Auditing knowledge graph construction method and system based on large language model technology

Abstract

The invention discloses an audit knowledge graph construction method and system based on a large language model technology, and relates to the technical field of artificial intelligence, and the method comprises the following steps: obtaining audit text content, analyzing the audit text content to obtain different types of main entities, and dividing the main entities into different text fragments, obtaining a main text statement formed by the main entity and the corresponding relationship under each fragment; obtaining text level data and noise influence data under different text segments, constructing a text statement deviation model based on a neural network algorithm, and obtaining a text statement deviation degree under each segment; performing preliminary screening according to the text statement deviation degree and a preset text content accurate range to obtain a preliminary text content set; and performing secondary screening on the preliminary text content set and the main text statement based on semantic similarity detection to obtain a screened final text content set to construct the audit knowledge graph. By analyzing and screening the deviation degree of the statements, the problems of poor accuracy and complexity in the auditing knowledge graph are solved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to the field of audit knowledge graph construction methods and system technologies based on large language model technology. Background Art

[0002] Auditing is an independent and systematic inspection process that aims to evaluate and verify the financial statements, operational activities, compliance and other aspects of a company, organization or individual to ensure their accuracy, legality and effectiveness. Auditing can involve financial auditing, internal auditing, compliance auditing, information system auditing and other fields, and is usually conducted by professional auditors or auditing institutions.

[0003] Auditing involves a wide variety of knowledge sources. To improve auditing efficiency and enhance data analysis capabilities, this can be achieved by constructing an auditing knowledge graph. The auditing knowledge graph is a graphical model for representing and managing knowledge in the auditing field. It connects auditing-related concepts, terms, processes, standards and other information in the form of nodes and edges through a graph, thereby constructing a structured and queryable knowledge system. This graph can not only help audit professionals to efficiently understand and apply auditing-related knowledge, but also improve the efficiency of decision support and risk assessment in auditing practice.

[0004] The above disclosed technical solutions have at least the following technical problems: The traditional audit knowledge graph has a relatively simple relationship, contains less information and relationships, and has a relatively simple structure with poor accuracy, which may not be able to fully cover various complex scenarios and problems in the audit field. This results in incomplete information coverage and limited thrust capacity of the audit knowledge graph, and it cannot adapt to complex audit tasks. In view of the above problems, the present invention proposes a solution. Summary of the invention

[0005] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an audit knowledge graph construction method and system based on large language model technology, which solves the problems of poor accuracy and complexity in the construction of the audit knowledge graph by analyzing and screening the deviation of sentences.

[0006] To achieve the above object, the present invention provides the following technical solutions: The method for constructing an audit knowledge graph based on a large language model technology comprises the following steps: obtaining audit text content, parsing the audit text content to obtain main entities of different categories, dividing the audit text content into different text segments according to the main entities of different categories, and obtaining a main text sentence consisting of the main entity and the corresponding relationship under each segment; obtaining text level data and noise influence data under different text segments; constructing a text sentence deviation model based on a neural network algorithm according to the text level data and the noise influence data, and obtaining the text sentence deviation degree under each segment; performing preliminary screening according to the text sentence deviation degree and a preset text content accuracy range to obtain a preliminary text content set; performing secondary screening on the preliminary text content set and the main text sentence based on semantic similarity detection to obtain a final text content set after screening, and constructing an audit knowledge graph based on the final text content set.

[0007] In a preferred embodiment, the audit text content is parsed to obtain main entities of different categories, the audit text content is divided into different text segments according to the main entities of different categories, and the main text sentence consisting of the main entity and the corresponding relationship under each segment is obtained. The specific steps are as follows: obtain the attributes of the entity in the audit text content, and identify different entity types by clustering analysis on the entity attributes; take the different entity types as the main entity, and divide the audit text into different text segments with each main entity as the content core according to the main entity; obtain multiple relationship descriptions corresponding to each main entity, and based on contextual semantic understanding, obtain the frequency of all relationship descriptions, and take the relationship description with the largest frequency as the relationship description with the highest intensity; combine the main entity with the relationship description with the highest intensity to construct a combined sentence, and set the combined sentence as the main text sentence under the current segment.

[0008] In a preferred embodiment, the text-level data includes a vocabulary-level difference coefficient and a syntactic structure difference coefficient. The vocabulary-level difference coefficient is specifically obtained by the following method: obtain multiple sub-entities and corresponding relationships under the current segment, splice the sub-entities and corresponding relationships to obtain spliced ​​sub-entity sentences; perform word segmentation on the sub-entity sentences and the main text sentences to obtain main text keywords and sub-entity keywords; perform word frequency analysis based on the main text keywords through the sub-entity keywords to obtain common-frequency keywords; convert the main text keywords, sub-entity keywords and common-frequency keywords into sub-entity vocabulary vectors, main text vocabulary vectors and common-frequency vocabulary vectors through a word vector model; obtain the common-frequency vocabulary proportion based on the total main text vocabulary vector; calculate the vocabulary-level difference coefficient based on the sub-entity vocabulary vector and the main text vocabulary vector combined with the common-frequency vocabulary proportion based on a preset vocabulary-level difference formula.

[0009] In a preferred embodiment, the syntactic structure difference coefficient is specifically obtained by the following method: obtaining the main text sentence and the sub-entity sentence of the current segment; analyzing the main text sentence and the sub-entity sentence based on dependency syntax to generate a corresponding dependency tree structure, wherein the dependency tree structure includes a main text sentence syntax tree and a sub-entity statement syntax tree; finding the number of structural nodes that match the main text sentence and the sub-entity sentence in the dependency tree structure according to the dependency tree structure; obtaining the total number of nodes in the dependency tree structure, and calculating the syntactic structure difference coefficient based on a preset syntactic structure difference calculation formula in combination with the number of matching structural nodes.

[0010] In a preferred embodiment, the noise impact data includes a noise impact coefficient, and the noise impact coefficient is specifically obtained by the following method: obtaining historical noise impact data, performing data cleaning and feature extraction on the historical noise impact data to obtain the degree of influence of the noise data on the deviation of the text sentence; performing regression analysis on the degree of influence of the deviation of the text sentence to obtain a signal strength influence factor and a noise strength influence factor; obtaining signal strength data and noise strength data for each text segment; and calculating the noise impact coefficient based on the signal strength influence factor and the noise strength influence factor based on a preset noise impact formula.

[0011] In a preferred embodiment, the preliminary screening is performed according to the text sentence deviation degree and the preset text content accurate range to obtain a preliminary text content set, and the specific steps are as follows: obtain multiple text sentence deviation degrees, calculate the multiple text sentence deviation degrees to obtain the average of the text sentence deviation degrees and the standard deviation of the text sentence deviation degrees; obtain the maximum and minimum values ​​of the text content accurate range based on the average of the text sentence deviation degrees and the standard deviation of the text sentence deviation degrees; set the maximum and minimum values ​​as boundaries as the text content accurate range; judge based on the text content accurate range according to the text sentence deviation degree; if the text sentence deviation degree is greater than the minimum value and less than the maximum value, it is within the text content accurate range; if the text sentence deviation degree is less than the minimum value or greater than the maximum value, it is outside the text content accurate range; integrate all text sentences within the text content accurate range to obtain a preliminary text content set.

[0012] In a preferred embodiment, the preliminary text content set and the main text sentence are subjected to secondary screening based on semantic similarity detection to obtain a final text content set after screening, and an audit knowledge graph is constructed based on the final text content set. The specific steps are as follows: obtaining a preliminary text content set, preprocessing the preliminary text content set and the main text sentence, and obtaining all sentence vectors and main text sentence vectors in the preliminary text content set based on a preset BERT model; respectively calculating the cosine similarity of the sentence vectors in the preliminary text content set and the main text sentence vectors; collecting the text content set sentences corresponding to the sentence vectors in the text content set with a cosine similarity of 1 to construct a final text content set; storing and managing the final text content set based on a Neo4j graph database, and constructing a knowledge graph.

[0013] One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. By comprehensively analyzing differences in vocabulary, syntactic structure and noise, an efficient and accurate audit knowledge graph deviation model is constructed, which can automatically identify and optimize the semantic relationship and structure of the text, improve the accuracy, complexity and scalability of graph construction, and enhance the robustness and flexibility of the system.

[0014] 2. By combining text deviation analysis and semantic similarity detection, using the BERT model and Neo4j graph database, it is possible to efficiently and accurately screen, integrate and manage text content, build a high-quality audit knowledge graph, and have the advantages of automated and intelligent text screening and knowledge management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the structure of the audit knowledge graph construction method based on the large language model technology provided in an embodiment of the present application.

[0016] Figure 2 A schematic diagram of the structure of the audit knowledge graph construction system based on the large language model technology provided in the embodiment of the present application. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] Embodiment 1, Figure 1 The schematic diagram of the structure of the audit knowledge graph construction method based on the large language model technology provided in the embodiment of the present application includes the following steps: S1, obtain the audit text content, divide the audit text content into different text segments, parse the audit text content of different text segments to obtain main entities of different categories, and obtain the main entity and the corresponding relationship under each segment to form a main text sentence.

[0019] In this example, the audit text content is obtained, and the audit text content includes audit standards, audit reports and cases. Audit reports refer to various types of audit reports, including financial statement audit reports, internal audit reports, compliance audit reports, etc. Audit cases refer to classic audit failure cases, cases of illegal audit behavior, practical cases of audit procedures and methods, etc., various forms of evidence, such as documents, bills, contracts, accounting vouchers, etc., audit working papers used by auditors to verify the authenticity and compliance of the audit object, financial statements and accounting standards, including balance sheets, income statements, cash flow statements, etc. Auditors usually need to analyze and verify these financial information and audit methods, and describe various methods commonly used in the audit process, such as sampling, interviews, observations, tests, etc., corporate financial information and operating data, etc. The audit text content is parsed to obtain different categories of main entities, and the audit text content is divided into different text segments according to different categories of main entities, and the main text sentences consisting of the main entities and corresponding relationships under each segment are obtained. The specific steps are as follows: Obtain the attributes of entities in the audit text content, and identify different entity types by clustering the attributes of the entities; The different entity types are taken as main entities, and the audit text is divided into different text segments with each main entity as the content core according to the main entity; Obtain multiple relationship descriptions corresponding to each main entity, obtain the frequencies of all relationship descriptions based on contextual semantic understanding, and take the relationship description with the largest frequency as the relationship description with the highest intensity; Combine the main entity with the highest intensity relationship description to construct a combined sentence, and set the combined sentence as the main text sentence under the current fragment.

[0020] It should be noted that dividing the audit text content into different text fragments helps to ensure that the structure of the graph is clear, the semantics are accurate, and it has good query efficiency. The attributes of entities and the relationships between them are important bases for dividing entity categories. An important criterion for entity division is the similarities and differences of attributes. If two entities have similar attributes, they may belong to the same category. The division of attributes can help you organize related entities into different hierarchies or categories. The main entity refers to the core entity in the knowledge graph, usually the most important and representative entity in a specific field. For example, the "audit object" in the audit knowledge graph may be a main entity, representing a specific goal that needs to be paid attention to during the audit process. In the audit field, common entities include: audit objects (such as companies or financial statements), audit standards, auditors, audit methods, audit reports, etc.

[0021] S2, obtain text-level data and noise impact data for different text segments. Among them, the text-level data includes the vocabulary-level difference coefficient and the syntactic structure difference coefficient. The vocabulary-level difference coefficient refers to the degree of vocabulary difference between the relevant text sentences constructed by splicing the secondary entity and the corresponding relationship description in the current text segment and the main text sentence in the segment. Analyzing the vocabulary-level difference coefficient has the following advantages for enhancing the accuracy and complexity of the construction of the audit knowledge graph: Improve semantic accuracy: The vocabulary-level difference coefficient helps measure the degree of vocabulary differences between different audit documents or audit data. By quantifying the differences between different audit texts, the vocabulary-level difference coefficient can help the audit knowledge graph identify the subtle differences in keywords, terms, and concepts during the construction process. This helps to accurately capture the ambiguity and professionalism of terms in different contexts in the audit field and improve the semantic accuracy of the graph.

[0022] Optimize the hierarchical structure of the knowledge graph: When constructing an audit knowledge graph, the vocabulary-level difference coefficient can be used to measure the similarity or difference between different concepts, helping to clarify which concepts are similar and which are related but not exactly the same. This difference measurement can optimize the hierarchical structure of the graph so that similar audit concepts are clustered together, while different concepts can be reasonably separated, avoiding the knowledge graph from being too complex or having inaccurate associations.

[0023] For example, in the auditing field, although "compliance audit", "financial audit", and "information system audit" are all part of auditing, they have significant differences at the lexical level. The lexical level difference coefficient can identify these differences and correctly map them to different levels or categories in the knowledge graph.

[0024] Enhance the scalability of the knowledge graph: With the continuous development of the audit field and the continuous application of new technologies, new audit methods, tools, regulations and technologies continue to emerge. By using the vocabulary level difference coefficient, the difference between new and old audit concepts can be effectively identified, so that new knowledge can be more accurately integrated into the existing knowledge graph. In this way, the knowledge graph can maintain accuracy while maintaining adaptability and scalability to emerging fields.

[0025] For example, as the application of artificial intelligence in auditing continues to increase, the lexical level difference coefficient can help distinguish the differences between the concepts of "artificial intelligence auditing" and traditional "manual auditing", ensuring that new fields can be smoothly integrated into the knowledge graph.

[0026] Improve the diversity and richness of the knowledge graph: In the audit field, different audit tasks, audit objectives, audit methods and audit tools may vary greatly. By analyzing the differences at the vocabulary level, the vocabulary level difference coefficient can help identify these differences, thereby enriching the content of the audit knowledge graph and making the graph more diverse. This can not only improve the coverage of the audit knowledge graph, but also make it more flexible in dealing with different audit scenarios and support multiple types of audit analysis and decision-making.

[0027] Improve the balance between the complexity and simplicity of the knowledge graph: The vocabulary-level difference coefficient can help builders make more reasonable choices in the knowledge graph. When the differences between different concepts are large, the vocabulary-level difference coefficient can prompt that these concepts need to be further split to avoid mistakenly merging them into the same node; conversely, when the differences are small or the concepts are similar, the vocabulary-level difference coefficient can prompt that these concepts can be integrated into one node. In this way, the vocabulary-level difference coefficient helps to achieve the reasonable stratification and integration of the knowledge graph, thereby maintaining the integrity of the information while avoiding the graph being too complex or too simple, ensuring its efficiency and practicality.

[0028] Improve the effectiveness of the automated construction process: The construction of modern audit knowledge graphs often relies on automated technologies such as natural language processing (NLP) and machine learning (ML). In this automated process, the vocabulary-level difference coefficient can be used as an important metric to help the automated system determine which text data have large differences, thereby more accurately separating or integrating different concepts. With this automated difference coefficient analysis, the accuracy and efficiency of the knowledge graph construction process can be effectively improved.

[0029] The specific method for obtaining the vocabulary-level difference coefficient is as follows: Obtain multiple sub-entities and corresponding relationships under the current segment, and concatenate the sub-entities and the corresponding relationships to obtain a concatenated sub-entity statement; Perform word segmentation on the secondary entity sentences and the main text sentences to obtain the main text key words and the secondary entity key words; Through the sub-entity key words, the frequency analysis is performed based on the main text key words to obtain the co-frequency words; The main text key words, sub-entity key words and co-frequency words are converted into sub-entity word vectors, main text word vectors and co-frequency word vectors through a word vector model; The co-frequency vocabulary vector is based on the total main text vocabulary vector to obtain the co-frequency vocabulary ratio; The secondary entity vocabulary vector and the main text vocabulary vector are combined with the proportion of co-frequency words based on a preset vocabulary level difference formula to calculate the vocabulary level difference coefficient.

[0030] The specific calculation formula for the proportion of co-frequency words is as follows:

[0031] The specific calculation formula of the vocabulary level difference coefficient is as follows:

[0032] In the formula, is the coefficient of variation at the lexical level, For the The secondary entity vocabulary vectors of the sentences, is the main text vocabulary vector, is the proportion of co-frequency words, For the The co-frequency word vectors of sentences, where =1,2,3,...,R, R is an integer, N is the total number of statements, is the total main text vocabulary vector.

[0033] It should be noted that the word vector model is a technology in natural language processing (NLP) that is used to map words into a continuous vector space. These vectors represent the semantic information of words and can capture the similarity, relevance, and contextual relationship between words. The core idea of ​​the word vector model is to represent each word as a vector and train the model so that words with similar semantics are closer in the vector space.

[0034] The syntactic structure difference coefficient refers to the difference in grammatical structure between the main text sentence and the secondary entity sentence. The difference between sentences is evaluated by comparing the grammatical composition of sentences, the arrangement of sentence components, and the differences in grammatical relations. Analyzing the syntactic structure difference coefficient has the following advantages for enhancing the accuracy and complexity of the construction of the audit knowledge graph: Improve the accuracy of semantic extraction: The syntactic structure difference coefficient helps to reveal the grammatical and semantic differences between different texts by quantitatively analyzing the structure of text sentences. This is especially important for semantic extraction in knowledge graphs. Literature and cases in the audit field often use complex legal terms, accounting concepts, and detailed audit rules. Through the analysis of the syntactic structure difference coefficient, key concepts, terms, and their relationships can be more accurately extracted from complex sentences, avoiding the error of ignoring deep semantics due to superficial language similarities.

[0035] Optimize the identification and classification of entity relationships: The core of the audit knowledge graph is to identify and accurately represent the relationship between entities (such as audit standards, audit methods, audit reports, etc.). The syntactic structure difference coefficient can help analyze the relative position and relationship of each entity in the sentence, and then infer the grammatical dependency between different entities. This is very important for entity classification and relationship extraction in the graph. For example, when analyzing audit literature, the relationship between "audit report" and "audit conclusion" can be more clearly identified through syntactic structure analysis, and the correct edges and nodes can be reasonably established in the knowledge graph.

[0036] For example, in the process of constructing an audit knowledge graph, if the system can recognize the syntactic differences between the two terms "audit report" and "audit opinion", it can more accurately establish the relationship between them and distinguish their different semantic categories, thereby improving the accuracy of graph construction.

[0037] Handling complex text and ambiguity: Auditing literature often contains many professional terms, legal provisions, and complex narrative structures, which makes text understanding and processing extremely complex. The syntactic structure difference coefficient can help the system identify and handle ambiguity and ambiguity in sentences. By quantifying the differences between different syntactic structures, the system can better identify which sentences have the same grammatical structure and which have potential differences in meaning, thereby effectively reducing false associations or misclassifications in the knowledge graph.

[0038] For example, some audit documents may contain both "management" and "auditors" as subjects, but due to differences in grammatical structure, they may carry different semantics. Through syntactic analysis, the system can accurately distinguish these polysemous expressions and avoid mistakenly confusing them as the same entity.

[0039] Enhance the complexity and flexibility of the knowledge graph: The application of the syntactic structure difference coefficient not only helps to improve the accuracy of the knowledge graph, but also enhances its flexibility and scalability in the face of complex scenarios. By identifying subtle syntactic differences between different texts, the system can flexibly respond to changing regulations and standards in the audit field, making the knowledge graph more adaptable when updating.

[0040] For example, the syntactic structure in the text may change. The analysis of the syntactic structure difference coefficient helps the system automatically identify these changes and update the entities and relationships in the graph in a timely manner to ensure that the graph always stays synchronized with the latest audit practices.

[0041] The syntactic structure difference coefficient is specifically obtained as follows: Get the main text sentence and sub-entity sentence of the current fragment; Analyze the main text sentence and the secondary entity sentence based on dependency syntax to generate a corresponding dependency tree structure, wherein the dependency tree structure includes a main text sentence syntax tree and a secondary entity sentence syntax tree; Find the number of structure nodes that match the main text sentence and the secondary entity sentence in the dependency tree structure according to the dependency tree structure; The total number of nodes in the dependency tree structure is obtained, and the syntactic structure difference coefficient is calculated based on the preset syntactic structure difference calculation formula in combination with the number of matching structure nodes.

[0042] The specific calculation formula of the syntactic structure difference coefficient is as follows:

[0043] In the formula, is the syntactic structure difference coefficient, is the number of matching structure nodes, is the total number of nodes in the dependency tree structure.

[0044] It should be noted that dependency syntactic analysis is a syntactic analysis method in natural language processing, which aims to reveal the dependency relationship between words in a sentence. It constructs a tree structure by identifying the grammatical relationship between each word (or term) in a sentence and other words, indicating how these words are related to each other. The generation of the dependency tree structure can be understood as constructing the dependency relationship between each word and other words based on the sentence structure, grammatical rules, grammatical models, etc. Through the identified dependency relationship, a tree structure is finally generated to represent the grammatical structure in the sentence. The dependency tree indicates that each word and its syntactic relationship (subject-predicate relationship, modification relationship, object relationship, etc.) are connected into a tree structure through dependency relationships. The nodes of the dependency tree represent the vocabulary, and the edges of the dependency tree represent the syntactic relationship between the vocabulary. According to the dependency tree structure, the number of structural nodes matching the main text sentence and the secondary entity sentence in the dependency tree structure is found, which can be understood as obtaining the number of structural nodes matching between the two grammatical tree structures through the similarity between the nodes and edges in the grammatical tree of the main text sentence and the grammatical tree of the secondary entity sentence.

[0045] Noise impact data includes the noise impact coefficient. The noise impact coefficient refers to the fact that during the construction of the knowledge graph, the impact of noise data can significantly increase the deviation of the text content. Noise affects the quality of the text, resulting in a decrease in the accuracy and completeness of the information. The knowledge graph relies on information extracted from the text. If the original text contains a large amount of irrelevant or erroneous information, the knowledge graph may store this inaccurate information as facts, resulting in deviation. Analyzing the noise impact coefficient has the following advantages for enhancing the accuracy and complexity of the construction of the audit knowledge graph: Improve accuracy: In audit data, noise may come from data inconsistency, redundancy or errors. The noise impact coefficient helps to identify these unreliable parts and reduce their interference with the final knowledge graph through weighting or suppression mechanisms. For example, in audit data, if the association strength of a node or edge is affected by noise, these abnormal data can be weakened by adjusting the noise impact coefficient, thereby improving the accuracy of the graph. By controlling the noise impact coefficient, the uncertainty of the data can be better managed and reasoning errors caused by noise can be avoided. Audit knowledge graphs often require complex reasoning and pattern recognition. The interference of noise may lead to deviations in the reasoning results. Reducing the impact of noise can ensure that the reasoning process is more accurate.

[0046] Optimizing the complexity of the graph: Noise often increases the complexity of the graph because it causes the generation of redundant nodes, edges, or relationships, making the graph large and difficult to manage. By analyzing the noise impact coefficient, this redundant information can be effectively identified and removed, simplifying the structure of the knowledge graph and reducing the interference of irrelevant information. In the audit knowledge graph, noise may increase the computational cost, especially in the reasoning and query process of the graph. The noise impact coefficient can be used to optimize the processing of the graph, making the query and reasoning process more efficient, thereby reducing the consumption of computing resources.

[0047] Enhanced robustness: Audit data is often multi-source, multi-dimensional, and contains a certain amount of noise. The noise impact coefficient helps to improve the robustness of the knowledge graph when facing imperfect data. For example, when facing missing values, erroneous data, or heterogeneous data, the noise impact coefficient can help evaluate which parts of the noise can be ignored and which require special attention, thereby improving overall stability and reliability. Through the analysis of the noise impact coefficient, the graph can tolerate a certain amount of error when facing abnormal situations or deviations, and will not be disturbed by extreme noise, maintaining a high performance.

[0048] Support incremental updates: As audit data is updated and the graph continues to evolve, the impact of noise may also change. The noise impact coefficient can play an important role in the incremental update of the knowledge graph. By dynamically adjusting the noise impact coefficient, the structure and content of the graph can be gradually optimized during the update process to ensure the accuracy and effectiveness of the graph. The noise impact coefficient provides an adaptive mechanism that allows the knowledge graph to flexibly adjust the noise weight when facing new data or new types of noise, ensuring that the graph can still maintain high performance in a changing environment.

[0049] Enhance semantic relevance: Optimize the semantic links of the knowledge graph. Noise not only affects the quality of nodes, but may also affect the relationship between nodes. By adjusting the noise impact coefficient, it can help identify meaningless or erroneous relationships, thereby improving the semantic association between nodes in the graph and making the graph more accurate and expressive. Improve the quality of data fusion. In the audit knowledge graph, data often comes from different sources, and there may be inconsistent formats or unclear concepts. The noise impact coefficient helps balance the impact of different data sources, thereby improving the consistency and correctness of the data fusion process and avoiding interference from irrelevant or misleading information.

[0050] The noise influence coefficient is specifically obtained as follows: Acquire historical noise impact data, perform data cleaning and feature extraction on the historical noise impact data to obtain the degree of influence of the noise data on the deviation of the text sentence; The influence degree of text sentence deviation is subjected to regression analysis to obtain the signal strength influence factor and the noise strength influence factor; Obtain signal strength data and noise strength data for each text segment; The noise impact coefficient is calculated based on the signal strength impact factor and the noise strength impact factor based on a preset noise impact formula.

[0051] The noise influence coefficient is specifically calculated as follows:

[0052] In the formula, is the noise influence coefficient, is the influence of noise on the deviation of text sentences, is the signal strength data, is the signal strength influencing factor, is the noise intensity data, is the noise intensity influencing factor, The inherent complexity of the data affects the data.

[0053] S3, builds a text sentence deviation model based on the neural network algorithm according to the text level data and noise impact data, and obtains the text sentence deviation degree under each fragment.

[0054] In this example, the text sentence deviation model has the following specific steps: Acquire historical text level data and historical noise impact data, and pre-process the historical text level data and historical noise impact data; The preprocessed data is divided into a training set and a test set. Based on the neural network algorithm, the training set is used as the input item to train the model, and the degree of influence of text-level data and noise-affected data on the deviation of text sentences is trained. The weights are automatically iterated and adjusted, and the corresponding weights are assigned to the text-level data and the noise-affected data, respectively, to build a text sentence deviation model. According to the text sentence deviation model, the text sentence deviation degree of each fragment is obtained.

[0055] The specific calculation formula of the text sentence deviation model is as follows:

[0056] In the formula, is the text sentence deviation, is the coefficient of variation at the lexical level, is the syntactic structure difference coefficient, is the noise influence coefficient, is the weight coefficient of the vocabulary level difference coefficient, is the weight coefficient of the syntactic structure difference coefficient, is the noise impact coefficient weight coefficient.

[0057] S4, perform preliminary screening according to the text sentence deviation and the preset text content accuracy range to obtain a preliminary text content set. The specific steps are as follows: Obtaining multiple text sentence deviations, and calculating the multiple text sentence deviations to obtain an average of the text sentence deviations and a standard deviation of the text sentence deviations; Obtaining the maximum and minimum values ​​of the accuracy range of the text content based on the average of the deviation degree of the sentence and the standard deviation of the text sentence deviation degree; The maximum value and the minimum value are used as boundaries to set the accurate range of the text content; Judging based on the accuracy range of the text content according to the text sentence deviation degree; If the text sentence deviation is greater than the minimum value and less than the maximum value, it is within the accurate range of the text content; If the text sentence deviation is less than the minimum value or greater than the maximum value, it is outside the accurate range of the text content; All text sentences within the accurate range of text content are integrated to obtain a preliminary text content set.

[0058] The maximum value of the accurate range of the text content is calculated by the following formula:

[0059] The minimum value of the accurate range of the text content, the specific calculation formula is as follows:

[0060] In the formula, is the maximum value of the accurate range of the text content. is the average deviation of text sentences, is a constant, indicating the multiple of the deviation, is the standard deviation of the text sentence deviation.

[0061] S5, performing secondary screening on the preliminary text content set and the main text sentence based on semantic similarity detection to obtain the screened final text content set, and constructing an audit knowledge graph based on the final text content set. The specific steps are as follows; Obtaining a preliminary text content set, preprocessing the preliminary text content set and the main text sentence, and obtaining all sentence vectors in the preliminary text content set and the main text sentence vector based on a preset BERT model; Calculate the cosine similarity of the sentence vectors in the preliminary text content set and the main text sentence vectors respectively; Collect the sentences in the text content set corresponding to the sentence vectors in the text content set with a cosine similarity of 1 to construct the final text content set; The final text content set is stored and managed based on the Neo4j graph database, and a knowledge graph is constructed.

[0062] The specific calculation formula of the cosine similarity is as follows:

[0063] In the formula, is the cosine similarity, is the sentence vector for the preliminary text content concentration, is the main text sentence vector.

[0064] It should be noted that BERT is a pre-trained language model. BERT can capture the subtle semantics of vocabulary through context and generate a vector representation at the sentence level. The output of the BERT model is usually a high-dimensional vector representation of the sentence. Neo4j is an open source graph database based on graph theory. It is particularly good at processing structured data with complex relationships between them. Unlike traditional relational databases (such as MySQL or PostgreSQL), Neo4j uses a graph model to store data, which is particularly suitable for representing and querying highly related data. Neo4j uses nodes, relationships, and properties to build graphs, which is suitable for application scenarios that require fast search in a large number of complex relationships.

[0065] Embodiment 2, Figure 2 The schematic diagram of the structure of the audit knowledge graph construction system based on the large language model technology provided in the embodiment of the present application includes a text segmentation module, a data acquisition module, a model construction module, a preliminary screening module, and a secondary screening and knowledge graph construction module, and there are connections between the modules: A text segmentation module is used to obtain audit text content, parse the audit text content to obtain main entities of different categories, divide the audit text content into different text segments according to the main entities of different categories, and obtain the main text sentence consisting of the main entity and the corresponding relationship under each segment; A data acquisition module is used to obtain text-level data and noise impact data under different text segments; A model building module is used to build a text sentence deviation model based on a neural network algorithm according to text level data and noise impact data, and obtain the text sentence deviation degree under each fragment; A preliminary screening module is used to perform preliminary screening based on the text sentence deviation and the preset text content accuracy range to obtain a preliminary text content set; The secondary screening and knowledge graph construction module is used to perform secondary screening on the preliminary text content set and the main text sentence based on semantic similarity detection to obtain the final text content set after screening, and to construct an audit knowledge graph based on the final text content set.

[0066] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0067] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product.

[0068] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0069] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0070] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0071] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The audit knowledge graph construction method based on large language model technology is characterized by: The steps include: Acquire audit text content, parse the audit text content to obtain main entities of different categories, divide the audit text content into different text segments according to the main entities of different categories, and obtain the main text sentence consisting of the main entity and the corresponding relationship under each segment; Obtain text-level data and noise impact data for different text segments; According to the text level data and noise impact data, a text sentence deviation model is constructed based on the neural network algorithm to obtain the text sentence deviation degree in each segment; Perform preliminary screening based on the text sentence deviation and the preset text content accuracy range to obtain a preliminary text content set; The preliminary text content set and the main text sentence are subjected to secondary screening based on semantic similarity detection to obtain the screened final text content set, and an audit knowledge graph is constructed based on the final text content set.

2. The audit knowledge graph construction method based on large language model technology according to claim 1 is characterized in that: The audit text content is parsed to obtain main entities of different categories, the audit text content is divided into different text segments according to the main entities of different categories, and the main text sentences consisting of the main entities and the corresponding relationships under each segment are obtained. The specific steps are as follows: Obtain the attributes of entities in the audit text content, and identify different entity types by clustering the attributes of the entities; The different entity types are taken as main entities, and the audit text is divided into different text segments with each main entity as the content core according to the main entity; Obtain multiple relationship descriptions corresponding to each main entity, obtain the frequencies of all relationship descriptions based on contextual semantic understanding, and take the relationship description with the largest frequency as the relationship description with the highest intensity; Combine the main entity with the highest intensity relationship description to construct a combined sentence, and set the combined sentence as the main text sentence under the current fragment.

3. The audit knowledge graph construction method based on large language model technology according to claim 1 is characterized in that: The text level data includes the vocabulary level difference coefficient and the syntactic structure difference coefficient. The specific method for obtaining the vocabulary level difference coefficient is as follows: Obtain multiple sub-entities and corresponding relationships under the current segment, and concatenate the sub-entities and the corresponding relationships to obtain a concatenated sub-entity statement; Perform word segmentation on the secondary entity sentences and the main text sentences to obtain the main text key words and the secondary entity key words; Through the sub-entity key words, the frequency analysis is performed based on the main text key words to obtain the co-frequency words; The main text key words, sub-entity key words and co-frequency words are converted into sub-entity word vectors, main text word vectors and co-frequency word vectors through a word vector model; The co-frequency vocabulary vector is based on the total main text vocabulary vector to obtain the co-frequency vocabulary ratio; The secondary entity vocabulary vector and the main text vocabulary vector are combined with the proportion of co-frequency words based on a preset vocabulary level difference formula to calculate the vocabulary level difference coefficient.

4. The audit knowledge graph construction method based on large language model technology according to claim 3 is characterized in that: The syntactic structure difference coefficient is specifically obtained as follows: Get the main text sentence and sub-entity sentence of the current fragment; Analyze the main text sentence and the secondary entity sentence based on dependency syntax to generate a corresponding dependency tree structure, wherein the dependency tree structure includes a main text sentence syntax tree and a secondary entity sentence syntax tree; Find the number of structure nodes that match the main text sentence and the secondary entity sentence in the dependency tree structure according to the dependency tree structure; The total number of nodes in the dependency tree structure is obtained, and the syntactic structure difference coefficient is calculated based on the preset syntactic structure difference calculation formula in combination with the number of matching structure nodes.

5. The audit knowledge graph construction method based on large language model technology according to claim 1 is characterized in that: The noise impact data includes the noise impact coefficient. The specific method for obtaining the noise impact coefficient is as follows: Acquire historical noise impact data, perform data cleaning and feature extraction on the historical noise impact data to obtain the degree of influence of the noise data on the deviation of the text sentence; The influence degree of text sentence deviation is subjected to regression analysis to obtain the signal strength influence factor and the noise strength influence factor; Obtain signal strength data and noise strength data for each text segment; The noise impact coefficient is calculated based on the signal strength impact factor and the noise strength impact factor based on a preset noise impact formula.

6. The audit knowledge graph construction method based on large language model technology according to claim 1 is characterized in that: The preliminary screening is performed based on the text sentence deviation and the preset text content accuracy range to obtain a preliminary text content set. The specific steps are as follows: Obtaining multiple text sentence deviations, and calculating the multiple text sentence deviations to obtain an average of the text sentence deviations and a standard deviation of the text sentence deviations; Obtaining a maximum value and a minimum value of the text content accuracy range based on the average value of the text sentence deviation degree and the standard deviation of the text sentence deviation degree; The maximum value and the minimum value are used as boundaries to set the accurate range of the text content; Judging based on the accuracy range of the text content according to the text sentence deviation degree; If the text sentence deviation is greater than the minimum value and less than the maximum value, it is within the accurate range of the text content; If the text sentence deviation is less than the minimum value or greater than the maximum value, it is outside the accurate range of the text content; All text sentences within the accurate range of text content are integrated to obtain a preliminary text content set.

7. The audit knowledge graph construction method based on large language model technology according to claim 1 is characterized in that: The preliminary text content set and the main text sentence are subjected to secondary screening based on semantic similarity detection to obtain the final text content set after screening, and the audit knowledge graph is constructed based on the final text content set. The specific steps are as follows; Obtaining a preliminary text content set, preprocessing the preliminary text content set and the main text sentence, and obtaining all sentence vectors in the preliminary text content set and the main text sentence vector based on a preset BERT model; Calculate the cosine similarity of the sentence vectors in the preliminary text content set and the main text sentence vectors respectively; Collect the sentences in the text content set corresponding to the sentence vectors in the text content set with a cosine similarity of 1 to construct the final text content set; The final text content set is stored and managed based on the Neo4j graph database, and a knowledge graph is constructed.

8. The audit knowledge graph construction method based on large language model technology according to claim 3 is characterized in that: The specific calculation formula for the proportion of co-frequency words is as follows: The specific calculation formula of the vocabulary level difference coefficient is as follows: In the formula, is the coefficient of variation at the lexical level, For the The secondary entity vocabulary vectors of the sentences, is the main text vocabulary vector, is the proportion of co-frequency words, For the The co-frequency word vectors of sentences, where =1,2,3,...,R, R is an integer, N is the total number of statements, is the total main text vocabulary vector.

9. The audit knowledge graph construction method based on large language model technology according to claim 5 is characterized in that: The noise influence coefficient is specifically calculated as follows: In the formula, is the noise influence coefficient, is the influence of noise on the deviation of text sentences, is the signal strength data, is the signal strength influencing factor, is the noise intensity data, is the noise intensity influencing factor, The inherent complexity of the data affects the data.

10. A system using the audit knowledge graph construction method based on large language model technology as described in any one of claims 1 to 9, characterized in that: It includes text segmentation module, data acquisition module, model construction module, preliminary screening module, secondary screening and knowledge graph construction module. There are connections between modules: A text segmentation module is used to obtain audit text content, parse the audit text content to obtain main entities of different categories, divide the audit text content into different text segments according to the main entities of different categories, and obtain the main text sentence consisting of the main entity and the corresponding relationship under each segment; A data acquisition module is used to obtain text-level data and noise impact data under different text segments; A model building module is used to build a text sentence deviation model based on a neural network algorithm according to text level data and noise impact data, and obtain the text sentence deviation degree under each fragment; A preliminary screening module is used to perform preliminary screening based on the text sentence deviation and the preset text content accuracy range to obtain a preliminary text content set; The secondary screening and knowledge graph construction module is used to perform secondary screening on the preliminary text content set and the main text sentence based on semantic similarity detection to obtain the final text content set after screening, and to construct an audit knowledge graph based on the final text content set.

Images