the concepts of the present disclosure to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0034] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0035] The text recommendation system based on large model interpretation of audit reports disclosed in the present invention is as follows: Figure 1As shown, it includes: an audit knowledge graph construction large model unit, a report key description information extraction unit, and an audit text recommendation large model unit.

[0036] The audit knowledge graph constructs a large model unit, which is used to perform semantic analysis on the text content of the audit report, identify and extract the key entities and relationships therein, and construct a structured audit knowledge graph based on the key entities and their relationships. Among them, the key entities identified and extracted by the audit knowledge graph construction large model unit include: audit objects, audit objectives, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; the relationships describe the mutual relationships between the above key entities.

[0037] Among them, in order to ensure that the audit knowledge graph construction big model can accurately understand the specific domain knowledge in the audit report, so as to effectively and accurately extract the key entities and their relationships, the big model must be targeted for training and fine-tuning. The following are the detailed training and fine-tuning steps. First, prepare the data set to collect and organize a data set containing diverse and representative audit reports; the data set should cover various audit objects, audit objectives, audit-related facts, and audit results, as well as various types of audit standards, various audit issues and related audit risks under audit rules; then, select a pre-trained model based on the Transformer architecture as the basis, such as a pre-trained big model based on the BERT or GPT architecture; then, use the following steps to fine-tune the pre-trained big model based on the Transformer architecture: Input the audit report text into the Transformer In the pre-trained large model, extract the feature representation of the audit report text, add a new hierarchy based on the Transformer pre-trained large model, which is specifically used to identify and classify the key entities and their relationships in the audit report; in the fine-tuning process, for the entity and relationship recognition tasks, use the cross entropy loss function to calculate the difference between the predicted label and the true label of the entity type and entity relationship, and use the gradient descent algorithm, such as the Adam optimizer, to update the weight of the model; perform model evaluation on the fine-tuned pre-trained large model, and evaluate the model performance on the validation set to ensure that the model can not only perform well on the training set, but also generalize to unseen data. In the complete fine-tuning process, we will train the model for multiple rounds, and evaluate it on the validation set after each round of training; if the performance of the model does not improve, we will adjust the learning rate, so as to finally form the audit knowledge graph construction large model.

[0038] The audit knowledge graph construction big model adopts the big model under the deep learning framework to perform semantic analysis on the text content of the audit report. In this process, the audit report is first preprocessed, including removing irrelevant information, word segmentation, part-of-speech tagging, etc., so that the model can better understand the text content; then, the big model captures the key information in the text through its internal attention mechanism. The big model outputs the type probability distribution of each word, identifies the words belonging to the key entities such as the audit object, audit target, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; by setting the threshold of the probability distribution and using post-processing technology, the identified words are combined into complete entities; further, the big model will also analyze the relationship between these key entities to lay the foundation for the subsequent knowledge graph construction. For example, if the audit report mentions that "the company's financial statements are false", the big model will identify "company" as the audit object, "financial statements" as the audit target, and "false reporting" as the audit-related fact, and analyze the logical relationship between them.

[0039] The audit knowledge graph construction model processes the extracted key entities and relationships into a structured data representation to form the audit knowledge graph. Specifically, firstly, entity normalization is performed to normalize the extracted entities to ensure the consistency and accuracy of the entity names, for example, all the "financial statements" mentioned are unified into a standard entity name; then, relationship modeling is performed to establish a relationship model based on the relationship between entities analyzed by the large model. The relationships covered by the relationship model may include causal relationships, attribution relationships, influence relationships, comparative relationships, etc.; graph construction is performed to integrate entities and their relationships into the knowledge graph using a graph database or graph computing framework, such as Neo4j or its equivalent tools, with each entity as a node in the graph and each relationship as an edge connecting the nodes. For example, if the audit report points out that "the false reporting of financial statements has led to the false increase of company profits", then "financial statements" and "company profits" are two entity nodes, and the causal relationship between them is added to the audit knowledge graph as an edge. In this way, the audit knowledge graph not only provides a structured representation for the audit report, but also provides rich semantic information for subsequent query and text recommendation tasks.

[0040] The report key description information extraction unit is used to perform refined information extraction for each audit report template that can be recommended, and to identify and extract report key description information whose correlation value with key entities in the audit knowledge graph is higher than a threshold.

[0041] The report key description information extraction unit is associated with an audit report template library, which stores desensitized and declassified audit report texts as recommended audit report templates. The report key description information extraction unit performs association analysis on any audit report template in the library and key entities in the constructed audit knowledge graph to determine text in the audit report template that is highly correlated with entities in the audit knowledge graph.

[0042] Specifically, the report key description information extraction unit takes each sentence in the audit report template as a unit and calculates the correlation value between the unit and the entity in the audit knowledge graph as follows:

[0043]

[0044] in Represents a sentence unit in an audit report template and audit entities in the knowledge graph The correlation value of and are the vector representations of the sentence unit and entity in the word space, and The modulus lengths of the vectors representing sentence units and entities respectively. For sentence units in the audit report template whose correlation values ​​with key entities in the audit knowledge graph are higher than a preset threshold, they are extracted as key description information of the report.

[0045] The audit text recommendation big model unit is used to receive the audit description content input by the user, combine it with the audit knowledge graph, identify the key entities and their relationships in the audit description content, automatically generate recommended prompt words through prompt word engineering, and guide the audit text recommendation big model to respond to the recommended prompt words, score the applicability of the key description information of the report, and recommend the audit report template with the highest score.

[0046] Specifically, the audit text recommendation model unit identifies and extracts key entities and their relationships in the audit description content based on the audit description content input by the user and the audit knowledge graph. First, the audit description content input by the user is preprocessed, including the removal of stop words, word segmentation, and part-of-speech tagging. Subsequently, the named entity recognition (NER) technology is used to identify the key entities in the audit description content, and the relationships between the key entities are extracted based on the audit knowledge graph. For the entities identified in the audit description content, they are represented as entity sets ,in Indicates entity and The relationship between For example, if the audit description entered by the user is "annual audit of the company's financial statements, focusing on cash flow and inventory management", after preprocessing this description, key entities such as "financial statements", "annual audit", "cash flow" and "inventory management" are identified; then, the algorithm further analyzes the relationship between these entities, for example, there is an "audit object" relationship between "annual audit" and "financial statements".

[0047] The audit text recommendation model unit automatically writes the recommendation prompt words including the recommended task description and the recommended boundary restrictions based on the extracted key entities and their relationships, providing clear guidance information for the audit report template recommendation. For the recommended task description, the identified key entities and relationships are mapped to the template of the audit task description. This mapping process uses the following formula to represent the generation of the task description: ,in Represents the recommendation task description, function It means selecting a suitable description template based on the entity set and relationship set, and filling in the entity name to form a recommended task description. In order to ensure the accuracy of the recommendation results, boundary restrictions need to be defined; for the recommended boundary restrictions, the recommended boundary restrictions will be generated based on the context of the entities and relationships, combined with the constraint condition set containing specific rules of the audit field.

[0048] The audit text recommendation model of this unit responds to the recommendation prompts, scores the applicability of the key descriptive information of the report, and recommends the audit report template with the highest score. The goal of this step is to evaluate the degree of match between each audit report template and the recommended prompts, so as to determine which template is most suitable for the current audit needs. The specific process of implementing this scoring mechanism is as follows: First, the key descriptive information in each audit report template is compared with the recommended prompts. The audit text recommendation model compares the similarity between the recommended prompts and the key descriptive information, and outputs the applicability score. The calculation method of the applicability score is:

[0049]

[0050] in represents the suitability score, is the number of entities and relations in the recommended prompt words, is the weight value of the i-th entity, Represents the i-th entity evaluated by the large model With key description information The similarity.

[0051] After completing the applicability scoring of all audit report templates, the template with the highest score is selected to be recommended to the user. Specifically, all audit report templates are sorted according to the applicability score calculated in the previous step. The higher the score, the higher the position of the template in the list. The audit report template with the highest score is selected as the main recommendation.

[0052] See also Figure 2 The present invention also provides a text recommendation method for interpreting audit reports based on a large model, including:

[0053] Audit knowledge graph construction step S1: Build a big model through the audit knowledge graph, perform semantic analysis on the text content of the audit report, identify and extract the key entities and relationships therein, and build a structured audit knowledge graph based on the key entities and their relationships. Among them, the key entities identified and extracted by the audit knowledge graph construction big model include: audit objects, audit objectives, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; the relationships describe the mutual relationships between the above key entities.

[0054] Step S2 of extracting key descriptive information of the report: For each audit report template available for recommendation, perform refined information extraction, identify and extract key descriptive information of the report whose correlation value with key entities in the audit knowledge graph is higher than the threshold.

[0055] Audit text recommendation step S3: Receive the audit description content input by the user, combine it with the audit knowledge graph, identify the key entities and their relationships in the audit description content, automatically generate recommended prompt words through prompt word engineering, and guide the audit text recommendation model to respond to the recommended prompt words, score the applicability of the key description information of the report, and recommend the audit report template with the highest score.

[0056] Among them, the audit knowledge graph construction step S1 also specifically includes: targeted training and fine-tuning of the audit knowledge graph to construct a large model, including: data set preparation, collecting and organizing a data set containing diverse and representative audit reports; pre-training model selection, selecting a pre-training large model based on the Transformer architecture as a basis; inputting the audit report text into the Transformer pre-training large model, extracting the feature representation of the audit report text, and adding a new hierarchy on the basis of the Transformer pre-training large model, which is specifically used to identify and classify key entities and their relationships in the audit report; in the fine-tuning process, for entity and relationship recognition tasks, the cross entropy loss function is used to calculate the difference between the predicted label and the true label of the entity type and entity relationship, and the gradient descent algorithm is used to update the model weight; the fine-tuned pre-training large model is evaluated.

[0057] Among them, the audit knowledge graph construction step S1 also specifically includes: performing entity normalization, normalizing the extracted entities; performing relationship modeling, establishing a relationship model based on the relationship between entities analyzed by the large model; performing graph construction, using a graph database or a graph computing framework, integrating entities and their relationships into the knowledge graph, with each entity as a node in the graph and each relationship as an edge connecting the nodes.

[0058] The step S2 of extracting key description information of the report specifically includes: taking each sentence in the audit report template as a unit, calculating the correlation value between the unit and the entity in the audit knowledge graph as follows:

[0059]

[0060] in Represents a sentence unit in an audit report template and audit entities in the knowledge graph The correlation value of and are the vector representations of the sentence unit and entity in the word space, and The modulus lengths of the vectors representing sentence units and entities respectively; for sentence units in the audit report template whose correlation values ​​with key entities in the audit knowledge graph are higher than a preset threshold, they are extracted as key description information of the report.

[0061] The audit text recommendation step S3 specifically includes: preprocessing the audit description content input by the user; identifying the key entities in the audit description content through named entity recognition technology, and extracting the relationship between the key entities based on the audit knowledge graph; for the entities identified in the audit description content, representing them as entity sets ,in Indicates entity and The relationship between ; Based on the extracted key entities and their relationships, the recommended prompt words containing the recommended task description and the recommended boundary restrictions are automatically written to provide clear guidance information for the audit report template recommendation; Among them, for the recommended task description, the identified key entities and relationships are mapped to the template of the audit task description. This mapping process uses the following formula to represent the generation of the task description: ,in Represents the recommendation task description, function It means selecting a suitable description template based on the entity set and relationship set, and filling in the entity name to form a recommended task description; for the recommended boundary restrictions, the recommended boundary restrictions will be generated based on the context of the entities and relationships, combined with a set of constraints that contain specific rules in the audit field.

[0062] The audit text recommendation step S3 specifically includes: comparing the key description information in each audit report template with the recommended prompt words, and the audit text recommendation model compares the similarity between the recommended prompt words and the key description information, thereby outputting a suitability score; the calculation method of the suitability score is:

[0063]

[0064] in represents the suitability score, is the number of entities and relations in the recommended prompt words, is the weight value of the i-th entity, Represents the i-th entity evaluated by the large model With key description information The similarity.

[0065] Therefore, the present invention is based on a specially trained big model and utilizes the semantic understanding and generation capabilities of the big model to realize the process of in-depth interpretation of audit reports, construction of audit knowledge graphs, extraction of key descriptive information in reports, and audit text matching and recommendation, thereby realizing automated and intelligent interpretation of audit reports, providing users with high-quality text recommendations, and improving the overall efficiency and quality of audit work.

[0066] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The description and examples are to be regarded as exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0067] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A text recommendation system for interpreting audit reports based on a large model, characterized in that: include: The audit knowledge graph construction large model unit is used to perform semantic analysis on the text content of the audit report, identify and extract the key entities and relationships therein, and construct a structured audit knowledge graph based on the key entities and their relationships; wherein the key entities identified and extracted by the audit knowledge graph construction large model unit include: audit object, audit objective, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; the relationships describe the mutual relationships between the above key entities; The report key description information extraction unit is used to refine the information of each audit report template that can be recommended, identify and extract the report key description information with a correlation value higher than the threshold with the key entity in the audit knowledge graph; The audit text recommendation large model unit is used to receive the audit description content input by the user, identify the key entities and their relationships in the audit description content in combination with the audit knowledge graph, automatically generate recommended prompt words through prompt word engineering, and guide the audit text recommendation large model to respond to the recommended prompt words, score the applicability of the key description information of the report, and recommend the audit report template with the highest score. The audit text recommendation large model unit pre-processes the audit description content input by the user; identifies the key entities in the audit description content through named entity recognition technology, and extracts the relationships between the key entities based on the audit knowledge graph; the entities identified in the audit description content are represented as entity sets ,in Indicates entity and The relationship between ; Based on the extracted key entities and their relationships, the recommended prompt words containing the recommended task description and the recommended boundary restrictions are automatically written to provide clear guidance information for the audit report template recommendation; Among them, for the recommended task description, the identified key entities and relationships are mapped to the template of the audit task description. This mapping process uses the following formula to represent the generation of the task description: ,in Represents the recommendation task description, function It means that a suitable description template is selected according to the entity set and the relationship set, and the entity name is filled in to form a recommended task description; for the recommended boundary restrictions, the recommended boundary restrictions are generated according to the context of the entity and the relationship, combined with the constraint condition set containing specific rules of the audit field; the key description information in each audit report template is compared with the recommended prompt words. The audit text recommendation model compares the similarity between the recommended prompt words and the key description information, and outputs the applicability score; the calculation method of the applicability score is: in represents the suitability score, is the number of entities and relations in the recommended prompt words, is the weight value of the i-th entity, Represents the i-th entity evaluated by the large model With key description information The similarity.

2. The system according to claim 1, characterized in that The audit knowledge graph is constructed by the following method to carry out targeted training and fine-tuning of the audit knowledge graph. The method includes: preparing a data set, collecting and organizing a data set containing diverse and representative audit reports; selecting a pre-trained model, and selecting a pre-trained model based on the Transformer architecture as a basis; inputting the audit report text into the Transformer pre-trained model, extracting the feature representation of the audit report text, and adding a new hierarchy on the basis of the Transformer pre-trained model, which is specifically used to identify and classify key entities and their relationships in the audit report; in the fine-tuning process, for entity and relationship recognition tasks, a cross-entropy loss function is used to calculate the difference between the predicted label and the true label of the entity type and entity relationship, and a gradient descent algorithm is used to update the weight of the model; and a model evaluation is performed on the fine-tuned pre-trained model.

3. The system according to claim 2, characterized in that The report key description information extraction unit takes each sentence in the audit report template as a unit and calculates the correlation value between the unit and the entity in the audit knowledge graph as follows: in Represents a sentence unit in an audit report template and audit entities in the knowledge graph The correlation value of and are the vector representations of the sentence unit and entity in the word space, and The modulus lengths of the vectors representing sentence units and entities respectively; for sentence units in the audit report template whose correlation values ​​with key entities in the audit knowledge graph are higher than a preset threshold, they are extracted as key description information of the report.

4. A text recommendation method for interpreting audit reports based on a large model, characterized in that: include: Audit knowledge graph construction steps: construct a large model through the audit knowledge graph, perform semantic analysis on the text content of the audit report, identify and extract the key entities and relationships therein, and construct a structured audit knowledge graph based on the key entities and their relationships; wherein the key entities identified and extracted by the audit knowledge graph construction large model include: audit object, audit objective, audit-related facts, audit issues, audit risks, audit rules, audit standards and audit conclusions; the relationships describe the mutual relationships between the above key entities; Extraction step of key descriptive information of the report: for each audit report template that can be recommended, perform refined information extraction, identify and extract key descriptive information of the report whose correlation value with key entities in the audit knowledge graph is higher than the threshold; Audit text recommendation step: receiving the audit description content input by the user, combining it with the audit knowledge graph, identifying the key entities and their relationships in the audit description content, automatically generating recommended prompt words through prompt word engineering, and guiding the audit text recommendation model to respond to the recommended prompt words, scoring the applicability of the key description information of the report, and recommending the audit report template with the highest score. The audit text recommendation step specifically includes: preprocessing the audit description content input by the user; identifying the key entities in the audit description content through named entity recognition technology, and extracting the relationships between the key entities based on the audit knowledge graph; for the entities identified in the audit description content, representing them as entity sets ,in Indicates entity and The relationship between ; Based on the extracted key entities and their relationships, the recommended prompt words containing the recommended task description and the recommended boundary restrictions are automatically written to provide clear guidance information for the audit report template recommendation; Among them, for the recommended task description, the identified key entities and relationships are mapped to the template of the audit task description. This mapping process uses the following formula to represent the generation of the task description: ,in Represents the recommendation task description, function It means that a suitable description template is selected according to the entity set and the relationship set, and the entity name is filled in to form a recommended task description; for the recommended boundary restrictions, the recommended boundary restrictions are generated according to the context of the entity and the relationship, combined with the constraint condition set containing specific rules of the audit field; the key description information in each audit report template is compared with the recommended prompt words. The audit text recommendation model compares the similarity between the recommended prompt words and the key description information, and outputs the applicability score; the calculation method of the applicability score is: in represents the suitability score, is the number of entities and relations in the recommended prompt words, is the weight value of the i-th entity, Represents the i-th entity evaluated by the large model With key description information The similarity.

5. The method according to claim 4, characterized in that The audit knowledge graph construction step also specifically includes: targeted training and fine-tuning of the audit knowledge graph to construct a large model, including: data set preparation, collecting and organizing a data set containing diverse and representative audit reports; pre-training model selection, selecting a pre-training large model based on the Transformer architecture as a basis; inputting the audit report text into the Transformer pre-training large model, extracting the feature representation of the audit report text, and adding a new hierarchy on the basis of the Transformer pre-training large model, which is specifically used to identify and classify key entities and their relationships in the audit report; in the fine-tuning process, for entity and relationship recognition tasks, the cross entropy loss function is used to calculate the difference between the predicted label and the true label of the entity type and entity relationship, and the gradient descent algorithm is used to update the model weight; the fine-tuned pre-training large model is evaluated.

6. The method according to claim 5, characterized in that The audit knowledge graph construction steps also specifically include: performing entity normalization to normalize the extracted entities; performing relationship modeling to establish a relationship model based on the relationship between entities analyzed by the large model; performing graph construction to integrate entities and their relationships into the knowledge graph using a graph database or a graph computing framework, with each entity serving as a node in the graph and each relationship serving as an edge connecting the nodes.

7. The method according to claim 6, characterized in that The step of extracting key descriptive information of the report specifically includes: taking each sentence in the audit report template as a unit, calculating the correlation value between the unit and the entity in the audit knowledge graph as follows: in Represents a sentence unit in an audit report template and audit entities in the knowledge graph The correlation value of and are the vector representations of the sentence unit and entity in the word space, and The modulus lengths of the vectors representing sentence units and entities respectively; for sentence units in the audit report template whose correlation values ​​with key entities in the audit knowledge graph are higher than a preset threshold, they are extracted as key description information of the report.

Images