Finance and tax report interpretation method and device, computer device and storage medium

Abstract

This invention discloses a method, apparatus, computer equipment, and storage medium for interpreting financial and tax reports. The method includes: acquiring questions, reports, and identity information related to the financial and tax report; converting and processing the format of the financial and tax risk report data in the report, and parsing the open-source report of the report to obtain a parsing result; optimizing the wording of the question, identifying the intent, and determining the category; when the category is a question related to the report, searching for relevant information in a knowledge base to obtain retrieval results; integrating the question, the identity information, the retrieval results, and the parsing results into prompt words, and generating an answer using a large language model to obtain an interpretation result; and outputting the interpretation result. By implementing the method of this invention, the quality of financial and tax report interpretation can be improved, maintenance processes simplified, costs controlled, and ultimately, the accuracy and professionalism of the interpretation improved while reducing reading difficulty.

Description

Technical Field

[0002] This invention relates to computers, and more specifically to methods, apparatus, computer equipment, and storage media for interpreting financial and tax reports. Background Technology

[0004] The significance of interpreting financial and tax reports lies in providing a scientific basis for corporate decision-making through accurate analysis and understanding of financial and tax data, while ensuring tax compliance and optimizing financial management efficiency. In short, it aims to enhance a company's insight into its financial situation, support strategic planning, and guarantee compliant operations.

[0005] Existing interpretation methods include tax and finance rule engine systems, traditional natural language processing question-answering tools, large-scale language models for vertical domains, and large-scale models combined with knowledge base enhancements. These methods each have their own characteristics in practical applications, but they also face different challenges and limitations.

[0006] First, tax and finance rule engine systems rely on pre-defined rules and structured tax and finance knowledge databases, and are widely used in corporate tax compliance reviews. Representative products include SAP Tax Module and Oracle Hyperion. While these systems excel in standardized data processing, their heavy reliance on manual rule configuration and knowledge base maintenance leads to insufficient flexibility and limited ability to mine and interpret unstructured information, resulting in a high user barrier. Second, traditional NLP question-answering tools, such as retrieval-based or basic language model-based question-answering robots, are mainly used for simple policy query tasks. They often cannot integrate knowledge from multiple databases, and most do not support complex multi-turn dialogue processes. Third, large-scale language models in vertical domains fine-tune or pre-train basic models using extensive tax and finance expertise or internal materials to provide a more flexible interactive experience. While this approach can handle longer texts, it faces the challenge of maintaining the timeliness of the knowledge base due to the need for continuous data updates and technical adjustments, while also requiring high data quality, hardware resources, and professional team support. Finally, general-purpose large-scale models, combined with knowledge base retrieval mechanisms, employ retrieval-enhanced generative techniques to access specific knowledge in professional domains as a supplement to answering questions. However, in specific fields such as finance and taxation, the effectiveness of this method may be unsatisfactory and requires additional customization; in addition, when multiple search results are input, it may further limit the model's context processing capabilities, making it difficult to effectively coordinate information from different sources, thus producing contradictory or repetitive answers.

[0007] Therefore, it is necessary to design a new method to improve the quality of financial and tax report interpretation, simplify maintenance processes, control costs, and ultimately improve the accuracy and professionalism of interpretation while reducing the difficulty of reading. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, apparatus, computer equipment and storage medium for interpreting financial and tax reports.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: a method for interpreting financial and tax reports, including:

[0011] Obtain questions, reports, and identity information regarding financial and tax reporting;

[0012] The financial and tax risk report data in the report is converted and processed in a different format, and the open-source report of the report is parsed to obtain the parsing results;

[0013] Optimize the wording of the problem, identify the intent, and determine the category;

[0014] When the category is a question related to the report, relevant information is searched in the knowledge base to obtain search results;

[0015] The question, the identity information, the search results, and the parsing results are integrated into prompt words, and a large language model is used to generate the answer to obtain the interpretation result;

[0016] Output the interpretation results.

[0017] The further technical solution is as follows: The financial and tax risk report data in the report undergoes format conversion and processing, and the open-source report of the report is parsed to obtain the parsing results, including:

[0018] The financial and tax risk report data in the report is subjected to format conversion, complex fields, irrelevant fields, long tables, and further format conversion to obtain the authorized report processing result;

[0019] The open-source report is processed according to a pattern to obtain the unauthorized report processing result;

[0020] The parsing results include the results of authorized report processing and the results of unauthorized report processing.

[0021] The further technical solution is as follows: The process involves format conversion, complex field and irrelevant field handling, long table formatting, and further format conversion of the financial and tax risk report data to obtain the authorized report processing result, including:

[0022] The financial and tax risk report data in the report is converted to obtain JSON format text;

[0023] Irregularly abbreviated financial and tax indicator names or chapter titles in the JSON format text are represented by short field names to convert them into semantically clear expressions, thus obtaining the translation results;

[0024] Redundant fields that are not helpful for financial and tax risk analysis or whose information already exists in the knowledge base are removed from the translation results to obtain the elimination results;

[0025] The elimination results are then transposed by reducing the number of repeated column names to obtain the transposed result.

[0026] The transposed results are organized into a standard JSON format, and multi-level chapter headings are added to obtain the authorization report processing results.

[0027] The further technical solution is as follows: the open-source report of the report is processed according to a pattern to obtain the unauthorized report processing result, including:

[0028] The open-source report in the quick mode of the aforementioned report uses the result of an open-source PDF to Markdown converter as the unauthorized report processing result;

[0029] The open-source report of the precise pattern of the report is divided into equal parts according to length and input into a large language model in parallel for semantic segmentation and labeling to obtain preliminary processing results;

[0030] The text is further segmented based on the delimiters added according to the preliminary processing results. Then, it is input into a large language model for detailed analysis and summarization to correct the positional misalignment problem that may be caused by PDF parsing, so as to obtain the summary results of each paragraph.

[0031] The summaries of the paragraphs are merged in their original order to form the unauthorized report processing result.

[0032] The further technical solution is as follows: optimizing the description of the problem, identifying the intent, and obtaining the category includes:

[0033] The problem statement is optimized using a large language model to obtain an optimized problem.

[0034] The optimized problems are analyzed to determine whether they are related to financial and tax reporting, and the results of the analysis are categorized to obtain a class.

[0035] The further technical solution is as follows: after optimizing the description of the problem and identifying the intent to obtain the category, it also includes:

[0036] When the category is a non-report-related question, the question, the identity information, and the parsing result are integrated into a prompt word, and an answer is generated using a large language model to obtain the interpretation result, and the interpretation result is output.

[0037] The further technical solution is that the interpretation result is output in Markdown format and has a fixed ending.

[0038] The present invention also provides a financial and tax report interpretation device, comprising:

[0039] The acquisition unit is used to acquire questions, reports, and identity information related to financial and tax reports.

[0040] The parsing unit is used to convert and process the financial and tax risk report data in the report, and to parse the open-source report of the report to obtain the parsing results;

[0041] An optimization unit is used to optimize the description of the problem, identify the intent, and determine the category;

[0042] The retrieval unit is used to search for relevant information in the knowledge base when the category is a question related to the report, so as to obtain retrieval results;

[0043] The interpretation unit is used to integrate the question, the identity information, the search results, and the parsing results into prompt words, and use a large language model to generate an answer to obtain the interpretation result;

[0044] The output unit is used to output the interpretation result.

[0045] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0046] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0047] The beneficial effects of this invention compared to existing technologies are as follows: This invention obtains users' questions related to financial and tax reports, the reports themselves, and users' identity information. First, it converts and processes the data in the reports and performs precise parsing on the open-source reports to extract key information. Next, it optimizes the wording of users' questions and identifies their intentions, classifying and confirming whether they are relevant to the report content. For queries related to the reports, it searches for supplementary information in the knowledge base to enrich the accuracy and professionalism of the answers. Then, it integrates users' questions, identity information, retrieved relevant information, and report parsing results into a prompt word template, and uses a specially trained large language model to generate detailed and easy-to-understand answers. This process not only improves the quality and professionalism of financial and tax report interpretation but also simplifies the maintenance process, effectively controls costs, and ultimately achieves the goal of improving interpretation accuracy, enhancing professionalism, and significantly reducing the difficulty for non-professionals to read and understand financial and tax reports.

[0048] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram illustrating an application scenario of the financial and tax report interpretation method provided in this embodiment of the invention.

[0052] Figure 2 A flowchart illustrating the financial and tax report interpretation method provided in this embodiment of the invention. Figure 1 ;

[0053] Figure 3 A flowchart illustrating the financial and tax report interpretation method provided in this embodiment of the invention. Figure 2 ;

[0054] Figure 4 A schematic diagram of the parsing process provided in the embodiments of the present invention;

[0055] Figure 5 This is a schematic diagram of the retrieval process provided in an embodiment of the present invention;

[0056] Figure 6 A flowchart illustrating the integration of prompt words provided in an embodiment of the present invention;

[0057] Figure 7 A schematic block diagram of a financial and tax report interpretation device provided in an embodiment of the present invention;

[0058] Figure 8 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0062] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0063] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0064] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the financial and tax report interpretation method provided in an embodiment of the present invention. Figure 2This is a schematic flowchart illustrating the method for interpreting financial and tax reports provided in this embodiment of the invention. The method is applied to a server. The server interacts with the terminal, first performing format conversion and processing on the report data, including simplifying complex fields, removing redundant information, and transposing long tables to improve data clarity and usability. Authorized and unauthorized reports are processed separately to obtain parsing results. Next, a large language model is used to optimize the user's question expression, accurately identify and categorize the question intent, and search for supplementary information in the knowledge base for report-related questions. After integrating all relevant information, the large language model is used again to generate accurate and professional interpretation answers. Finally, the results are output in a concise Markdown format with a fixed ending. This not only improves the quality and professionalism of the financial and tax report interpretation but also simplifies the maintenance process and controls costs through automated processing, thereby significantly reducing reading difficulty and improving the overall accuracy and accessibility of the interpretation.

[0065] Figure 2 This is a flowchart illustrating the method for interpreting financial and tax reports provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S160.

[0066] S110. Obtain questions, reports, and identity information regarding financial and tax reporting.

[0067] In this embodiment, a user question refers to a question raised by the user.

[0068] The questions raised by users may involve the content of financial and tax reports, such as the explanation of specific indicators and risk assessments; or they may be unrelated to the content of the reports, such as inquiries about service-related information or technical support.

[0069] User questions serve as initial input to optimize wording and identify intent, determining subsequent processing paths (whether to search the knowledge base, how to parse the report, etc.).

[0070] Uploaded reports include license reports and open-source reports.

[0071] Authorization reports are typically structured message texts sourced from trusted databases. These reports have a clear hierarchical structure but may contain redundant information, requiring further processing to improve readability and analytical efficiency.

[0072] Open-source reports are typically unauthorized reports uploaded as PDF files. Due to the flexibility of the PDF format, parsing it presents challenges, requiring different processing modes (fast mode or precise mode) to ensure the accuracy and completeness of information extraction.

[0073] The report provides raw data support for subsequent report analysis and information extraction, and is one of the key bases for generating the final answer.

[0074] Customized identity information is personalized information provided by service purchasers to reflect specific service backgrounds or identifiers in responses. This enhances the professionalism and relevance of responses, especially when discussing service details and access control. It also helps protect intellectual property and maintain service brand consistency.

[0075] Through step S110, this method comprehensively acquires the essential information needed to understand and respond to user needs. This process not only lays the foundation for subsequent report processing, issue preprocessing, and knowledge base retrieval, but also ensures the accuracy, professionalism, and relevance of the output, thereby effectively reducing the difficulty for non-professionals to understand complex financial and tax risk reports and providing valuable insights and suggestions.

[0076] S120. Convert and process the format of the financial and tax risk report data in the report, and parse the open source report of the report to obtain the parsing results.

[0077] In this embodiment, the parsing result refers to the structured and easily understandable data output generated after a series of processing steps on the authorized and unauthorized reports.

[0078] In one embodiment, please refer to Figure 4 The above-mentioned step S120 may include steps S121 to S122.

[0079] S121. Perform format conversion, complex field, irrelevant field, long table, and further format conversion on the financial and tax risk report data in the report to obtain the authorized report processing result.

[0080] In this embodiment, the authorization report processing result refers to the clear and structured authorization tax risk report text after processing such as format conversion, field translation, redundant information removal, and long table transposition.

[0081] In one embodiment, step S121 described above may include steps S1211 to S1215.

[0082] S1211. Convert the financial and tax risk report data in the report to obtain JSON format text.

[0083] In this embodiment, JSON format text refers to message body text extracted from the original data and converted into a hierarchical structure to facilitate subsequent processing and analysis.

[0084] This process converts the authorized user's financial and tax risk report data from its original storage format into JSON-like structured text, providing a foundation for subsequent processing.

[0085] S1212. Represent the irregularly abbreviated financial and tax indicator names or chapter titles in the JSON format text using short field names to convert them into semantically clear expressions to obtain translation results.

[0086] In this embodiment, the translation result refers to the list of field names that have been converted from irregularly abbreviated financial and tax indicator names or chapter titles into a list that is semantically clear and concise.

[0087] The specific operations include:

[0088] Translate irregular abbreviation fields into semantically clear expressions, such as converting "fp_amt" to "invoice amount" and "inflated_income_risk_for_4y" to "risk of inflated income for the past 4 years".

[0089] Provided the field names are clear and descriptive, retain the version that minimizes the number of input tokens for the model. For example, for fields like "rate" or "risk_hint," if the length of the translated token exceeds the original English name, retain the original English field name to optimize model input efficiency.

[0090] S1213. Remove redundant fields from the translation results that are not helpful for financial and tax risk analysis or whose information already exists in the knowledge base, to obtain the removal results.

[0091] In this embodiment, the elimination result refers to the simplified report text after removing irrelevant or duplicate information from the financial and tax risk analysis, thus focusing more on the key content.

[0092] This process reduces redundant information and improves the conciseness of the report by removing irrelevant fields (such as legal basis, risk liability statements, etc.).

[0093] S1214. The removal results are transposed by reducing the number of repeated column names to obtain the transposed result.

[0094] In this embodiment, transpose means that the long table data, optimized by reducing the number of repetitions of table column names, is presented in a more compact form to display the same information.

[0095] Whether to perform a transpose operation depends on the number of rows and columns in the table. For example, a table with multiple rows of repeated column names can be transposed into a format where each row of column names corresponds to multiple rows of data, thereby reducing duplicate content in the JSON text.

[0096] Example:

[0097] Before transposition (110 characters):

[0098] "Debt Risk Situation":{

[0099] {"Year": 2021, "Debt Ratio": 0.08, "Risk Rating": "Low"},

[0100] {"Year": 2022, "Debt Ratio": 0.05, "Risk Rating": "Low"},

[0101] {"Year": 2023, "Debt Ratio": 0.50, "Risk Rating": "Medium"},

[0102] {"Year": 2024, "Debt Ratio": 3.00, "Risk Rating": "High"},

[0103] {"Year": 2025, "Debt Ratio": 0.10, "Risk Rating": "Low"}

[0104] }

[0105] After transposition (53 characters):

[0106] "Debt Risk Situation (Transfer)": {

[0107] "Year": [2021, 2022, 2023, 2024, 2025],

[0108] "Debt ratio": [0.08, 0.05, 0.50, 3.00, 0.10]

[0109] Risk Rating: ["Low", "Low", "Medium", "High", "Low"]

[0110] }

[0111] S1215. Organize the transposed result into a standard JSON format and add multi-level chapter titles to obtain the authorization report processing result.

[0112] This process includes:

[0113] Organize the transposed data into standard JSON format.

[0114] Add multi-level chapter headings in sequence (such as "I. Basic Company Information", "1.1 Shareholder Information", etc.) to reduce JSON nesting, making the report hierarchy clearer and facilitating subsequent model processing.

[0115] S122. The open-source report of the report is processed according to the pattern to obtain the unauthorized report processing result;

[0116] The parsing results include the results of authorized report processing and the results of unauthorized report processing.

[0117] In this embodiment, the unauthorized report processing result refers to the concise and standardized report text obtained by directly converting an open-source PDF report using the fast mode or conducting in-depth analysis using the precise mode.

[0118] In one embodiment, step S122 described above may include steps S1221 to S1224.

[0119] S1221. The open-source report of the quick mode of the report is processed as an unauthorized report using the result of an open-source PDF to Markdown converter.

[0120] In this embodiment, the fast mode directly uses the conversion result of the open-source PDF to Markdown converter as the final output. This method is suitable for scenarios with high efficiency requirements, but may sacrifice some accuracy.

[0121] S1222. Divide the open-source report of the precise pattern of the report into equal parts according to length, and input them in parallel into a large language model for semantic segmentation and labeling to obtain preliminary processing results.

[0122] In this embodiment, the preliminary processing result mentioned above refers to the set of text fragments after segmenting the open source report text and adding delimiters through a large language model in precise mode.

[0123] The specific operations include:

[0124] The original report text was simply divided into equal parts by length (approximately 1000 words each), and the semantics of each part may be incomplete.

[0125] Each segment is input into a general large language model in parallel, instructing the model to add semantic delimiters (e.g., ...) to each segment.<split_sign> ).

[0126] S1223. Based on the preliminary processing results, the text is further segmented using the added delimiters. Then, it is input into a large language model for detailed analysis and summarization to correct any positional errors that may be caused by PDF parsing, in order to obtain the summary results for each paragraph.

[0127] In this embodiment, the summary results of each paragraph mentioned above refer to the accurate summary of each paragraph obtained after detailed analysis and correction by re-inputting it into the large language model based on the preliminary processing results.

[0128] The specific operations include:

[0129] The segments with added delimiters are merged in their original order to obtain the original report text containing multiple delimiters.

[0130] The text is segmented according to the position of the delimiter, and then input into the general large language model in parallel. The model is instructed to analyze in depth any text with possible misaligned positions (usually caused by PDF parsing errors), and to summarize the content of each segment in Markdown format.

[0131] S1224. Combine the summary results of each paragraph in their original order to form the unauthorized report processing result.

[0132] In this embodiment, this process ensures the completeness and accuracy of the report while optimizing the text structure for easier subsequent processing.

[0133] Through the above steps, both authorized and unauthorized reports are processed into structured and easy-to-understand parsing results, providing a high-quality data foundation for subsequent intelligent interpretation.

[0134] S130. Optimize the wording of the problem, identify the intent, and determine the category.

[0135] In this embodiment, the category refers to classifying the optimized problem into two categories: "report-related" or "non-report-related".

[0136] Step S130 aims to optimize and analyze the questions raised by users to ensure that the questions are clearly stated and semantically unambiguous, and to accurately identify whether the questions are related to financial and tax reports, thereby classifying them for subsequent processing.

[0137] In one embodiment, please refer to Figure 4 The above-mentioned step S130 may include steps S131 to S132.

[0138] S131. Optimize the expression of the problem using a large language model to obtain the optimized problem.

[0139] In this embodiment, the optimized problem refers to a version of the problem that has been processed by a large language model, making the expression clearer, the semantics more explicit, and the problem easier to understand.

[0140] Specifically, a large language model is used to process the user's original question, making its expression clearer, more concise, and easier to understand.

[0141] Input the specific question posed by the user; utilize a general large language model to perform grammatical correction, vocabulary replacement, and sentence restructuring on the input question to eliminate ambiguity and ensure accurate expression. Output the optimized question, characterized by clearer semantics and a more reasonable structure, facilitating subsequent understanding and processing by models or humans.

[0142] S132. Analyze the optimized problem to determine whether it is related to financial and tax reporting, and classify it according to the analysis results to obtain a category.

[0143] In this embodiment, the optimized problem is analyzed in depth to determine whether the problem involves the content of financial and tax reports, and is classified as "report-related" or "non-report-related" accordingly.

[0144] Enter the optimized problem text;

[0145] The intent of the optimized question is identified using a general large language model to determine whether the core theme of the question revolves around financial and tax reports.

[0146] Based on the results of intent recognition, the problems are divided into two main categories:

[0147] Report-related: If the question involves specific financial and tax risk indicators, chapter titles, or other elements directly related to the report content, it falls into this category.

[0148] Non-report related: If the question does not involve specific report content, but is about service background, usage methods, or other aspects, it is classified into this category.

[0149] Output a category label indicating whether the issue is "report-related" or "non-report-related," providing a basis for the next step of the processing.

[0150] In this embodiment, "category" refers to the problem types classified based on the above analysis results, mainly divided into two categories:

[0151] Report related: These questions are raised in response to specific content in the uploaded financial and tax risk report, and may include, but are not limited to, financial data analysis, risk assessment, and policy interpretation.

[0152] Non-report related: These questions are not related to any specific report, but are more about service operation guidelines, technical support, customization information, and other aspects.

[0153] This classification mechanism allows the system to more accurately route user questions to the appropriate processing modules, thereby improving overall work efficiency and service quality. Simultaneously, it ensures that different types of queries receive the most suitable answer format, enhancing the professionalism and satisfaction of the user experience.

[0154] S140. When the category is a question related to the report, search for relevant information in the knowledge base to obtain search results.

[0155] In this embodiment, please refer to Figure 5The search results refer to a collection of professional information that is closely related to the content of financial and tax reports and has been selected from the knowledge base based on user questions, after similarity matching and multi-level information integration.

[0156] Specifically, based on the optimized user question, the most relevant professional information is retrieved from a pre-built vector knowledge base. This information includes, but is not limited to, tax and financial regulations and policies, explanations of professional terms, and official Q&As from tax authorities, ensuring the professionalism and accuracy of the answers.

[0157] Semantic analysis techniques are used to perform similarity matching between user questions and content in the knowledge base. By calculating similarity scores between texts, the knowledge entries most closely related to the user's question are selected, thereby providing the user with accurate information support.

[0158] The search results are not merely a list of individual knowledge points, but a multi-layered information structure formed after comprehensive processing by an intelligent agent. This includes legal provisions directly related to the user's question, background knowledge that may affect the understanding of the answer, and comparative analysis with other similar cases, in order to provide the user with a comprehensive perspective.

[0159] The system dynamically adjusts and prioritizes information based on its relevance and importance to the user's question. Key information points are given higher weight and displayed first in the final answer to help users quickly grasp the core content.

[0160] All search results must be reviewed according to security rules to ensure that they do not contain any content that violates laws, regulations, or ethical standards, while protecting the underlying information of the model from being unintentionally disclosed.

[0161] Therefore, in this embodiment, the search results are not just a simple query response, but a carefully designed and processed collection of information, designed to provide professional answers that best meet the user's needs, while also ensuring the security and legality of the information.

[0162] S150. Integrate the question, the identity information, the search results, and the parsing results into prompt words, and use a large language model to generate an answer to obtain the interpretation result.

[0163] In this embodiment, please refer to Figure 6 The interpretation result refers to the professional and accurate answer generated by integrating prompts from user questions, identity information, knowledge base search results, and report analysis content, using a large language model, and outputting an overview in Markdown format with a fixed ending. In short, it is a structured, concise, and clear professional answer provided after comprehensive analysis of user questions regarding financial and tax reports.

[0164] The process integrates the user's question, customized identity information, relevant information retrieved from the knowledge base, and the analysis results of financial and tax reports into a prompt. This prompt is then fed into a large language model to generate an answer tailored to the user's question. This process aims to combine all relevant information and context to ensure that the answer is accurate, targeted, and provides expert insights.

[0165] Specifically, for questions related to the report, prompt template II will be used. This template includes the following parts:

[0166] User questions: Specific questions that need to be answered.

[0167] Identity information: Customized identity information provided by the service purchaser to reflect the specific service background or to answer questions about the underlying information of the model.

[0168] Search results: Information highly relevant to the user's question, selected from the vector knowledge base. This information may include laws and policies, explanations of professional terms, etc., which helps to improve the professionalism of the answer.

[0169] Analysis Results: Based on the financial and tax report content processed in the previous steps, including translated complex fields, simplified information after removing irrelevant fields, and transposed long tables, the results ensure that the model can perform analysis based on clear and structured report data.

[0170] S160. Output the interpretation results.

[0171] The interpretation results are output in Markdown format and have a fixed ending.

[0172] The interpretation result refers to the detailed answer generated through the above process. It not only includes a specific response to the user's question but also adheres to the following output requirements:

[0173] Presented in a concise Markdown format, making it easy to read and understand.

[0174] Includes a fixed ending to the answer, such as: "Please note that AI cannot guarantee the complete accuracy of the answer; the actual situation still requires further judgment based on professional knowledge." This ending is intended to remind users that although AI provides professional analysis and suggestions, the final decision still depends on the knowledge and judgment of professionals.

[0175] In summary, steps S150 and S160 together realize the entire process from receiving user questions to generating and outputting professional and accurate interpretation results, thereby lowering the threshold for non-professionals to understand financial and tax risk reports and providing strong support for solving corporate financial and tax risk issues.

[0176] In one embodiment, when the category is a non-report-related question, the question, the identity information, and the parsing result are integrated into a prompt word, and an answer is generated using a large language model to obtain the interpretation result, and step S160 is executed.

[0177] First, a mechanism is needed to identify or categorize user-submitted questions as "non-report-related." This typically involves Natural Language Processing (NLP) techniques, analyzing the question's content and context to classify it into different categories. If it's determined that the question is not report-related, proceed to the next step.

[0178] Once it's confirmed that the problem is not report-related, the next step is to integrate information from the following aspects:

[0179] Question: What specific question did the user raise?

[0180] Identity information: Basic information about the questioner, such as their role (customer, employee, etc.), company, etc. This information helps to provide a more personalized and accurate answer.

[0181] Analysis results: This refers to the results obtained after analyzing the report.

[0182] The above information is integrated to form a prompt word, which will be passed as input to the large language model in order to generate a more accurate answer.

[0183] Using integrated prompts as input, the system leverages advanced Large Language Models (LLMs) to generate answers. These models, trained on vast amounts of text data, are capable of understanding and generating high-quality human language, thus producing both professional and easily understandable answers based on the provided prompts.

[0184] Ultimately, the answer generated by the large language model is the interpretation result for the original question. This answer should be structured and, as far as possible, meet the user's query requirements.

[0185] In this embodiment, please refer to Figure 3 Users first ask questions about the financial and tax report, which may be directly related to the report's content or irrelevant. Simultaneously, users upload their financial and tax reports, which are categorized as authorized reports (stored as text messages) and open-source reports (PDF files). Furthermore, service purchasers may provide customized identity information to reflect specific service backgrounds in their responses.

[0186] For authorization reports, the system reads data from the trusted database and generates a message body text in a JSON-like format. The system then performs a series of processing steps on this data: translating irregularly abbreviated fields into semantically clear expressions while retaining shorter field names to conserve tokens; removing fields irrelevant to financial and tax risk analysis or already existing in the knowledge base; transposing long tables to reduce the repetition of column names, thereby shortening the message body text; and finally converting the message body into standard JSON format and adding multi-level section headings to create a clear and structured report text.

[0187] For open-source reports, the system offers two processing modes: a fast mode and a precise mode. The fast mode directly uses the result from the PDF to Markdown converter as output. The precise mode is more complex: first, the original report text is divided into equal parts, approximately 1000 words each, and input into a general-purpose large language model in parallel to add delimiters; then, the text is segmented according to the delimiters, and input into the model again for in-depth analysis and summarization, correcting any positional errors that might have been caused by PDF parsing; finally, the segmented summary results are merged in their original order to obtain the final, organized report.

[0188] The system uses a general large language model to optimize user questions, making their semantics clearer and their expressions more concise. Next, the system determines whether the question is relevant to the content of financial and tax reports and categorizes it accordingly. For report-related questions, the system performs a similarity search in a vector knowledge base, filtering out information with high relevance to improve the professionalism of the answers. The knowledge base covers regulations and policies, explanations of professional terms, and official Q&As from the tax bureau.

[0189] Based on the question category, the system generates different prompt words and inputs them into the corresponding model to generate the answer. For non-report-related questions, the system concatenates the user's question and customized identity information into prompt word template I, and inputs it into a general dialogue model to generate the answer. The answer follows security rules and includes a fixed ending. For report-related questions, the system concatenates the user's question, knowledge base search results, report parsing content, and customized identity information into prompt word template II, and inputs it into a fine-tuned and enhanced large language model (such as the "Deep Search" financial and tax large model) to generate a detailed answer. The answer also follows security rules and includes a fixed ending.

[0190] The system returns the solutions generated based on the above steps to the user, ensuring a professional, accurate, and efficient resolution of the user's questions regarding financial and tax reports. Through this process, the intelligent agent achieves professional interpretation of financial and tax reports, lowers the barrier for non-professionals to understand financial and tax risk reports, and provides precise advice and insights.

[0191] The method's input includes user questions, uploaded reports, and customized identity information. The processing steps encompass organizing uploaded reports, processing user questions, retrieving knowledge bases (if necessary), inserting prompts with the processing results and customized identity information, and finally inputting the information into the corresponding model to obtain the answer. This technical solution aims to provide a more professional, accurate, and efficient intelligent interpretation solution for financial and tax risk reports. It not only lowers the barrier for non-professionals to read and understand financial and tax risk reports but also provides more insights and comprehensive suggestions for addressing and handling corporate financial and tax risk issues.

[0192] For example, for authorized users, the company stores their financial and tax risk report-related data in a trusted database. When retrieved, the data is organized into a JSON-like message body according to the chapter structure of the visual report. This type of message body has the advantage of a clear hierarchical structure, but it contains a lot of redundant information and requires further processing.

[0193] Translate irregular abbreviation fields into semantically clear expressions. For example, “fp_amt” should be translated into “invoice amount”, and “inflated_income_risk_for_4y” should be translated into “risk of inflated income for the past 4 years”.

[0194] Provided the field names are clear and meaningful, the version that minimizes the number of input tokens for the large model is retained. Since the number of tokens for Chinese and English characters differs in the encoding methods of downstream large language models, shorter field name representations are chosen to conserve dialogue context windows. For example, "rate" can be translated as "ratio," "proportion," or "speed," but the length of these encoded Chinese tokens exceeds the length of the English word itself. Therefore, the original English field names can be retained, allowing the large model to understand the meaning from the context.

[0195] Fields that are irrelevant to financial and tax risk analysis, or whose information already exists in the knowledge base or downstream models, such as legal basis and risk liability statements, will be discarded.

[0196] The message body length is reduced by minimizing the repetition of table column names in the JSON text. Transposition is controlled based on the number of rows and columns in the table.

[0197] Convert the message body to standard JSON format and add multi-level chapter headings in sequence (such as "I. Basic Enterprise Information", "1.1 Shareholder Information", etc.) to form the final report text that is passed to the downstream model. This helps to reduce JSON nesting and make the report hierarchy clearer.

[0198] For clients who haven't yet authorized access, this workflow offers the ability to upload open-source PDFs, allowing users to experience the report interpretation agent based on the "Deep Search" financial and tax big data model. The highly flexible encoding of PDF files has always presented a significant challenge for accurate and efficient parsing. To balance efficiency and accuracy, this workflow uses a rule-matching open-source PDF-to-Markdown converter, offering either a fast or precise mode based on user needs.

[0199] Quick mode directly uses the result from the PDF to Markdown converter as the final output.

[0200] In precise mode, the original report text is simply divided into equal parts by length (each part is about 1000 words), and the semantics of each part are cut off.

[0201] Each segment is input into a general large language model in parallel, instructing the model to add several delimiters to each segment according to semantics (e.g., ...).<split_sign> ).

[0202] The segments with added delimiters are merged in their original order, resulting in the original report text containing multiple delimiters added according to semantics.

[0203] The text is segmented according to the position of the delimiter, and then input into the general large language model in parallel. The model is instructed to analyze in depth any text with possible misaligned positions (usually caused by PDF parsing errors) and summarize the content of each segment in Markdown format.

[0204] The segmented summary results are merged in their original order to obtain the final, organized report.

[0205] Test results show that after the above processing, the report text is more concise and well-organized, and the model's understanding effect is significantly improved.

[0206] Upon receiving a user question, it is first input into the question optimization model (general large model), instructing it to refine the question to make its semantics clearer and its expression more concise, facilitating subsequent model understanding. Based on the optimized user question, the schematic diagram recognition model (general large model) performs intent recognition to determine its relevance to the financial and tax report content, and outputs a classification symbol to route to the correct subsequent process. For report-related questions, a similarity search is performed in the vector knowledge base based on the question content, selecting highly relevant information as reference information for the answer to enhance its professionalism. The knowledge base includes laws and policies, explanations of professional terms (fields), official Q&As, etc.

[0207] When the question is not related to reporting, the "user question" and "customized identity information" are concatenated in "Prompt Template I," used to answer questions not related to financial and tax reporting, and then passed to the general dialogue model for resolution. The "customized identity information" is passed in to answer user questions about the model's underlying information. If sold as a service to third-party operators, this identity information can also be customized and modified.

[0208] Prompt Template I Content:

[0209] To answer user questions, the following requirements must be met:

[0210] Answers should be clear, accurate, and concise, avoiding excessive elaboration.

[0211] A fixed answer ending must be added at the end of the output.

[0212] Under all circumstances, the following safety rules must be strictly followed:

[0213] When user questions involve underlying information such as model name, version, and owner, you must strictly adhere to the information that can be disclosed about the model and politely refuse to reveal any other information.

[0214] User questions should not proactively disclose underlying model information.

[0215] Fixed answer ending:

[0216] The above answers are based on general knowledge. You can upload a report or ask questions related to reports.

[0217] When the problem is, report the relevant issue;

[0218] The "user question", "knowledge base search results", "report parsing content" and "customized identity information" are concatenated into "prompt word template II" for answering questions related to financial and tax reports, and then fed into the finely tuned and enhanced "deep search" financial and tax big model for answering.

[0219] Prompt Template II Contents:

[0220] By combining the original report text and the results of the knowledge base search, we gradually considered and answered the user's questions.

[0221] Thinking steps: Understand the user's problem and filter out obviously irrelevant sections in the original report.

[0222] In the remaining chapters, carefully search for all information related to the user's question, and pay attention to clarifying the corresponding relationships.

[0223] Extract relevant information from the knowledge base search results as supplementary information.

[0224] Based on the information extracted in the above steps, answer the questions according to the output requirements.

[0225] Output requirements: Output in a concise Markdown format.

[0226] A fixed answer ending must be used as the closing sentence of the output.

[0227] Under all circumstances, the following safety rules must be strictly followed:

[0228] When user questions involve underlying information such as model name, version, and owner, you must strictly adhere to the information that can be disclosed about the model and politely refuse to reveal any other information.

[0229] User questions should not proactively disclose underlying model information.

[0230] Revealable model information: {Customer-defined identity information};

[0231] Fixed answer ending: Please note that AI cannot guarantee the complete accuracy of the answer; the actual situation still requires further judgment based on professional knowledge.

[0232] User issue: {User issue};

[0233] Knowledge base search results: {search results};

[0234] Original report: {Post-processing report}.

[0235] The aforementioned method for interpreting financial and tax reports involves obtaining user-related questions, the report itself, and user identity information. First, the data in the report undergoes format conversion and processing, followed by precise parsing of the open-source report to extract key information. Next, the wording of user questions is optimized, their intent identified, and their relevance to the report content categorized. For report-related queries, supplementary information is retrieved from the knowledge base to enhance the accuracy and professionalism of the answers. Finally, user questions, identity information, retrieved relevant information, and report parsing results are integrated into a prompt word template, and a specially trained large language model is used to generate detailed and easy-to-understand answers. This process not only improves the quality and professionalism of financial and tax report interpretation but also simplifies the maintenance process, effectively controls costs, and ultimately achieves the goals of improving interpretation accuracy, enhancing professionalism, and significantly reducing the difficulty for non-professionals to read and understand financial and tax reports.

[0236] Figure 7 This is a schematic block diagram of a financial and tax report interpretation device 300 provided in an embodiment of the present invention. Figure 7As shown, corresponding to the above-described method for interpreting financial and tax reports, the present invention also provides a financial and tax report interpretation device 300. This device 300 includes a unit for performing the above-described method for interpreting financial and tax reports, and can be configured in a server. Specifically, please refer to... Figure 7 The financial and tax report interpretation device 300 includes an acquisition unit 301, a parsing unit 302, an optimization unit 303, a retrieval unit 304, an interpretation unit 305, and an output unit 306.

[0237] The acquisition unit 301 is used to acquire questions, reports, and identity information related to the financial and tax report; the parsing unit 302 is used to perform format conversion and processing on the financial and tax risk report data of the report, and to parse the open-source report of the report to obtain the parsing result; the optimization unit 303 is used to optimize the expression of the question, identify the intent, and determine the category; the retrieval unit 304 is used to search for relevant information in the knowledge base when the category is a question related to the report to obtain the retrieval result; the interpretation unit 305 is used to integrate the question, the identity information, the retrieval result, and the parsing result into prompt words, and to generate an answer using a large language model to obtain the interpretation result; the output unit 306 is used to output the interpretation result.

[0238] In one embodiment, the parsing unit 302 includes:

[0239] The authorized processing subunit is used to perform format conversion, complex field, irrelevant field, long table, and further format conversion on the financial and tax risk report data of the report to obtain the authorized report processing result; the open source processing subunit is used to process the open source report of the report according to the pattern to obtain the unauthorized report processing result; wherein, the parsing result includes the authorized report processing result and the unauthorized report processing result.

[0240] In one embodiment, the authorization processing subunit includes:

[0241] The report includes a format conversion module to convert the financial and tax risk report data into JSON format text; a translation module to represent irregularly abbreviated financial and tax indicator names or chapter titles in the JSON format text using short field names to convert them into semantically clear expressions; a removal module to remove redundant fields from the translation results that are not helpful for financial and tax risk analysis or whose information already exists in the knowledge base; a transposition module to transpose the removal results into a long table by reducing the number of repeated column names; and a re-conversion module to organize the transposed results into standard JSON format and add multi-level chapter titles to obtain the authorized report processing result.

[0242] In one embodiment, the open-source subunit includes:

[0243] The conversion module is used to process the open-source report in the quick mode of the report using the result of an open-source PDF to Markdown converter as the unauthorized report processing result; the segmentation module is used to divide the open-source report in the precise mode of the report into equal parts according to length, and input them in parallel into a large language model for semantic segmentation and tagging to obtain preliminary processing results; the re-segmentation module is used to further segment the text according to the segmentation symbols added in the preliminary processing results, and then input them into the large language model for detailed analysis and summary, correcting possible positional errors caused by PDF parsing, to obtain the summary results of each paragraph; the merging module is used to merge the summary results of each paragraph in the original order to form the unauthorized report processing result.

[0244] In one embodiment, the optimization unit 303 includes:

[0245] The expression optimization subunit is used to optimize the expression of the problem through a large language model to obtain an optimized problem; the classification subunit is used to analyze the optimized problem to determine whether it is related to financial and tax reports, and classify it according to the analysis results to obtain a category.

[0246] In one embodiment, the tax report answering device further includes:

[0247] The non-report processing unit is used to integrate the question, the identity information, and the parsing result into a prompt word when the category is a non-report-related question, and to generate an answer using a large language model to obtain an interpretation result, and to execute the output of the interpretation result.

[0248] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned financial and tax report interpretation device 300 and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0249] The aforementioned financial and tax report interpretation device 300 can be implemented as a computer program, which can, for example... Figure 8 It runs on the computer device shown.

[0250] Please see Figure 8 , Figure 8 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0251] See Figure 8The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0252] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform a method for interpreting financial and tax reports.

[0253] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0254] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for interpreting financial and tax reports.

[0255] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0256] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps:

[0257] The process involves: acquiring questions, reports, and identity information related to financial and tax reports; converting and processing the financial and tax risk report data in the reports, and parsing the open-source reports to obtain parsing results; optimizing the wording of the questions, identifying intent, and determining categories; when the category is a question related to the report, searching for relevant information in the knowledge base to obtain retrieval results; integrating the questions, identity information, retrieval results, and parsing results into prompt words, and using a large language model to generate answers to obtain interpretation results; and outputting the interpretation results.

[0258] The interpretation results are output in Markdown format and have a fixed ending.

[0259] In one embodiment, when the processor 502 performs the steps of format conversion and processing of the financial and tax risk report data in the report, and parses the open-source report of the report to obtain the parsing result, the specific implementation is as follows:

[0260] The financial and tax risk report data in the report is processed by format conversion, complex field removal, irrelevant field removal, long table removal, and further format conversion to obtain the authorized report processing result; the open source report of the report is processed according to the pattern to obtain the unauthorized report processing result.

[0261] The parsing results include the results of authorized report processing and the results of unauthorized report processing.

[0262] In one embodiment, when the processor 502 performs the steps of format conversion, complex field, irrelevant field, long table, and further format conversion on the financial and tax risk report data of the report to obtain the authorized report processing result, the specific steps are as follows:

[0263] The financial and tax risk report data in the report is converted into JSON format text. Irregularly abbreviated financial and tax indicator names or chapter titles in the JSON format text are represented by short field names to convert them into semantically clear expressions, resulting in a translation result. Redundant fields that are not helpful for financial and tax risk analysis or whose information already exists in the knowledge base are removed from the translation result, resulting in a rejection result. The rejection result is then transposed into a long table by reducing the number of repeated column names, resulting in a transposed result. The transposed result is then organized into a standard JSON format and multi-level chapter titles are added to obtain the authorized report processing result.

[0264] In one embodiment, when the processor 502 implements the step of processing the open-source report of the report according to the pattern to obtain the processing result of the unauthorized report, the specific steps are as follows:

[0265] The open-source report in the quick mode of the aforementioned report uses the result of an open-source PDF to Markdown converter as the unauthorized report processing result; the open-source report in the precise mode of the aforementioned report is divided into equal parts by length and input into a large language model in parallel for semantic segmentation and tagging to obtain preliminary processing results; the text is further segmented based on the segmentation symbols added according to the preliminary processing results, and then input into the large language model for detailed analysis and summarization to correct any positional errors that may have been caused by PDF parsing, in order to obtain summary results for each paragraph; the summary results of each paragraph are merged in their original order to form the unauthorized report processing result.

[0266] In one embodiment, when the processor 502 implements the step of optimizing the problem statement and identifying intent to obtain the category, it specifically implements the following steps:

[0267] The problem is formulated by optimizing the expression of the problem using a large language model to obtain an optimized problem; the optimized problem is analyzed to determine whether it is related to financial and tax reporting, and the results of the analysis are classified to obtain a category.

[0268] In one embodiment, after implementing the steps of optimizing the problem statement and identifying intent to obtain a category, the processor 502 further implements the following steps:

[0269] When the category is a non-report-related question, the question, the identity information, and the parsing result are integrated into a prompt word, and an answer is generated using a large language model to obtain the interpretation result, and the interpretation result is output.

[0270] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0271] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0272] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps:

[0273] The process involves: acquiring questions, reports, and identity information related to financial and tax reports; converting and processing the financial and tax risk report data in the reports, and parsing the open-source reports to obtain parsing results; optimizing the wording of the questions, identifying intent, and determining categories; when the category is a question related to the report, searching for relevant information in the knowledge base to obtain retrieval results; integrating the questions, identity information, retrieval results, and parsing results into prompt words, and using a large language model to generate answers to obtain interpretation results; and outputting the interpretation results.

[0274] The interpretation results are output in Markdown format and have a fixed ending.

[0275] In one embodiment, when the processor executes the computer program to perform format conversion and processing of the financial and tax risk report data in the report, and parses the open-source report of the report to obtain the parsing result, the processor specifically implements the following steps:

[0276] The financial and tax risk report data in the report is processed by format conversion, complex field removal, irrelevant field removal, long table removal, and further format conversion to obtain the authorized report processing result; the open source report of the report is processed according to the pattern to obtain the unauthorized report processing result.

[0277] The parsing results include the results of authorized report processing and the results of unauthorized report processing.

[0278] In one embodiment, when the processor executes the computer program to perform format conversion, complex field, irrelevant field, long table, and further format conversion on the financial and tax risk report data of the report to obtain the authorized report processing result, the processor specifically implements the following steps:

[0279] The financial and tax risk report data in the report is converted into JSON format text. Irregularly abbreviated financial and tax indicator names or chapter titles in the JSON format text are represented by short field names to convert them into semantically clear expressions, resulting in a translation result. Redundant fields that are not helpful for financial and tax risk analysis or whose information already exists in the knowledge base are removed from the translation result, resulting in a rejection result. The rejection result is then transposed into a long table by reducing the number of repeated column names, resulting in a transposed result. The transposed result is then organized into a standard JSON format and multi-level chapter titles are added to obtain the authorized report processing result.

[0280] In one embodiment, when the processor executes the computer program to implement the step of processing the open-source report of the report according to the pattern to obtain the processing result of the unauthorized report, the processor specifically implements the following steps:

[0281] The open-source report in the quick mode of the aforementioned report uses the result of an open-source PDF to Markdown converter as the unauthorized report processing result; the open-source report in the precise mode of the aforementioned report is divided into equal parts by length and input into a large language model in parallel for semantic segmentation and tagging to obtain preliminary processing results; the text is further segmented based on the segmentation symbols added according to the preliminary processing results, and then input into the large language model for detailed analysis and summarization to correct any positional errors that may have been caused by PDF parsing, in order to obtain summary results for each paragraph; the summary results of each paragraph are merged in their original order to form the unauthorized report processing result.

[0282] In one embodiment, when the processor executes the computer program to implement the formulation of the problem and identify the intent to obtain the category step, it specifically implements the following steps:

[0283] The problem is formulated by optimizing the expression of the problem using a large language model to obtain an optimized problem; the optimized problem is analyzed to determine whether it is related to financial and tax reporting, and the results of the analysis are classified to obtain a category.

[0284] In one embodiment, after executing the computer program to implement the formulation of the problem and identify the intent to obtain the category step, the processor further implements the following steps:

[0285] When the category is a non-report-related question, the question, the identity information, and the parsing result are integrated into a prompt word, and an answer is generated using a large language model to obtain the interpretation result, and the interpretation result is output.

[0286] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0287] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0288] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0289] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0290] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0291] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. Methods for interpreting financial and tax reports, characterized by: include: Obtain questions, reports, and identity information regarding financial and tax reporting; The financial and tax risk report data in the report is converted and processed in a different format, and the open-source report of the report is parsed to obtain the parsing results; Optimize the wording of the problem, identify the intent, and determine the category; When the category is a question related to the report, relevant information is searched in the knowledge base to obtain search results; The question, the identity information, the search results, and the parsing results are integrated into prompt words, and a large language model is used to generate the answer to obtain the interpretation result; Output the interpretation results.

2. The method for interpreting financial and tax reports according to claim 1, characterized in that, The process involves format conversion and processing of the financial and tax risk report data, and parsing the open-source report to obtain the parsing results, including: The financial and tax risk report data in the report is subjected to format conversion, complex fields, irrelevant fields, long tables, and further format conversion to obtain the authorized report processing result; The open-source report is processed according to a pattern to obtain the unauthorized report processing result; The parsing results include the results of authorized report processing and the results of unauthorized report processing.

3. The method for interpreting financial and tax reports according to claim 2, characterized in that, The process of converting the financial and tax risk report data in the report, removing complex fields, irrelevant fields, and long tables, and then re-converting the format to obtain the authorized report processing result includes: The financial and tax risk report data in the report is converted to obtain JSON format text; Irregularly abbreviated financial and tax indicator names or chapter titles in the JSON format text are represented by short field names to convert them into semantically clear expressions, thus obtaining the translation results; Redundant fields that are not helpful for financial and tax risk analysis or whose information already exists in the knowledge base are removed from the translation results to obtain the elimination results; The elimination results are then transposed by reducing the number of repeated column names to obtain the transposed result. The transposed results are organized into a standard JSON format, and multi-level chapter headings are added to obtain the authorization report processing results.

4. The method for interpreting financial and tax reports according to claim 3, characterized in that, The process of processing the open-source report according to a pattern to obtain the unauthorized report processing result includes: The open-source report in the quick mode of the aforementioned report uses the result of an open-source PDF to Markdown converter as the unauthorized report processing result; The open-source report of the precise pattern of the report is divided into equal parts according to length and input into a large language model in parallel for semantic segmentation and labeling to obtain preliminary processing results; The text is further segmented based on the delimiters added according to the preliminary processing results. Then, it is input into a large language model for detailed analysis and summarization to correct the positional misalignment problem that may be caused by PDF parsing, so as to obtain the summary results of each paragraph. The summaries of the paragraphs are merged in their original order to form the unauthorized report processing result.

5. The method for interpreting financial and tax reports according to claim 1, characterized in that, The optimization of the problem statement, identifying intent, and obtaining categories includes: The problem statement is optimized using a large language model to obtain an optimized problem. The optimized problems are analyzed to determine whether they are related to financial and tax reporting, and the results of the analysis are categorized to obtain a class.

6. The method for interpreting financial and tax reports according to claim 1, characterized in that, After optimizing the problem statement, identifying the intent, and obtaining the category, the process further includes: When the category is a non-report-related question, the question, the identity information, and the parsing result are integrated into a prompt word, and an answer is generated using a large language model to obtain the interpretation result, and the interpretation result is output.

7. The method for interpreting financial and tax reports according to claim 1, characterized in that, The interpretation results are output in Markdown format and have a fixed ending.

8. A financial and tax report interpretation device, characterized in that, include: The acquisition unit is used to acquire questions, reports, and identity information related to financial and tax reports. The parsing unit is used to convert and process the financial and tax risk report data in the report, and to parse the open-source report of the report to obtain the parsing results; An optimization unit is used to optimize the description of the problem, identify the intent, and determine the category; The retrieval unit is used to search for relevant information in the knowledge base when the category is a question related to the report, so as to obtain retrieval results; The interpretation unit is used to integrate the question, the identity information, the search results, and the parsing results into prompt words, and use a large language model to generate an answer to obtain the interpretation result; The output unit is used to output the interpretation result.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images