Auditing report generation method oriented to financial management system
By fine-tuning the large language model and building an external knowledge base, combining the LDA theme model and the big model to generate financial management system audit reports, the time-consuming and lack of targeted problems in the generation of audit reports in the existing technology are solved, and efficient and professional Audit report generated.
Patent Information
- Application Number
- CN202411061475.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-29
AI Technical Summary
The existing financial management system audit report generation method is time-consuming and labor-intensive, prone to human errors, and is difficult to achieve full coverage, over-reliance on manual labeling of data, and lacks targeted audit suggestions and conclusions.
A fine-tuning large language model and building an external knowledge base are used to design a standardized audit report generation method, determine the audit dimensions through the LDA theme model, build a financial management system instruction dialogue data set, and use the big model to generate a standardized audit report.
It significantly improves the speed and quality of information processing of financial management systems, realizes the automated generation and efficient summary of audit reports, provides reasonable suggestions and conclusions, and ensures the standardization and professionalization of reports.
Smart Images

Figure CN119047446A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of financial management, and more specifically to a method for generating an audit report for a financial management system. Background Art
[0002] In the field of financial management, the audit report of the financial management system plays a vital role. It aims to independently review the financial statements of financial institutions and issue audit opinions. High-quality audit reports can significantly improve the credibility of financial information and help stakeholders make more informed decisions.
[0003] Current audit reports usually require auditors to write them manually, a process that is not only time-consuming and labor-intensive, but also prone to human errors. Especially when faced with a large amount of textual materials on financial management systems, auditors need to spend a lot of time reviewing and summarizing these materials, which often cover complex rules and regulations. And because of the large number of textual materials on financial management systems, auditors can often only review them by sampling, which means that there may be important information or clauses that have not been reviewed. This sampling review method makes it difficult to achieve full coverage of system audits.
[0004] There are complex relationships between different financial management system texts, which are crucial to understanding the overall system architecture. However, manually combing through these relationships is very complicated and time-consuming, and it is easy to miss important related information.
[0005] There may be differences between the specialized terms in financial management systems and those in general management systems, even though they may have the same meaning. Such differences require auditors to invest extra effort to identify and understand these terms, thus increasing the complexity of the audit work.
[0006] In addition, current machine learning models tend to over-rely on manually labeled data when handling audit tasks, which limits the generalization ability of the model. Moreover, the audit content generated by these models is relatively simple, lacking targeted audit recommendations and conclusions, and cannot meet the needs of audit work.
[0007] Therefore, how to design an audit report generation method for financial management systems that can quickly and accurately organize and review financial management systems and significantly improve the processing speed and quality of financial management system information is an urgent problem that technical personnel in this field need to solve. Summary of the invention
[0008] In view of this, the present invention provides a method for generating audit reports for financial management systems. By combining the fine-tuning of the large language model method with the construction of an external knowledge base, a standardized financial management system audit report prompt template is designed. The text generation capability and information retrieval technology of the large model are used to fill in the report prompt template to generate a standardized financial management system audit report.
[0009] In order to achieve the above object, the present invention adopts the following technical solution:
[0010] A method for generating an audit report for a financial management system comprises the following steps:
[0011] S1. Obtain a financial management system text dataset and perform preprocessing to obtain a preprocessed financial management system text dataset;
[0012] S2. Based on the preprocessed financial management system text dataset, determine the audit dimension of the financial management system through the LDA topic model, and use the big model to build the financial management system instruction dialogue dataset;
[0013] S3. Fine-tune the big model using the financial management system instruction dialogue dataset to obtain the audit big model;
[0014] S4. Combined with the audit dimension of financial management system, construct the audit report dimension and the corresponding report prompt template, input the audit documents to be processed into the audit big model, and generate the corresponding audit report.
[0015] Among them, in S1, the preprocessing includes: removal of preset characters and punctuation, word segmentation, removal of stop words, stem extraction and word form restoration, entity recognition and normalization, semantic enhancement and anomaly detection.
[0016] Furthermore, in S2, the audit dimensions of the financial management system are determined by the LDA topic model, including:
[0017] S211. Use the LDA topic model to process the preprocessed financial management system text data set to obtain relevant topic descriptions and topic distribution;
[0018] S212. Based on the relevant topic descriptions and topic distribution, high-frequency vocabulary and keyword analysis is performed under each topic, each topic is named, and topics related to the financial management system are identified.
[0019] Furthermore, the S211 includes:
[0020] A topic was randomly assigned to each word in each financial management system document;
[0021] The topic assignment is iteratively optimized through the Gibbs sampling algorithm. The probability of each word in the document belonging to each topic is recalculated, and the topic assignment of each word is continuously adjusted.
[0022] Until the convergence condition is reached, combined with the optimal number of topics, the word distribution of each topic and the topic distribution of each financial management system document are output.
[0023] Furthermore, the optimal number of topics is determined by calculating the perplexity of the model under different numbers of topics, including:
[0024] The preprocessed financial management system text dataset is divided into training set and test set, and the range of the number of topics is determined;
[0025] Use the training set data to train LDA models with different numbers of topics, and combine the test set data to calculate the perplexity of models with different numbers of topics:
[0026]
[0027]
[0028] Where p(w) is the probability of word w appearing in the test set, p(z|d) is the probability of topic z appearing given document d, p(w|z) is the probability of word w appearing given topic z, M is the number of documents in the test set data, and N is the probability of word w appearing given document d. d is the number of words in document d;
[0029] Based on the perplexity calculation results of models with different number of topics, the number of topics with the smallest perplexity is selected as the optimal number of topics.
[0030] Furthermore, in S2, a large model is used to construct a financial management system instruction dialogue dataset, including:
[0031] S221. Process the preprocessed financial management system text data set using the big model to generate corresponding questions and answers; the questions and answers cover the main content and details of the text;
[0032] S222. Check questions and answers for accuracy, consistency and compliance with industry regulations through an automatic verification mechanism;
[0033] S223. Based on the inspection results, the questions and answers are corrected and optimized;
[0034] S224. Arrange the corrected and optimized questions and answers into a unified data format to obtain a financial management system instruction dialogue data set.
[0035] Furthermore, the S3 includes:
[0036] S31, dividing the financial management system instruction dialogue data set into a training set and a validation set;
[0037] S32, determine the large model and set hyperparameters; the hyperparameter settings include: learning rate, training rounds and batch size;
[0038] S33. Input the training set and validation set into the big model after the hyperparameters are set, and train the big model in combination with the LORA strategy to obtain the audit big model.
[0039] Furthermore, the LORA strategy includes low-rank decomposition of part of the weight matrix of the model to reduce the number of required training parameters; and using an adaptive learning rate strategy and incremental training method to update the low-rank matrix parameters.
[0040] Furthermore, in S4, the audit report dimensions and the corresponding report prompt template are constructed, including:
[0041] S411. Use crawler programs to obtain historical audit reports, archive them, and classify them;
[0042] S412. Based on the archived and classified historical audit reports, use natural language processing algorithms to identify and extract basic structures;
[0043] S413. Determine the audit dimensions, combine the basic structures identified and extracted, and construct the initial report prompt template;
[0044] S414: Automated verification and feedback optimization are performed on the report prompt template to obtain a report prompt template.
[0045] Furthermore, in S4, the audit document to be processed is input into the audit model to generate a corresponding audit report, including:
[0046] S421. Control the audit big model to process the pending audit documents through instructions; the instructions include: first instruction: extract the audit object, audit time and financial management system name from the pending audit documents; second instruction: generate a query according to the audit dimension of the financial management system, and retrieve relevant summaries, audit background and purpose and financial management system name in the external knowledge base, and reorganize the retrieved information through the financial management system audit big model; third instruction: generate audit content, suggestions and conclusions based on the financial management system name and the financial management system name in the external knowledge base;
[0047] S422. Integrate all generated content with the audit report dimensions and corresponding report prompt templates, and use document generation tools to generate audit reports in different formats.
[0048] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:
[0049] Beneficial effects:
[0050] 1. By performing word segmentation on the text data set of financial management system and removing special characters and punctuation marks, the data quality is guaranteed, providing a clean data set for subsequent analysis. The use of word segmentation tools specifically for the financial field ensures that financial proper nouns and terms can be correctly identified and segmented, improving the professionalism and accuracy of data processing.
[0051] 2. Use topic models (such as LDA) to analyze the clean data set to determine the key themes and dimensions of the financial management system, which helps to identify the most relevant keywords for the financial management system. The optimal number of topics is determined by calculating the perplexity of the model with different numbers of topics, which ensures the effectiveness and accuracy of the model.
[0052] 3. Fine-tune the pre-trained large model using the constructed financial management system instruction dialogue dataset, so that it has professional knowledge of financial management system auditing, and improves the model's contextual understanding and generation capabilities in this field. The LORA strategy is used in the fine-tuning process, so that the model can learn and adapt to new tasks more efficiently, while maintaining a small model size and reducing resource consumption.
[0053] 4. Design standardized audit report prompt templates to ensure that the content of the generated audit report is closer to the actual audit requirements, and realize the automatic generation and efficient summary of audit reports by integrating the content generated by the big model with the audit report dimensions and report prompt templates. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.
[0055] Figure 1 A flow chart of a method for generating an audit report for a financial management system provided by an embodiment of the present invention;
[0056] Figure 2 A flowchart for determining the audit dimension of a financial management system through an LDA topic model provided in an embodiment of the present invention;
[0057] Figure 3 A flow chart of constructing a financial management system instruction dialogue data set using a large model provided by an embodiment of the present invention;
[0058] Figure 4 A flowchart of obtaining a large audit model provided by an embodiment of the present invention;
[0059] Figure 5 A flow chart of constructing audit report dimensions and corresponding report prompt templates provided in an embodiment of the present invention;
[0060] Figure 6 A flowchart for generating a corresponding audit report is provided for an embodiment of the present invention. DETAILED DESCRIPTION
[0061] 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.
[0062] like Figure 1 As shown, this embodiment provides a method for generating an audit report for a financial management system, comprising the following steps:
[0063] S1. Obtain a financial management system text dataset and perform preprocessing to obtain a preprocessed financial management system text dataset;
[0064] S2. Based on the preprocessed financial management system text dataset, determine the audit dimension of the financial management system through the LDA topic model, and use the big model to build the financial management system instruction dialogue dataset;
[0065] S3. Fine-tune the big model using the financial management system instruction dialogue dataset to obtain the audit big model;
[0066] S4. Combined with the audit dimension of financial management system, construct the audit report dimension and the corresponding report prompt template, input the audit documents to be processed into the audit big model, and generate the corresponding audit report.
[0067] This method can quickly and accurately organize and review financial management systems, significantly improve the efficiency and quality of generating audit reports on financial management systems, and provide reasonable suggestions and conclusions in the reports, thus achieving standardization and professionalization of audit reports.
[0068] The above steps and related technical features are further described in detail below:
[0069] In the present embodiment S1, the preprocessing includes: removal of preset characters and punctuation marks, word segmentation, removal of stop words, stem extraction and word form restoration, entity recognition and normalization, semantic enhancement and anomaly detection.
[0070] Specifically, regular expressions are used to remove special characters and punctuation marks in the text, such as <>, &, etc., to reduce noise interference. A word segmentation tool suitable for the field of financial management systems is used, such as NLTK (Natural Language Toolkit) or jieba分词 (a word segmentation tool specifically for Chinese), to ensure that financial proper nouns and terms can be correctly identified and segmented. Word segmentation is the process of splitting continuous text into meaningful lexical units, and this step is crucial for subsequent topic modeling because it directly affects the input quality of the topic model.
[0071] Remove stop words in the text, such as commonly used but less meaningful words like "的", "和", "是", etc., to facilitate subsequent topic modeling. And apply stemming or lemmatization techniques to convert words into their basic forms or roots, which helps reduce the diversity brought by synonyms and improve the model's performance.
[0072] Identify financial entities in the text, such as financial institution names, regulation names, etc., through named entity recognition (NER) technology, and normalize them to ensure that the same entity is represented consistently in different documents. Entity recognition is particularly important for financial texts because it can help identify key financial concepts and entities, while normalization helps ensure data consistency. And perform semantic analysis on the text, adding synonyms or related terms with a financial professional background to enhance the information density of the text. Semantic enhancement can help capture the deep meaning in the text, especially when dealing with professional terms in the financial field, which is crucial for improving the accuracy and relevance of the topic model.
[0073] Finally, use statistical methods to identify and process outliers in the text, such as content or formats that significantly do not conform to the characteristics of financial texts. Outlier detection helps filter out irrelevant or incorrect information and ensure the quality of the dataset. This step can ensure the high quality and consistency of the dataset, thus providing a reliable basis for subsequent topic modeling and large model fine-tuning.
[0074] As Figure 2 shown, in this embodiment S2, the audit dimensions of the financial management system are determined through the LDA topic model, including:
[0075] S211. Use the LDA topic model to process the preprocessed financial management system text dataset to obtain relevant topic descriptions and topic distribution situations;
[0076] S212. Based on the relevant topic descriptions and topic distribution, high-frequency words and keywords under each topic are analyzed, each topic is named, and topics related to the financial management system are identified. For example, the topic words generated by the LDA topic model, such as keywords such as "adaptability", "update", "frequency", and "cycle", are summarized as the audit dimension of untimely revision of the system, and keywords such as "regulatory compliance", "internal standards", and "industry standards" are summarized as non-compliant system content, which are ultimately used to evaluate and improve the financial management system.
[0077] The S211 includes:
[0078] A topic was randomly assigned to each word in each financial management system document;
[0079] The topic assignment is iteratively optimized through the Gibbs sampling algorithm. The probability of each word in the document belonging to each topic is recalculated, and the topic assignment of each word is continuously adjusted.
[0080] Until the convergence condition is reached, combined with the optimal number of topics, the word distribution of each topic and the topic distribution of each financial management system document are output.
[0081] In the LDA topic model, the convergence condition of the Gibbs sampling algorithm usually refers to the stabilization of the model state, that is, the topic allocation no longer changes significantly or reaches the preset maximum number of iterations. Specifically, the basis for judging convergence includes: Number of iterations: Set a fixed number of iterations and stop iterating when this number is reached. Topic allocation stability: Observe the changing trend of topic allocation. If the topic allocation does not change significantly in consecutive rounds of iterations, it is considered that convergence has been achieved. KL divergence: Calculate the KL divergence (Kullback-Leibler Divergence) between two rounds of iterations. When the KL divergence is less than a certain threshold, the model is considered to have converged. Topic distribution similarity: Compare the topic distribution of the previous round and the next round. If the difference between the two is small, the model is considered to have converged.
[0082] Once one of the conditions is met, the iteration stops and the word distribution of each topic and the topic distribution of each financial management system document are output.
[0083] The optimal number of topics is determined by calculating the perplexity of the model under different numbers of topics, including:
[0084] The preprocessed financial management system text dataset is divided into training set and test set, and the range of the number of topics is determined; the prior knowledge of the field is used to set a reasonable range of the number of topics, which is estimated according to the classification and professional field of the financial management system. And combined with dynamic range adjustment: a wider range of the number of topics is set, and then the range is further narrowed according to the results of preliminary experiments to find a more accurate optimal number of topics.
[0085] Use the training set data to train LDA models with different numbers of topics, and combine the test set data to calculate the perplexity of models with different numbers of topics:
[0086]
[0087]
[0088] Where p(w) is the probability of word w appearing in the test set, p(z|d) is the probability of topic z appearing given document d, p(w|z) is the probability of word w appearing given topic z, M is the number of documents in the test set data, and N is the probability of word w appearing given document d. d is the number of words in document d;
[0089] Based on the perplexity calculation results of models with different number of topics, the number of topics with the smallest perplexity is selected as the optimal number of topics.
[0090] In step S211 of this embodiment, different numbers of topics (e.g., 3, 5, 10, 15) are pre-set in the financial management system text training set to train the topic model. The trained topic models with different numbers of topics are then verified on the financial management system text test set, and the perplexity index is used to determine the exact number of topics.
[0091] p(w|z) is the probability of word w under the condition of determining topic z. The topic model generation is the number of specified words corresponding to each topic. The topic can be further determined based on the words, that is, the dimension of the audit. For example, the topic model with 5 topics has the lowest perplexity. The number of topics is determined to be 5, and the dimensions of the audit are summarized to be 5.
[0092] It can extract key information and topic structures from a large amount of financial management system text data, ensuring the quality and accuracy of the data and revealing the underlying structure and patterns behind the data. The use of topic models can effectively improve the pertinence and practicality of audit reports by deeply analyzing and understanding each audit dimension.
[0093] like Figure 3 As shown in the figure, a large model is used to construct a dialogue dataset of financial management system instructions, including:
[0094] S221. Use the big model to process the preprocessed financial management system text data set to generate corresponding questions and answers; the questions and answers cover the main content and details of the text; for example, to answer a certain financial regulatory system regulation, the generated question is "What is the main content of the regulation?" and the corresponding detailed answer.
[0095] S222. Check questions and answers for accuracy, consistency and compliance with industry regulations through an automatic verification mechanism;
[0096] S223. Based on the inspection results, the questions and answers are corrected and optimized;
[0097] S224. Arrange the corrected and optimized questions and answers into a unified data format to obtain a financial management system instruction dialogue data set.
[0098] The large model in this embodiment is a pre-trained model, which is a deep learning model that has been pre-trained on a large amount of unlabeled text data, such as BERT, RoBERTa, GPT series, etc. These models capture rich language structure and semantic information through self-supervised learning tasks (such as masked language model, cloze, etc.).
[0099] It uses the Transformer architecture with multiple layers of encoders or decoders, which can process long sequences of text and capture complex contextual dependencies. It is trained through tasks such as the Masked Language Model (MLM) and the Causal Language Model (CLM), enabling the model to predict masked words or generate continuous text.
[0100] In this step, a large model is used to automatically generate questions and answers, and a financial management system instruction dialogue dataset is constructed. Compared with traditional manual methods, this process is more efficient, accurate and consistent, and can fully cover content details and is easy to maintain and expand, thereby ensuring high data set quality and enhancing model performance, significantly improving the quality and efficiency of audit work.
[0101] like Figure 4 As shown, this embodiment S3 includes:
[0102] S31, dividing the financial management system instruction dialogue data set into a training set and a validation set;
[0103] S32, determine the large model and set hyperparameters; the hyperparameter settings include: learning rate, training rounds and batch size;
[0104] S33. Input the training set and validation set into the big model after the hyperparameters are set, and train the big model in combination with the LORA strategy to obtain the audit big model.
[0105] The LORA strategy reduces the number of required training parameters by performing low-rank decomposition on some weight matrices in the large model, thereby reducing the computational cost. This includes: performing low-rank decomposition on some weight matrices of the model to reduce the number of required training parameters; and using an adaptive learning rate strategy and incremental training method to update low-rank matrix parameters.
[0106] In the specific LORA method, the original weight matrix W is approximated by two smaller matrices U and V, that is, W≈U·V T , the matrices U and V are smaller in shape, reducing the number of required parameters. The size (rank) of the low-rank matrix is an adjustable hyperparameter, and the optimal value can be determined by experiment. During the fine-tuning phase, only the low-rank matrix is updated, while the original weight matrix W remains unchanged. In addition, the use of adaptive learning rate strategies (such as AdamW optimizer) and incremental training methods can further accelerate the training process and reduce memory consumption.
[0107] like Figure 5 As shown, in this embodiment S4, the audit report dimension and the corresponding report prompt template are constructed, including:
[0108] S411. Use a crawler program to obtain historical audit reports, and archive and classify them; the historical audit reports include historical audit reports obtained from multiple sources such as internal databases, partner organizations, and third-party data providers.
[0109] S412. Based on the archived and classified historical audit reports, use natural language processing algorithms to identify and extract basic structures;
[0110] S413. Determine the audit dimensions, combine the basic structures identified and extracted, and construct the initial report prompt template;
[0111] S414: Automated verification and feedback optimization are performed on the report prompt template to obtain a report prompt template.
[0112] In this embodiment, the audit report prompt template includes at least: audit object: {audit object}; audit time: {audit time}; summary: {summary}; audit background and purpose: {audit background and purpose}; audit content: {audit content}; conclusion: {conclusion}; suggestion: {suggestion}.
[0113] The construction of report prompt templates significantly improves the efficiency and quality of audit report generation, reduces the workload of auditors, and ensures the professionalism and consistency of audit reports.
[0114] like Figure 6 As shown, the audit documents to be processed are input into the audit model to generate the corresponding audit report, including:
[0115] S421. Control the audit big model to process the pending audit documents through instructions; the instructions include: first instruction: extract the audit object, audit time and financial management system name from the pending audit documents; second instruction: generate a query according to the audit dimension of the financial management system, and retrieve relevant summaries, audit background and purpose and financial management system name in the external knowledge base, and reorganize the retrieved information through the financial management system audit big model; third instruction: generate audit content, suggestions and conclusions based on the financial management system name and the financial management system name in the external knowledge base;
[0116] The construction of the external knowledge base in this step includes: integrating the public financial management system database and related documents, and using efficient and accurate text segmentation technology to process these materials, determining the block size to be 256 tokens, setting the overlap between each block to be around 20 characters, and dynamically adjusting the block size and the number of overlaps according to the complexity and content length of the document; applying named entity recognition strategies to annotate important terms, institution names, etc., and linking the identified entities to authoritative external knowledge sources; converting the segmented text into numerical vectors, and storing these vectors in an efficient vector database system that supports multimodal data.
[0117] S422. Integrate all generated content with the audit report dimensions and corresponding report prompt templates, and use document generation tools to generate audit reports in different formats.
[0118] In the specific steps of generating the corresponding audit report, it processes the audited documents through the instruction-driven audit big model, which can automatically extract key elements and retrieve relevant information in the carefully constructed external knowledge base based on the audit dimension of the financial management system, and then generate detailed audit content, suggestions and conclusions; the external knowledge base uses advanced technologies such as efficient text segmentation, named entity recognition and vectorized storage to ensure the accuracy and speed of information retrieval; finally, all generated content and audit report templates are integrated, and document generation tools are used to output a variety of audit report formats, which greatly improves the automation level and accuracy of audit work, while shortening the audit cycle and reducing the risk of human errors.
[0119] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0120] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for generating an audit report for a financial management system, characterized in that: The following steps are involved: S1. Obtain a financial management system text dataset and perform preprocessing to obtain a preprocessed financial management system text dataset; S2. Based on the preprocessed financial management system text dataset, determine the audit dimension of the financial management system through the LDA topic model, and use the big model to build the financial management system instruction dialogue dataset; S3. Fine-tune the big model using the financial management system instruction dialogue dataset to obtain the audit big model; S4. Combined with the audit dimension of financial management system, construct the audit report dimension and the corresponding report prompt template, input the audit documents to be processed into the audit big model, and generate the corresponding audit report.
2. A method for generating an audit report for a financial management system according to claim 1, characterized in that: In S1, the preprocessing includes: removal of preset characters and punctuation, word segmentation, removal of stop words, stem extraction and word form restoration, entity recognition and normalization, semantic enhancement and anomaly detection.
3. A method for generating an audit report for a financial management system according to claim 1, characterized in that: In S2, the audit dimensions of the financial management system are determined through the LDA topic model, including: S211. Use the LDA topic model to process the preprocessed financial management system text data set to obtain relevant topic descriptions and topic distribution; S212. Based on the relevant topic descriptions and topic distribution, high-frequency vocabulary and keyword analysis is performed under each topic, each topic is named, and topics related to the financial management system are identified.
4. A method for generating an audit report for a financial management system according to claim 3, characterized in that: The S211 includes: A topic was randomly assigned to each word in each financial management system document; The topic assignment is iteratively optimized through the Gibbs sampling algorithm. The probability of each word in the document belonging to each topic is recalculated, and the topic assignment of each word is continuously adjusted. Until the convergence condition is reached, combined with the optimal number of topics, the word distribution of each topic and the topic distribution of each financial management system document are output.
5. A method for generating an audit report for a financial management system according to claim 4, characterized in that: The optimal number of topics is determined by calculating the perplexity of the model under different numbers of topics, including: The preprocessed financial management system text dataset is divided into training set and test set, and the range of the number of topics is determined; Use the training set data to train LDA models with different numbers of topics, and combine the test set data to calculate the perplexity of models with different numbers of topics: Where p(w) is the probability of word w appearing in the test set, p(z|d) is the probability of topic z appearing given document d, p(w|z) is the probability of word w appearing given topic z, M is the number of documents in the test set data, and N is the probability of word w appearing given document d. d is the number of words in document d; Based on the perplexity calculation results of models with different number of topics, the number of topics with the smallest perplexity is selected as the optimal number of topics.
6. A method for generating an audit report for a financial management system according to claim 1, characterized in that: In S2, a large model is used to construct a dialogue dataset of financial management system instructions, including: S221. Process the preprocessed financial management system text data set using the big model to generate corresponding questions and answers; the questions and answers cover the main content and details of the text; S222. Check questions and answers for accuracy, consistency and compliance with industry regulations through an automatic verification mechanism; S223. Based on the inspection results, the questions and answers are corrected and optimized; S224. Arrange the corrected and optimized questions and answers into a unified data format to obtain a financial management system instruction dialogue data set.
7. The method for generating an audit report for a financial management system according to claim 1, characterized in that: The S3 includes: S31, dividing the financial management system instruction dialogue data set into a training set and a validation set; S32, determine the large model and set hyperparameters; the hyperparameter settings include: learning rate, training rounds and batch size; S33. Input the training set and validation set into the big model after the hyperparameters are set, and train the big model in combination with the LORA strategy to obtain the audit big model.
8. A method for generating an audit report for a financial management system according to claim 7, characterized in that: The LORA strategy includes low-rank decomposition of some weight matrices of the model to reduce the number of required training parameters; And use adaptive learning rate strategy and incremental training method to update low-rank matrix parameters.
9. The method for generating an audit report for a financial management system according to claim 1, characterized in that: In S4, the audit report dimensions and the corresponding report prompt template are constructed, including: S411. Use crawler programs to obtain historical audit reports, archive them, and classify them; S412. Based on the archived and classified historical audit reports, use natural language processing algorithms to identify and extract basic structures; S413. Determine the audit dimensions, combine the basic structures identified and extracted, and construct the initial report prompt template; S414: Automated verification and feedback optimization are performed on the report prompt template to obtain a report prompt template.
10. The method for generating an audit report for a financial management system according to claim 1, characterized in that: In S4, the audit document to be processed is input into the audit model to generate a corresponding audit report, including: S421. Control the audit big model to process the pending audit documents through instructions; the instructions include: first instruction: extract the audit object, audit time and financial management system name from the pending audit documents; second instruction: generate a query according to the audit dimension of the financial management system, and retrieve relevant summaries, audit background and purpose and financial management system name in the external knowledge base, and reorganize the retrieved information through the financial management system audit big model; third instruction: generate audit content, suggestions and conclusions based on the financial management system name and the financial management system name in the external knowledge base; S422. Integrate all generated content with the audit report dimensions and corresponding report prompt templates, and use document generation tools to generate audit reports in different formats.
Citation Information
Cited By
Knowledge-driven underground building design scheme intelligent examination method, system and equipment
CN120216671A
