Auditing question and answer method and system

Abstract

The invention discloses an audit question and answer method and system. The method comprises the following steps: in response to question content of a user, carrying out vector search in multi-source data based on cosine similarity to obtain an information block which is most matched with the question content; and based on the most matched information block, generating an answer for answering the question content. The technical problem that audit query answers are inaccurate is solved.

Description

Technical Field

[0001] The present invention relates to a depth field, and in particular, to an audit question-and-answer method and system. Background Art

[0002] As the complexity of auditing business increases, efficient processing of massive multi-source data and accurate acquisition of information during the audit process have become the focus of industry attention. In the existing technology, auditors usually rely on manual retrieval or keyword matching methods to extract relevant information from audit databases or documents. However, these methods have problems with low information extraction efficiency and inaccurate matching results when faced with a mixture of structured and unstructured data, which can easily lead to audit misjudgments or omissions of key data. In addition, some systems based on traditional retrieval algorithms cannot fully utilize the semantic correlation between data, resulting in low answer quality and difficulty in meeting the actual needs of complex audit scenarios.

[0003] In response to the above problems, there is an urgent need for a method that can efficiently and accurately retrieve relevant information blocks in multi-source data based on the content of user questions, so as to improve the answer quality of the audit question and answer system, solve the technical problem of inaccurate audit data acquisition, and thus provide more intelligent support for audit work.

[0004] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0005] The embodiment of the present invention provides an audit question and answer method and system to at least solve the technical problem of inaccurate audit query answers.

[0006] According to one aspect of an embodiment of the present invention, an audit question and answer method is provided, comprising: in response to a user's question content, performing a vector search in multi-source data based on cosine similarity to obtain an information block that best matches the question content; based on the best matching information block, generating an answer for answering the question content.

[0007] According to another aspect of an embodiment of the present invention, an audit question and answer system is also provided, including: a retrieval module, configured to respond to the user's question content, perform a vector search in multi-source data based on cosine similarity, and obtain an information block that best matches the question content; a generation module, configured to generate an answer for answering the question content based on the best matching information block.

[0008] In the embodiment of the present invention, in response to the user's question content, a vector search is performed in multi-source data based on cosine similarity to obtain the information block that best matches the question content; based on the best matching information block, an answer for answering the question content is generated. Through the above technology, the technical problem of inaccurate audit query answers is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0010] Figure 1 is a flow chart of an audit question and answer method according to an embodiment of the present invention;

[0011] Figure 2 is an architectural diagram of an audit question-and-answer system according to an embodiment of the present invention;

[0012] Figure 3 is a flowchart of another optional audit question and answer method according to an embodiment of the present invention;

[0013] Figure 4 is a flowchart of another optional audit question and answer method according to an embodiment of the present invention;

[0014] Figure 5 is a flow chart of another audit question-and-answer system according to an embodiment of the present invention;

[0015] Figure 6 A schematic diagram of the structure of an electronic device suitable for implementing the embodiments of the present disclosure is shown. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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 should fall within the scope of protection of the present invention.

[0017] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0018] According to an embodiment of the present invention, a method embodiment of an audit question and answer method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0019] Figure 1 is an audit question and answer method according to an embodiment of the present invention, such as Figure 1 As shown, the method comprises the following steps:

[0020] Step S102 : in response to the user's question content, a vector search is performed in the multi-source data based on cosine similarity to obtain an information block that best matches the question content.

[0021] Before performing vector search in multi-source data based on cosine similarity, the audit file uploaded by the user is obtained and pre-processed. For example, the audit file is formatted and document content is extracted from the audit file after format conversion; the extracted document content is processed into blocks to obtain multiple text blocks; each of the multiple text blocks is converted into a vector and stored in a local vector storage.

[0022] Afterwards, the similarity between the vector corresponding to the question content and the vector in the multi-source data is measured by calculating the angle between the vector corresponding to the question content and the vector in the multi-source data; when the similarity is greater than a similarity threshold, the information block corresponding to the corresponding vector in the multi-source data is determined to be the most matching information block.

[0023] Step S104: generating an answer to the question based on the best matching information block.

[0024] Add context to the best matching information block to generate a prompt; based on the prompt, use the relevant knowledge text to generate the answer. For example, the system performs context expansion on the best matching information block, retrieves context information related to it, such as adjacent text blocks or other text data on the same topic, and merges it with the best matching information block to generate a more complete prompt. The prompt will combine the user's question content and the expanded information block, for example: "The user asked 'XXX', the relevant information is 'XXX', please generate an answer based on the above content." The system then inputs the prompt into the pre-trained language model to generate the final answer content. In this process, the system will optimize the generated answer in combination with the relevant knowledge text to ensure that the content of the answer is not only highly relevant to the user's question, but also clear in language expression and semantically coherent.

[0025] The present application embodiment provides an audit question-answering system. Figure 2 As shown, it includes: a vector knowledge base 22, a langchain module 24, an LLM model 26 and a UI interface 28.

[0026] The vector knowledge base 22 is the core data storage component in the audit question-answering system, which is used to efficiently store and retrieve knowledge related to audit tasks. The knowledge base uses a vectorized representation method to convert documents, reports, regulations, standards, and historical data in the audit field into high-dimensional vectors and store them in the database. The vector representation can effectively capture the semantics and context of information, so that the system can understand the query intent and text semantics when performing information retrieval, not only based on keyword matching.

[0027] The LangChain module 24 (also called the langChain framework) is the core framework of the RAG application. This embodiment realizes efficient linkage between large language models, audit knowledge bases, the Internet, and external APIs based on the LangChain framework. LangChain constitutes the core of the architecture, covering six key components: indexes, memory, prompts, chains, models, and agents, which jointly support the application of large language models. With the help of LangChain, dynamic interaction with audit knowledge bases, LLMs, the Internet, and external resources can be achieved. In the embodiment of the present application, LangChain accurately analyzes complex audit queries, extracts relevant information, and generates human-understandable responses by integrating organized retrieval and generation steps. In addition, by integrating semantic search technology and knowledge base, the accuracy of retrieval results and the relevance of answers are greatly improved, solving the problems of inaccurate information, slow updates, and opaque answers that may be encountered in practical applications of large language models.

[0028] The LLM model 26 is used for long-form text capture, multi-language processing, multi-turn dialogue management, text generation, common sense reasoning, and information extraction. The LLM model can deeply and comprehensively understand complex audit issues and provide natural and fluent responses similar to human dialogue.

[0029] The embodiment of the present application combines the LangChain framework with the LLMs model, and uses precise prompts designed specifically for LLMs to greatly enhance the performance of the model when processing audit tasks. This combination not only guides LLMs to make relevant and accurate responses to audit tasks such as risk assessment, data analysis, or compliance checks, but also dynamically adjusts the information retrieval and generation process to meet the needs of specific fields. By building an efficient and powerful RAG question-and-answer application, the LangChain framework ensures the accuracy of information retrieval and the consistency of generated answers, while the addition of LLMs significantly improves the system's ability to process natural language.

[0030] The UI interface 28 is an interface for auditors to interact with the system. Through the UI interface, users can easily input audit-related questions and obtain fast and accurate responses from the system. The interface provides multiple interaction methods, including natural language input, voice input and text query, to ensure that it adapts to the needs and operating habits of different users.

[0031] This application uses advanced natural language processing technology to improve the system's understanding ability, uses machine learning and artificial intelligence technology to optimize the data processing process, and designs an intuitive and easy-to-use UI interface to enhance the user experience. Through the above architecture, the audit question-and-answer system can provide fast and accurate answers while ensuring the accuracy and transparency of information to meet the needs of audit professionals.

[0032] Figure 3 Another audit question-and-answer method according to an embodiment of the present invention is applied in Figure 2 On the system shown, the method comprises the following steps:

[0033] Step S302: document upload processing.

[0034] In the audit question and answer system, document upload processing not only enriches the system's knowledge base, but also ensures that the system can respond to the latest audit information needs in a timely manner. Through the document processing process, the system can convert the files uploaded by users into a machine-readable format, thereby effectively analyzing and integrating the latest data. This is essential for maintaining a dynamically updated audit environment that can quickly respond to real-time changes.

[0035] The document upload process specifically includes:

[0036] 1) Format conversion and content extraction. The system uses advanced preprocessing mechanisms to convert documents of various formats into a unified format and extract key information from them. This ensures that valuable information in the documents can be effectively identified and utilized by the system.

[0037] 2) Block processing of document content. In order to improve processing efficiency and retrieval accuracy, the system divides the huge text data into smaller and more manageable data blocks. This not only lays the foundation for in-depth analysis, but also significantly improves the efficiency of retrieval tasks.

[0038] 3) Vectorization and indexing of text blocks. Using embedding technology, the system converts text blocks into high-dimensional vector representations that capture the deep semantic features of the text. Through FAISS technology, these vectors are effectively stored in a database optimized for similarity retrieval, so that queries submitted by users can quickly get accurate answers.

[0039] The entire document upload process not only optimizes the accuracy of automated answers, but also greatly shortens the time it takes for users to obtain the information they need.

[0040] Step S304: multi-source information retrieval and matching.

[0041] In semantic search and data analysis, cosine similarity is used to evaluate the similarity between text vectors. Cosine similarity measures the degree of similarity between two vectors by calculating the angle between them. This method is very effective for comparing vectors in high-dimensional space because it only focuses on the direction of the vector rather than the size, making it highly reliable and accurate when processing text data. The following formula describes the process of cosine similarity calculation, where A·B is the dot product of two vectors, ||A|| and ||B|| are the Euclidean norms of the vectors (i.e., the lengths of the vectors).

[0042]

[0043] Where n is the number of vectors.

[0044] To further improve the efficiency and accuracy of retrieval, the Q&A application integrates multiple information sources from the Internet and professional audit knowledge bases. By performing similarity searches in these extensive data sources, the Q&A application can provide users with more comprehensive and in-depth answers. This cross-source information retrieval strategy not only broadens the scope of information that users can obtain, but also improves the comprehensiveness and depth of answers.

[0045] In the specific implementation process, the Q&A application uses vector database technologies such as Pinecone and Weaviate to build and optimize the search function of the knowledge base. This enables the system to quickly locate the most matching answer in a huge data set based on deep semantic understanding. In addition, in order to enhance the user experience, the Q&A application also integrates the Internet search mode to expand the search scope through external search engines, thereby providing users with richer and more diverse information sources.

[0046] In this application, the system not only displays information related to the user's query, but also improves the user experience and meets the user's personalized needs through a series of processing and optimization steps. From structured presentation of search results to providing direct citations and other diverse information display methods, it greatly facilitates users to obtain and understand the required information. The question-and-answer application uses semantic matching technology to quickly lock the most relevant information in the vector database, relying on deep learning algorithms to understand and match user query intent. The search results are personalized through text summaries, key information highlighting, etc., to improve information readability and enable users to quickly capture key content.

[0047] Step S306, output processing.

[0048] In the output processing link of the audit question-and-answer system, the system generates text answers that are closely related to user queries and are natural and fluent through the LLMs chain implemented by LangChain technology, emphasizing meeting user information needs and providing personalized interactive experience. The system uses a specific language model and configures prompt templates to guide the model to generate relevant text and control the subject scope of the generated content. Construct an LLMs chain instance to integrate the language model and prompt template to generate text that meets the requirements.

[0049] In addition, the audit question-answering system defines JSON architecture and chat prompt templates based on the structured output of chat history to help users understand the communication context. The structured output chain of Internet search results enhances the system's ability to obtain external information, defines the output format and initialization model, and uses language models and prompt templates to generate structured output.

[0050] In summary, the output processing link optimizes the quality of language generation and improves the efficiency and experience of users in obtaining information through detailed configuration and advanced technology. Through personalized output and structured information display, the system provides a rich interactive experience while meeting user needs.

[0051] This application provides an audit question and answer method based on RAG technology, which solves the problem of audit professionals extracting accurate information when faced with huge amounts of data. RAG technology combines the capabilities of LLMs with efficient information retrieval, significantly improves the accuracy of answers, and can quickly integrate expertise in specific fields. Through document upload processing, information retrieval, and output processing, the audit question and answer application can provide efficient and accurate answers to complex audit questions and optimize the interaction process between users and the system. This not only improves the efficiency and quality of audit work, but also provides a new path for solving future audit problems. The RAG technology of this application has good results in handling complex audit questions and improving the accuracy of answers, providing rich possibilities for the further development of the audit industry.

[0052] This application embodiment provides another audit question and answer method, such as Figure 4 As shown, the method comprises the following steps:

[0053] Step S402: receiving and processing unstructured data.

[0054] Convert unstructured data such as audit-related documents uploaded by users, Internet search content, and audit knowledge base into structured data that can be understood by machine learning models.

[0055] First, data format conversion and content extraction are performed. The format parsing module is used to convert the document into a unified text format. Subsequently, key information in the document, including titles, chapter contents, digital tables, and annotations, is extracted through natural language processing technology. For example, optical character recognition (OCR) is used to parse scanned files or image format documents to ensure the comprehensiveness and accuracy of content extraction. Then, redundant format characters (such as page numbers and footnotes) are removed and the text content is regularized to improve the efficiency of subsequent processing.

[0056] Next, the text is segmented. The long text is divided into multiple small blocks. Specifically, through syntactic analysis and semantic analysis, the semantics of each text block is ensured to be complete to avoid information omission. According to the input length limit of the model, the length of the text block is controlled within the processable range of the model. This processing method plays an important role in natural language processing (NLP) tasks and question-answering methods in the audit field. In the audit field, complex audit texts can be converted into structured small units through text segmentation, laying the foundation for subsequent analysis. It provides refined input for large language models, which is convenient for extracting key information or performing context matching. It is a key step in document processing and natural language understanding, especially in the scenario of audit question and answer, it can provide high-quality input for the model, improve analysis efficiency and answer accuracy.

[0057] Finally, text vectorization and index construction are performed. Each text block is converted into a high-dimensional vector representation using embedding technology. These vectors capture the deep semantic features of the text block, which facilitates subsequent retrieval and matching. The specific implementation includes: using the embedding interface to generate a vector representation of the text block, and then using fast similarity search technology to store the generated vector in a vector database to establish an efficient retrieval index. The construction of the index significantly improves the system's response speed and accuracy to user queries.

[0058] Step S404: information retrieval and query matching.

[0059] Convert text, images or other data into numerical vector representations, and calculate the similarity between different vectors to achieve fast and accurate information retrieval. In the audit question-answering scenario driven by a large language model, vector search can efficiently find content that is semantically related to user input (such as questions, queries) in massive audit data by combining semantic embedding technology.

[0060] Specifically, embedding generation is performed first. The key to vector search is to convert text data into vectors. This is usually done by a large language model or a specific embedding model. These models generate high-dimensional embedding vectors based on context and semantic information.

[0061] Next, similarity calculation is performed. In some embodiments, the similarity between the user query and each vector in the database can be calculated using methods such as cosine similarity or Euclidean distance, so as to find the best matching result. In this way, the semantic relevance between texts can be used to overcome the shortcomings of traditional keyword matching. In some other embodiments, the model can also be used for similarity matching. For example, the initial embedding vector group of the user query and the auxiliary embedding vector group of each vector in the database can be extracted respectively. For the user query, the model generates its feature representation, that is, the initial embedding vector group, which is used to capture the characteristics of the user query in the multidimensional semantic space. Similarly, for each vector in the database, the model generates an auxiliary embedding vector group to characterize its semantic features. The extraction process of these embedding vectors enables the model to model the intrinsic features of the text in the high-dimensional semantic space, laying the foundation for the subsequent steps. Subsequently, according to the set extension rank rule, the initial embedding vector group and the auxiliary embedding vector group are respectively extended in multiple dimensions. Next, a multi-channel convolution kernel is used to couple and fuse the features of the embedding vector group to obtain the coupled initial embedding group and auxiliary embedding group. Next, the coupling results are processed by applying a nonlinear activation function to further enhance the expressive power of semantic features and generate the activated initial embedding group and auxiliary embedding group. Finally, the activated embedding group is normalized to ensure the consistency of data scale, and finally a multidimensional feature matrix is ​​constructed. Next, high-order matrix decomposition operations are performed on the initial feature matrix and the auxiliary feature matrix respectively. Specifically, these matrices are unfolded along the extended dimension and decomposed into multiple sub-matrices. Singular value decomposition is performed on each sub-matrix to extract the core subspace representation and factor matrix. The core subspace representation gathers the key information of the matrix, while the factor matrix reveals its local structural characteristics. Subsequently, the main component differences between the initial feature matrix and the auxiliary feature matrix are calculated. By analyzing the similarity of the core subspace representation and the difference of the main components in the factor matrix, the subtle differences between the two sets of data can be accurately captured.

[0062] Step S406, generate answers and optimize the answers.

[0063] After completing the information retrieval, the final answer is generated based on the query context and matching results. First, the model is used to generate preliminary answers, and the style of the answers is optimized according to the user's language preference. The system uses prompt engineering in the LangChain framework to ensure that the generated answers are consistent with the language style of the audit field. In order to improve the readability of the answers, the system will summarize and refine the preliminary generated answers, highlight key information and eliminate redundant content. All data contained in the answers are accompanied by reference sources (such as knowledge bases, the Internet, documents), and embedded in the answers in the form of hyperlinks, making it convenient for users to trace the source of the information. This application converts unstructured data into a structured format, and then uses vector similarity matching technology to quickly locate the information most relevant to the user's query in massive data. In this way, not only the accurate extraction and processing of information is optimized, but also the system is ensured to operate stably and efficiently, providing a solid foundation for the entire audit question and answer system.

[0064] Through the above steps, audit professionals can quickly extract the accurate information they need when faced with huge amounts of data. This not only improves the efficiency and quality of audit work, but also provides a new path for solving future audit problems. RAG technology has demonstrated its potential in handling complex audit problems and improving the accuracy of answers, providing rich possibilities for the further development of the audit industry.

[0065] This application also provides an audit question and answer system, such as Figure 5 The system shown includes: a retrieval module 52, which is configured to respond to the user's question content and perform vector search in multi-source data based on cosine similarity to obtain the information block that best matches the question content; and a generation module 54, which is configured to generate an answer to the question content based on the best matching information block.

[0066] It should be noted that the audit question and answer system provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the audit question and answer system provided in the above embodiment belongs to the same concept as the audit question and answer method embodiment. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0067] Figure 6 FIG. 1 shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiment of the present disclosure. It should be noted that: Figure 6 The electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0068] like Figure 6As shown, the electronic device includes a central processing unit (CPU) 1001, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage part 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for system operation are also stored. The CPU 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.

[0069] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed.

[0070] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An audit question-and-answer method, characterized in that: include: In response to the user's question content, a vector search is performed in the multi-source data based on cosine similarity to obtain an information block that best matches the question content; Based on the best matching information block, an answer for answering the question content is generated.

2. The method according to claim 1, characterized in that Before performing vector search in multi-source data based on cosine similarity, the method further includes: obtaining an audit file uploaded by a user, and preprocessing the audit file.

3. The method according to claim 2, characterized in that Preprocessing the audit file includes: Converting the audit file into a new format and extracting the document content from the audit file after the conversion. The extracted document content is processed into blocks to obtain multiple text blocks; Each text block in the plurality of text blocks is converted into a vector and stored in a local vector storage.

4. The method according to claim 1, characterized in that: Based on the cosine similarity, a vector search is performed in the local vector storage to obtain the information block that best matches the question content, including: Measuring the similarity between the vector corresponding to the question content and the vector in the multi-source data by calculating the angle between the vector corresponding to the question content and the vector in the multi-source data; In the case where the similarity is greater than a similarity threshold, the information block corresponding to the corresponding vector in the multi-source data is determined to be the best matching information block.

5. The method according to claim 4, characterized in that Based on the best matching information block, generating answer information for answering the question content includes: Adding context to the best matching information block to generate a prompt; Based on the prompt, the answer is generated using relevant knowledge text.

6. The method according to claim 4, characterized in that The multi-source data includes at least one of the following: an embedded knowledge base, a local vector storage, and the Internet.

7. An audit question-answering system, characterized in that: include: A retrieval module is configured to respond to the user's question content and perform vector search in multi-source data based on cosine similarity to obtain an information block that best matches the question content; The generating module is configured to generate an answer for answering the question content based on the best matching information block.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored program, wherein when the program is executed, the device where the computer-readable storage medium is located is controlled to execute the method according to any one of claims 1 to 6.

9. A computer device, characterized in that: include: Memory and processor, The memory stores a computer program; The processor is used to execute the computer program stored in the memory, and when the computer program is run, the processor is enabled to execute the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images