Report generation method based on large model, electronic equipment and computer program product

Abstract

The invention discloses a report generation method based on a large model, electronic equipment and a computer program product, and relates to the technical field of artificial intelligence, the report generation method comprises the following steps: identifying a logic structure, a placeholder and a task dependency relationship in a preset report template, and initializing a context state carrier based on the logic structure, the placeholder and the task dependency relationship; obtaining a report generation requirement, converting the report generation requirement into SQL query data, and generating a cue word based on the SQL query data; and inputting the cue word into a preset generation model to obtain filling content, updating the filling content to the context state carrier, and filling placeholders of the report template based on the context state carrier to obtain a final report. According to the method, the defect of low efficiency of manual data arrangement is overcome, and the problem that a simple template filling tool cannot process a complex logical relationship is solved, so that the report generation efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a report generation method based on a large model, an electronic device and a computer program product. BACKGROUND

[0002] In the fields of financial analysis, policy research, industry research, etc., standardized reports are the core carriers of information transmission and decision support, and their generation efficiency and quality directly affect the timeliness and accuracy of business development. Traditional report generation relies on manual writing or simple template filling tools, which has efficiency bottlenecks. Manual writing requires a lot of time for data sorting, logical analysis and content organization, and is prone to human errors. Although simple template tools can achieve uniform formats, they lack deep analysis capabilities for report logical frameworks and cannot dynamically adjust content structures according to different business scenarios. The template reusability is poor, and it is difficult to adapt to diversified report requirements, which affects the efficiency of report generation.

[0003] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0004] The main purpose of the present application is to provide a report generation method based on a large model, an electronic device and a computer program product, aiming to solve the technical problem of low efficiency of current report generation methods.

[0005] To achieve the above purpose, the present application provides a report generation method based on a large model, which comprises: identifying the logical structure, placeholder and task dependency relationship in the preset report template, and initializing a context state carrier based on the logical structure, the placeholder and the task dependency relationship; obtaining report generation requirements and converting the report generation requirements into SQL query data, and generating prompt words based on the SQL query data; inputting the prompt words into a preset generation model to obtain filled content, updating the filled content to the context state carrier, and filling the context state carrier to the placeholder of the report template to obtain a final report.

[0006] In an embodiment, the step of identifying the logical structure, placeholder and task dependency relationship in the preset report template comprises: parsing chapters and titles in the report template to obtain the logical structure in the report template, and parsing the placeholder in the report template; determining the task dependency relationship based on the chapters and the titles, wherein the task dependency relationship is used to represent the context dependency between the chapters.

[0007] In an embodiment, before the step of generating the prompt word based on the SQL query data, the method further comprises: text cleaning is performed on the industry white paper and policy document to obtain cleaned text, a pre-trained language model is used to generate a text vector embedding based on the cleaned text, and the text vector embedding is stored in a vector database; The step of generating the prompt word based on the SQL query data comprises: Based on the report generation requirement, a target text vector embedding is found in the vector database, and a prompt word is generated based on the target text vector embedding and the SQL query data.

[0008] In an embodiment, the step of generating the prompt word based on the SQL query data comprises: A plurality of language vector embeddings corresponding to respective languages are generated, and each language vector embedding is stored in the vector database; The step of generating the prompt word based on the report generation requirement and finding the target text vector embedding in the vector database based on the target text vector embedding and the SQL query data comprises: Based on the report generation requirement, a target text vector embedding is found in the vector database, and a target language vector embedding is found in the vector database based on the target language of the report generation requirement; A prompt word is generated based on the target text vector embedding, the target language vector embedding, and the SQL query data.

[0009] In an embodiment, the generation model comprises a data analysis model and a text generation model; The step of inputting the prompt word into a preset generation model to obtain filled content comprises: A numerical prompt word in the prompt word is inputted into the data analysis model to obtain a statistical calculation result; A text prompt word in the prompt word is inputted into the text generation model to obtain an analysis text; Filled content is generated based on the statistical calculation result and the analysis text.

[0010] In an embodiment, the step of updating the filled content to the context state carrier and filling the placeholders in the report template based on the context state carrier to obtain a final report comprises: A report title is determined based on a dependency-free starting task, and the report title is written into the context state carrier to obtain a target carrier; A chapter list is traversed, the target carrier is updated by the filled content of each chapter, and the updated target carrier is filled into the placeholders in the report template to obtain a final report.

[0011] In an embodiment, the step of updating the filling content to the context state carrier and filling the placeholders in the report template based on the context state carrier to obtain a final report comprises: inputting the placeholder instructions into a business database to obtain business data; inputting the business data and the placeholder type corresponding to the business data into a preset generator to obtain multi-modal content; filling the multi-modal content into the placeholders in the report template based on the context state carrier to obtain a final report.

[0012] In an embodiment, after the step of updating the filling content to the context state carrier and filling the placeholders in the report template based on the context state carrier to obtain a final report, the method further comprises: inputting the final report into a sensitive information detection module to obtain a sensitive content detection result; if the sensitive content detection result indicates that there is sensitive content in the final report, performing desensitization processing on the final report by using a regular expression to obtain a desensitized report.

[0013] In addition, to achieve the above object, the present application further provides an electronic device, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and the computer program is configured to implement the steps of the data processing method as described above.

[0014] In addition, to achieve the above object, the present application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the steps of the report generation method as described above.

[0015] The application identifies the logical structure, placeholders and task dependency relationship in the report template, constructs a dynamically adjustable context state carrier, solves the problem that the traditional template tool cannot parse the document logical framework, enables the system to understand the hierarchical relationship and dependency constraint between chapters. Secondly, the generation requirements in natural language form are converted into structured SQL query data, realizing accurate mapping from requirements to data operation and overcoming the defect of low efficiency of manual data sorting. Based on the design of the SQL generation prompt word, the normativity of data retrieval is ensured, and the semantic clear input condition is provided for the subsequent generation model. By combining the filling content output by the generation model with the context state carrier, the multi-position content filling is completed under the premise of maintaining the overall logical consistency of the report, realizing the deep integration of automatic generation and manual template, and the continuous updating mechanism of the context state carrier in the application can dynamically transmit the dependency information between chapters, ensure the coherence of the generated content, break through the problem that the simple template filling tool cannot handle complex logical relationship, and improve the report generation efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings, which are incorporated herein and form part of the specification, illustrate embodiments consistent with the present application and, together with the description, further serve to explain the principles of the application.

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, for those skilled in the art, other drawings can also be obtained based on these drawings without creative labor.

[0018] Figure 1 The flowchart provided by the first embodiment of the report generation method based on the large model of the present application is shown in the figure. Figure 2 The flowchart provided by the second embodiment of the report generation method based on the large model of the present application is shown in the figure. Figure 3 The flowchart provided by the third embodiment of the report generation method based on the large model of the present application is shown in the figure. Figure 4 The brief flowchart of the report generation method based on the large model provided by an embodiment of the present application is shown in the figure. Figure 5 The module structure diagram of the report generation device based on the large model of the embodiment of the present application is shown in the figure. Figure 6 The device structure diagram of the hardware running environment involved in the report generation method based on the large model of the embodiment of the present application is shown in the figure.

[0019] The object implementation, functional characteristics and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0020] It should be understood that the specific embodiments described herein are merely intended to explain the technical solutions of the present application, and are not intended to limit the present application.

[0021] In order to better understand the technical solutions of the present application, the following will be described in detail in combination with the drawings of the specification and specific embodiments.

[0022] Since the traditional report generation depends on manual writing or simple template filling tools, there is an efficiency bottleneck. Manual writing needs to spend a lot of time on data arrangement, logical analysis and content organization, and is easy to miss due to human operation. Although the simple template tool can realize format uniformity, it lacks deep analysis ability of the report logical framework, cannot dynamically adjust the content structure according to different business scenarios, has poor template reusability, and is difficult to adapt to diversified report requirements, resulting in that the report generation efficiency is affected.

[0023] The present application provides a solution. By identifying the logical structure, placeholder and task dependency relationship in the report template, a context state carrier that can be dynamically adjusted is constructed, the problem that the traditional template tool cannot analyze the document logical framework is solved, so that the system can understand the hierarchical relationship and dependency constraint between chapters. Secondly, the generation requirement in natural language form is converted into structured SQL query data, realizing accurate mapping from requirement to data operation, overcoming the defect of low efficiency of manual data arrangement. Based on the design of SQL generation prompt word, the normativity of data retrieval is ensured, and the semantic clear input condition is provided for the subsequent generation model. By combining the filling content output by the generation model with the context state carrier, multi-position content filling is completed under the premise of keeping the overall logical consistency of the report, realizing the deep integration of automatic generation and manual template, and the continuous updating mechanism of the context state carrier can dynamically transmit the dependency information between chapters, ensuring the coherence of the generated content, breaking through the problem that the simple template filling tool cannot handle complex logical relationship, thereby improving the report generation efficiency.

[0024] It should be noted that the execution subject of the present embodiment can be a computing service device with report generation, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions. The present embodiment and the following embodiments will be described below taking the electronic device as an example.

[0025] Based on this, the present application provides a report generation method based on a large model. Referring to Figure 1 , Figure 1 is a flowchart of the first embodiment of the report generation method based on a large model of the present application.

[0026] In this embodiment, the report generation method based on a large model comprises steps S10-S30: Step S10, identify the logical structure, placeholder and task dependency relationship in the preset report template, and initialize the context state carrier based on the logical structure, the placeholder and the task dependency relationship.

[0027] The logical structure refers to the hierarchical relationship and arrangement order of the chapter titles in the template, which can be specifically implemented by using natural language processing technology to parse the title tags in the template, and is used to determine the organization framework of the report content. The placeholder refers to a symbol in the template that marks the content to be filled, which can be specifically implemented by matching the identifier wrapped in angle brackets through a regular expression, and is used to locate the content area that needs to be dynamically generated. The task dependency relationship refers to the data or logical association between different chapters, which can be specifically implemented by analyzing the nesting relationship and context keywords of the chapter titles, and is used to constrain the order and relevance of content generation. The context state carrier refers to a data structure that stores the intermediate results of chapter generation, which can be specifically implemented by using a JSON object or an in-memory database table, and is used to temporarily store the information of the generated chapters during the report generation process. The context state carrier at least includes: a generated content set and a task dependency relationship, wherein the generated content set stores the content corresponding to the placeholder in each completed filled chapter in the form of key-value pairs, and the task dependency relationship records the dependency relationship between chapters, which is used to schedule the generation order.

[0028] First, parse the chapter titles and placeholders of the report template to establish the logical structure containing the chapter hierarchy and dependency relationship. For example, when the "market analysis" chapter depends on the "data overview" chapter in the template, the system will preferentially generate the basic data part.

[0029] Step S20, obtain the report generation requirement, and convert the report generation requirement into SQL query data, and generate prompt words based on the SQL query data.

[0030] The SQL query data refers to the structured query instruction converted from the natural language requirement, which can be specifically implemented by mapping the user input to a database operation statement through a semantic parsing model, and is used to accurately extract business data.

[0031] Convert the report requirement submitted by the user into a SQL statement, extract related data from a business database, and generate prompt words containing data fields and semantic constraints by associating the SQL query result with the template placeholder.

[0032] Step S30, input the prompt words into a preset generation model to obtain filled content, update the filled content to the context state carrier, and fill the context state carrier to the placeholder of the report template to obtain the final report.

[0033] The model outputs analysis text and statistical results according to the prompt words, for example, converts the sales growth rate data into a trend description sentence. During the filling process, the context state carrier updates the key information of the generated chapter in real time, ensuring data consistency when subsequent chapters are referenced. When filling the "competition analysis" chapter, the system automatically obtains the regional sales data generated in the previous stage from the carrier, avoiding repeated queries or logical conflicts.

[0034] Specifically, the step of filling content into placeholders of the report template based on the context state carrier to obtain the final report includes: determining a chapter filling order based on the task dependency relationship in the context state carrier; traversing the chapter list, for each chapter, updating the filling content corresponding to the chapter to the context state carrier, and inserting the filling content into the corresponding placeholder of the report template according to the updated context state carrier, thereby generating the final report. Compared with the traditional template tool that cannot identify the dependency relationship between chapters, resulting in contradictions in the generated content. For example, after manually adjusting the chapter order, the reference relationship needs to be manually corrected one by one. However, the embodiment dynamically maintains the data association between chapters through the context state carrier, ensuring the coherence of content generation. In addition, the existing method requires manual writing of SQL query statements to extract data, while the embodiment reduces the data preparation time by automatically converting user requirements. Thus, the embodiment realizes the automatic logic analysis and dynamic content filling of the report generation process. Through structured identification of template elements and dependency relationships, the problem of poor template reusability of traditional tools is solved; through automatic conversion of requirements to SQL instructions, the time cost of manual data sorting is reduced; through real-time updating of the context state carrier, the logical consistency of multi-chapter content is ensured, thereby improving the generation efficiency of complex reports.

[0035] In a feasible embodiment, after the step S30 of inputting the prompt words into the preset generation model to obtain the filling content, and filling the filling content into placeholders of the report template based on the context state carrier to obtain the final report, the method further includes: Step S40, inputting the final report into a sensitive information detection module to obtain a sensitive content detection result.

[0036] The sensitive information detection module refers to an automatic processing unit for content screening based on a preset rule set, which can be implemented by using a keyword matching algorithm or a machine learning classifier, and is used to identify personal privacy data or business secret information contained in the report.

[0037] Step S50, if the sensitive content detection result indicates that there is sensitive content in the final report, performing desensitization processing on the final report through a regular expression to obtain a desensitized report.

[0038] The regular expression desensitization processing refers to locating sensitive fields and performing masking or replacing operations through pattern matching rules, which can be implemented by using a string replacement function combined with a regular expression engine, and is used to eliminate sensitive information while maintaining the integrity of the text structure.

[0039] Specifically, a sensitive information detection module is set at the end of the report generation process, and the full text of the report is scanned by loading a preset sensitive word library and detection rules. When content meeting the preset sensitive characteristics is detected, a desensitization processing flow is triggered, and a regular expression engine is called to perform pattern matching on the detected sensitive fields, for example, replacing consecutive digits in an ID number with asterisks while preserving the original data of non-sensitive fields. During the processing, a context state carrier is maintained to record the desensitization operation log, ensuring data traceability and reusability of the original report template.

[0040] Compared with the traditional report generation system which usually relies on manual review for sensitive information processing, there is a low efficiency and a risk of missed detection. The embodiment realizes closed-loop control of the sensitive information processing flow through the cooperative mechanism of programmed detection and automated desensitization, completes compliance verification without human intervention, and avoids misoperations that may be caused by traditional string replacement methods through precise matching of regular expressions. Thus, the embodiment realizes the automatic identification and processing of sensitive information in the report generation process, prevents the efficiency bottleneck and operation errors caused by manual review, and ensures that the output report meets the data security specification requirements. Through the precise matching mechanism of regular expressions, the directed masking of sensitive fields is completed while maintaining the logical structure of the original data, avoiding the data distortion problem that may be caused by traditional global replacement methods.

[0041] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as the above-mentioned first embodiment can be referred to the above introduction, and will not be described hereinafter. On this basis, please refer to Figure 2 , step S10: a step of identifying the logical structure, placeholders and task dependency relationship in the preset report template, comprising: Step S101, parsing chapters and titles in the report template to obtain the logical structure in the report template, and parsing the placeholders in the report template.

[0042] The logical structure refers to the hierarchical organization relationship of chapters and titles in the template, which can be implemented by using natural language processing technology to analyze the semantic hierarchy of chapter titles, and constructing a tree structure by identifying title serial numbers, font styles and position information.

[0043] Step S102, determining a task dependency relationship based on the chapters and the titles, wherein the task dependency relationship is used to represent the context dependency between the chapters.

[0044] Task dependency refers to the content correlation between chapters, which can be achieved by analyzing the semantic correlation of chapter titles using graph theory algorithms. A dependency graph is constructed by calculating the co-occurrence frequency of title keywords.

[0045] Specifically, during the chapter parsing process, the hierarchical information of the title is extracted through the semantic analysis module, such as the membership relationship between the main title and the sub-title, forming a structured framework. During the placeholder identification process, the pattern matching algorithm is used to locate the filling area, such as identifying the "sales_data" form of the marker. During the dependency relationship construction process, the execution order is established according to the semantic correlation of the chapter title, such as the "market analysis" chapter needs to be generated after the "data overview" chapter is completed, ensuring that the analysis conclusion references the correct data basis.

[0046] Compared with the traditional method of only parsing the format layout of the template and ignoring the logical structure, resulting in the fixed order of chapter generation. This embodiment dynamically constructs task dependency through semantic correlation analysis, such as when detecting that the "conclusion" chapter references a specific chart of the "experimental data" chapter, automatically establishing a cross-chapter dependency link, avoiding data reference errors. The placeholder in the prior art only supports single variable replacement, and this embodiment supports nested placeholder processing through structured parsing, such as identifying multi-level placeholders within table cells, thereby realizing accurate parsing of the template logical framework and solving the problem of mispositioning of generated content caused by incomplete understanding of the template structure.

[0047] In a feasible embodiment, before step S20: generating prompt words based on the SQL query data, it further includes: Step S01: performing text cleaning on industry white papers and policy documents to obtain cleaned text, generating text vector embedding based on the cleaned text through a pre-trained language model, and storing the text vector embedding into a vector database.

[0048] Text cleaning refers to the format standardization and redundant information filtering of unstructured documents, which can be achieved by using regular expression matching and stop word filtering to eliminate layout markers and irrelevant content in the original document. Text vector embedding refers to converting semantic information into high-dimensional numerical representation through a deep learning model, which can be achieved by using BERT or RoBERTa model to establish semantic association relationship between documents. Vector database refers to a database system that stores multi-dimensional vector data and provides similarity retrieval function, which can be built by using FAISS or Milvus framework to realize efficient retrieval of large-scale vector data.

[0049] Specifically, the industry white paper and policy document after text cleaning retains the core policy provisions and technical specifications, and the pre-trained language model is used to convert the cleaned text into vector data with semantic representation ability. These vector data are stored in a vector database supporting fast retrieval.

[0050] Step S20: generating a prompt word based on the SQL query data, including: step S201, generating a requirement to find a target text vector embedding in the vector database based on the report, and generating a prompt word based on the target text vector embedding and the SQL query data.

[0051] The target text vector embedding refers to the vector segment with the highest semantic correlation with the current report topic, which can be calculated by a cosine similarity algorithm and is used to extract industry specification content matching business requirements.

[0052] When a specific report generation request is received, similarity retrieval is performed in the vector database according to the request topic, and the most relevant policy provision vector representation of the current business scenario is obtained. The target text vector embedding, that is, the vector representation and the structured query data together constitute the input instruction of the generation model, so that the model can synchronously integrate business data and industry specification requirements when generating report content. For example, when generating a carbon emission report, the system automatically retrieves the carbon emission calculation standard vector in the latest environmental protection policy to ensure that the calculation method in the report meets the current regulations.

[0053] Compared with the traditional method of relying on manual review of policy documents for content writing, there is a risk of information lag and subjective bias. The present embodiment dynamically integrates the latest industry specifications through an automated vector retrieval mechanism and accurately locates relevant provisions using semantic matching technology, avoiding the omission of key information in the manual screening process. The keyword matching-based document retrieval method in the prior art cannot handle synonym and semantic extension problems. The present embodiment realizes accurate association at the semantic level through vector similarity calculation. Thus, the present embodiment can automatically extract industry policy information highly relevant to the report topic and ensure that the generated content meets the technical specification requirements of the professional field. The cleaned text data eliminates noise interference and improves the accuracy of subsequent processing; the vector storage mechanism supports fast retrieval of the latest policy provisions to ensure the timeliness of the report content; the semantic association technology realizes intelligent fusion of business data and industry specifications, reducing the misuse of professional terms caused by manual intervention.

[0054] In a feasible embodiment, before step S20: generating a prompt word based on the SQL query data, it further includes: Step S02, generating a plurality of language vector embeddings corresponding to each language, and storing each language vector embedding in the vector database; The language vector embedding refers to a mathematical representation of converting semantic features of different languages into a unified vector space expression. Specifically, a pre-trained language model can be used to encode multilingual corpus to achieve this feature, which allows semantic information of different languages to be calculated in the same vector space. In the process of generating prompt words, first, a pre-trained model is used to encode multilingual corpus to form a vector database containing language features.

[0055] Step S201: Based on the report generation requirement, find the target text vector embedding in the vector database, and generate prompt words based on the target text vector embedding and the SQL query data, including: Step S2011, based on the report generation requirement, find the target text vector embedding in the vector database, and find the target language vector embedding in the vector database based on the target language of the report generation requirement.

[0056] The target language vector embedding refers to the mapping result of the grammar rules and expression habits of a specific language in the vector space. Specifically, it can be achieved by training or fine-tuning a multilingual model on target language corpus. This feature can capture the structural features specific to the target language.

[0057] When receiving a report generation request containing a target language identifier, the text vector related to the industry policy and the language vector matching the target language are retrieved from the vector database in parallel. The text vector ensures the accuracy of professional terms, and the language vector captures the grammatical structure features.

[0058] Step S2012, based on the target text vector embedding, the target language vector embedding and the SQL query data, generate prompt words.

[0059] The target text vector embedding and the target language vector embedding are fused with the SQL query statement in the vector space to generate a prompt word input generation model that contains both accurate data query results and conforms to the expression habits of the target language. For example, when generating an English report, the language vector carries the tense and word order features of English. After combining with the Chinese policy text vector, the generation model can output an analysis paragraph that conforms to the English grammar specifications and is accurate in content.

[0060] Compared with the traditional cross-language report generation relying on manual translation or fixed templates, there are problems of semantic deviation and context rupture. The existing technology cannot dynamically adapt to the grammar structure of different languages, resulting in errors in the generated content or harsh expression. The embodiment solves the problem of semantic consistency in a multilingual environment by automatically adapting the expression rules of the target language through the retrieval mechanism of language vector embedding while maintaining the accuracy of the original data. Thus, the embodiment realizes dynamic adaptation of language features in the cross-language report generation process, ensuring that the generated content not only conforms to the grammar habits of the target language, but also accurately reflects the semantics of the original data. The scheme solves the problem of semantic deviation in multilingual prompt word generation, improving the professionalism and readability of cross-language reports.

[0061] Based on the first and / or second embodiments of the present application, in the third embodiment of the present application, the same or similar content as the above embodiment one and / or embodiment two can refer to the above introduction, and the subsequent will not be described again. On this basis, the generation model includes a data analysis model and a text generation model, please refer to Figure 3 Step S30: inputting the prompt word into a preset generation model to obtain a filled content, comprising: Step S301: inputting the numerical prompt word in the prompt word into the data analysis model to obtain a statistical calculation result.

[0062] The data analysis model refers to a calculation engine specially processing structured numerical data, which can be specifically implemented by a statistical software package or a machine learning regression algorithm library, and its function is to accurately mathematically operate the prompt word containing numerical indicators. The numerical prompt word refers to an operation instruction containing statistical parameters and business indicators, such as a mathematical expression of "calculating the Q3 sales growth rate". The statistical calculation result refers to the quantitative data output obtained by numerical operation, such as percentage, trend chart and other visual data. The analysis text refers to the coherent word content generated by semantic understanding, such as paragraph description of industry competition pattern.

[0063] When the system receives a prompt word of mixed data types, the semantic analysis module is first used to distinguish numerical prompt words from text prompt words. The numerical prompt word is routed to the data analysis model, and a preset statistical algorithm is called to perform accurate calculation, such as using a time series analysis model to process sales growth rate data.

[0064] Step S302: inputting the text prompt word in the prompt word into the text generation model to obtain an analysis text.

[0065] The text generation model refers to a generation engine for processing unstructured text data, which can be implemented by using a pre-trained language model based on a Transformer architecture. The text generation model is used for context-aware text generation of semantic prompts. The text prompt refers to a natural language instruction containing analysis requirements and description requirements, such as a semantic request “elaborate the influencing factors of market trends”. The text prompt is input into the text generation model to generate a logical paragraph through an attention mechanism, such as generating an industry risk analysis based on historical policy texts.

[0066] In step S303, the filling content is generated based on the statistical calculation result and the analysis text.

[0067] The output results of the data analysis model and the text generation model are aligned in format and connected in semantics in the content integration layer, for example, a statistical chart is inserted into a corresponding position of the analysis text to form a complete chapter content. Compared with the traditional method of using a single model to process mixed data types, numerical calculation is easily disturbed by text noise, resulting in a decrease in accuracy, and text generation is affected by mathematical symbols, resulting in a lack of semantic coherence. The embodiment establishes independent processing channels, the data analysis model focuses on mathematical operations to avoid semantic interference, and the text generation model focuses on language generation without being affected by formula symbols. The two types of processing processes are physically isolated at the data level and logically associated at the output level. Thus, the embodiment realizes the shunt processing of structured data and unstructured data, and eliminates the mutual interference in mixed data processing. The numerical calculation module ensures the mathematical rigor of the statistical result, for example, it avoids the influence of approximate expressions in the text description on the calculation accuracy when calculating the compound growth rate; the text generation module maintains the integrity of the semantic logic, for example, it is not disturbed by numerical formulas when analyzing market dynamics, thereby simultaneously improving the accuracy of data processing and the readability of text output.

[0068] In an embodiment, step S30: updating the filling content to the context state carrier, and filling the placeholders in the report template based on the context state carrier to obtain a final report, including: In step S304, the report title is determined based on the independent starting task, and the report title is written into the context state carrier to obtain a target carrier.

[0069] The dependency-free initial task refers to an initial generation task that is not dependent on other chapters. Specifically, a task scheduling algorithm can be used to identify a title generation task without a preceding dependency to implement this feature. This feature is used to explicitly report the starting point of the generated report to avoid circular dependency problems. The context state carrier refers to a data structure that stores the generated content and chapter association relationship. Specifically, a JSON object or XML document can be used to implement this feature. This feature is used to dynamically maintain the context transmission relationship between chapters. The target carrier refers to an iteratively updated context storage structure. Specifically, a key-value pair form can be used to store chapter content and association relationship to implement this feature. This feature is used to ensure that the latest context state is inherited when filling the chapter.

[0070] In this embodiment, a task scheduling algorithm is used to identify a dependency-free title generation task as an initial processing node. The generated title content is written to the initialized context state carrier to form a target carrier. Based on the chapter list generated by topological sorting, each chapter is processed in order. The filling content of the current chapter is merged and updated with the target carrier each time to form a new target carrier containing the latest context state.

[0071] Step S305, traverse the chapter list, update the target carrier through the filling content of each chapter, and fill the updated target carrier into the placeholder of the report template to obtain the final report.

[0072] The chapter list refers to a processing queue sorted by dependency relationship. Topological sorting algorithm can be used to generate chapter execution order to implement this feature. This feature is used to ensure that dependent chapters are processed before their preceding chapters.

[0073] During the traversal process, the context information of the preceding chapter is obtained based on the updated target carrier each time a chapter is processed, ensuring the logical coherence between chapters. Finally, the complete target carrier content is matched and filled into the placeholder of the report template to realize the dynamic fusion of context state and template structure.

[0074] Compared with the traditional method relying on artificial maintenance of chapter dependency relationship or filling the template with fixed order, the complex dependency scene cannot be dynamically handled, resulting in missing correlation or logical contradiction of chapter content. The existing template tool lacks context inheritance mechanism when filling the placeholder, and needs to manually adjust the content order repeatedly. The embodiment automatically maintains the chapter dependency relationship through the dynamically updated context state carrier, and inherits the context information of the previous chapter in real time during the filling process, eliminating the need for manual intervention. Thus, the embodiment solves the context break problem caused by improper handling of dependency relationship in the traditional method, and ensures the logical coherence between chapters by dynamically maintaining the context state. Topological sorting is used to generate a chapter list to avoid processing errors caused by circular dependencies, and the target carrier is updated iteratively to realize real-time inheritance of context information. In the financial analysis report generation scene, this method can automatically handle cross-chapter data reference problems, such as referencing the calculation results of the market capacity chapter in the industry trend chapter, without the need for manual data transfer to achieve content consistency.

[0075] In an embodiment, step S30: updating the filled content to the context state carrier, and filling the placeholders in the report template based on the context state carrier to obtain a final report, comprises: Step S306: inputting the placeholder instruction into the business database to obtain business data.

[0076] The multi-modal placeholder refers to an identifier in the template for marking the insertion position of non-text content, which can be implemented by a combination of symbols with type identifiers, for example, setting a placeholder containing a chart type identifier in the template. This feature enables the template to structurally define the insertion position of multimedia content.

[0077] Step S307: inputting the business data and the placeholder type corresponding to the business data into a preset generator to obtain multi-modal content.

[0078] The business database refers to a storage system for storing structured business data, which can be implemented by a relational database or a time series database. This feature provides raw data support for multi-modal content generation. The preset generator refers to a processing module that calls corresponding generation algorithms according to data types. It can be implemented by an integrated framework containing a visualization engine and a streaming media rendering engine. The generator can realize adaptive processing of different types of data. In this embodiment, the generator includes but is not limited to an icon generation API (Application Programming Interface), a code generation model, and / or an image generation model, etc., which is used to generate corresponding multi-modal content such as charts, code snippets, and / or images according to business data and placeholder types.

[0079] In this embodiment, when a placeholder with chart identification in the template is detected, the corresponding sales data set is extracted from the business database, and the data set and chart type are jointly input into the generator. The generator calls the visualization algorithm according to the chart type to generate a column chart and convert it into a vector graphics format. For audio and video placeholders, after obtaining the original media file from the business database, the generator calls the codec for format conversion and generates a stream media segment that can be embedded in the report.

[0080] In step S308, the multi-modal content is filled into the placeholders of the report template based on the context state carrier to obtain a final report.

[0081] During the filling process, the context state carrier records the statistical indicators of the generated chapters, so that subsequent chart generation can automatically associate the previous data, for example, automatically referencing the previously calculated growth rate indicators when generating market trend charts.

[0082] Compared with the traditional report generation method which only supports text placeholder processing and cannot automatically identify and generate multi-modal content such as charts, audio and video, manual data visualization operations and media file insertion are required. The present embodiment defines multi-modal placeholder types through structuring, establishes an automatic link between data acquisition and content generation, for example, automatically triggers media rendering processes when a video placeholder is detected, which saves time and cost of format conversion and content arrangement compared with manual operation. Thus, the present embodiment realizes the automatic generation and intelligent embedding of multi-modal content and solves the technical limitations of traditional template tools that only support single text format. For example, in an industry analysis report, the system can automatically generate a dynamically updated market trend chart and maintain data consistency with the text analysis content, avoiding data bias that may be caused by manual drawing. At the same time, the context state carrier maintains the logical association between cross-modal content, ensuring that the generated statistical charts can accurately reference the key indicator values calculated in the previous text, and improving the overall coherence of the report content.

[0083] For the purpose of facilitating the understanding of the implementation process of the report generation method based on large models obtained after the above-mentioned embodiment one, please refer to Figure 4 , Figure 4 A schematic diagram of the brief process of the report generation method based on large models is provided, specifically: 1. Task initialization and template analysis: The scheduling module first receives the initial user requirements and the parsed template task dependency graph (such as the dependency relationship of a document report, which can be defined as: [generate title] -> [generate keywords, generate abstract] -> [generate outline] -> [traverse outline to generate chapter text]). This provides the basis for subsequent processes. Then load the target template file, deeply analyze its logical structure (including chapter level, title system and placeholder distribution), accurately identify placeholders such as {{data_result}} and {{analysis_section}}, and build a structured writing element system to clearly define element types, relationships and dependency order, providing a structured framework for content filling.

[0084] 2. RAG knowledge base construction: Collect public knowledge materials such as industry white papers and policy documents. After text cleaning by NLP preprocessing, use pre-trained language models such as SBERT to generate vector embeddings, and store them in vector databases such as FAISS or Pinecone to form an efficient vector index library. At the same time, set metadata fields such as document source and publication date to support semantic-based precise search functions and provide knowledge support for report generation.

[0085] 3. Business data extraction: Use NLP technology to automatically convert user requirements into SQL statements. For example, you can use fine-tuned T5 or BERT Transformer models to perform intent recognition and structured analysis on user requirements, generate SQL statements that can be seamlessly executed, and accurately extract target business data to provide data for data modules in the report.

[0086] 4. Intelligent generation in stages: The scheduling module calls specialized large models to fill content based on task dependency graphs and dynamically updated JSON format context state carriers (ContextPayload). The context state carrier continuously aggregates generated content during task transfer as the core data link. The scheduling module first executes the initial node task (such as "generate title") without dependencies, encapsulates user requirements as prompt words to call specified models to generate results, and writes {"title":"About XXX Report"} and other content to the state carrier. Then, according to the dependency relationship, inject the generated content into the prompt word template of subsequent tasks such as "generate abstract" and "generate keywords", and initiate new model calls in parallel or series. In the text generation stage, the scheduling module traverses the chapter list output by "generate outline", integrates all historical information (title, abstract, business data, etc.) in the state carrier and RAG knowledge base retrieval results for each chapter, and calls specialized large models as needed to ensure the consistency and logical coherence of each module content through API requests or local model loading.

[0087] 5. Compliance and readability optimization: The system automatically detects sensitive information in the content, such as email, ID number, and applies mask processing to complete desensitization; at the same time, language polishing is carried out to optimize the fluency and accuracy of the sentence expression, and finally output the standardized report that meets the format specification, content professional and compliance, and complete the whole process generation closed loop.

[0088] It should be noted that the above examples are only for understanding the present application and do not constitute a limitation on the large model-based report generation method of the present application. Further simple transformations based on this technical concept are within the scope of protection of the present application.

[0089] The present application also provides a large model-based report generation device, please refer to Figure 5 , the large model-based report generation device comprises: The recognition module 10 is used to identify the logical structure, placeholder and task dependency relationship in the preset report template, and initialize the context state carrier based on the logical structure, the placeholder and the task dependency relationship; The generation module 20 is used to obtain the report generation requirement, and convert the report generation requirement into SQL query data, and generate prompt words based on the SQL query data; The filling module 30 is used to input the prompt words into the preset generation model to obtain the filling content, update the filling content to the context state carrier, and fill the placeholder in the report template based on the context state carrier to obtain the final report.

[0090] Optionally, the step of identifying the logical structure, placeholder and task dependency relationship in the preset report template comprises: The chapters and titles in the report template are parsed to obtain the logical structure in the report template, and the placeholders in the report template are parsed; The task dependency relationship is determined based on the chapters and the titles, wherein the task dependency relationship is used to represent the context dependency between the chapters.

[0091] Optionally, the device further comprises a storage module, which is used to: Text cleaning is performed on industry white papers and policy documents to obtain cleaned text, and a pre-trained language model is used to generate text vector embedding based on the cleaned text, and the text vector embedding is stored in a vector database; The generation module 20 is used to: Find the target text vector embedding in the vector database based on the report generation requirement, and generate prompt words based on the target text vector embedding and the SQL query data.

[0092] Optionally, the storage module is used to: generate a plurality of language respective language vector embeddings, and store the language vector embeddings into the vector database; The generating module 20 is configured to: find a target text vector embedding in the vector database based on the report generation requirement, and find a target language vector embedding in the vector database based on a target language of the report generation requirement; generate a prompt word based on the target text vector embedding, the target language vector embedding, and the SQL query data.

[0093] Optionally, the generation model comprises a data analysis model and a text generation model; and the filling module 30 is configured to: input a numerical prompt word in the prompt word into the data analysis model to obtain a statistical calculation result; input a text prompt word in the prompt word into the text generation model to obtain an analysis text; generate filling content based on the statistical calculation result and the analysis text.

[0094] Optionally, the filling module 30 is configured to: determine a report title based on an independent starting task, and write the report title into the context state carrier to obtain a target carrier; traverse a chapter list, update the target carrier by filling content corresponding to each chapter, and fill the updated target carrier into a placeholder of the report template to obtain a final report.

[0095] Optionally, the filling module 30 is configured to: input a placeholder instruction into a business database to obtain business data; input the business data and a placeholder type corresponding to the business data into a preset generator to obtain a multi-modal content; fill the multi-modal content into the placeholder of the report template based on the context state carrier to obtain a final report.

[0096] Optionally, the apparatus further comprises a desensitization module configured to: input the final report into a sensitive information detection module to obtain a sensitive content detection result; if the sensitive content detection result indicates that there is sensitive content in the final report, perform desensitization processing on the final report by using a regular expression to obtain a desensitized report.

[0097] The report generation device based on a large model provided in the application adopts the report generation method based on a large model in the above embodiments, and can solve the technical problem of low efficiency of the current report generation method. Compared with the prior art, the beneficial effects of the report generation device based on a large model provided in the application are the same as those of the report generation method based on a large model provided in the above embodiments, and other technical features of the report generation device based on a large model are the same as those disclosed in the above embodiments, and will not be repeated here.

[0098] The application provides an electronic device, which comprises at least one processor and a memory connected with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the report generation method based on a large model in the above embodiment one.

[0099] Reference will be made to the accompanying drawings Figure 6 which shows a structural schematic diagram of an electronic device suitable for implementing the embodiments of the application. The electronic device in the embodiments of the application can include but is not limited to mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (Personal Digital Assistant), PADs (Portable Application Description), PMPs (Portable Media Player), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), and the like, and fixed terminals such as digital TVs, desktop computers, and the like. Figure 6 The electronic device shown is only an example, and should not bring any limitation to the functions and use range of the embodiments of the application.

[0100] As Figure 6As shown, the electronic device can include a processing device 1001 (e.g., a central processor, a graphics processor, etc.) that can perform various appropriate actions and processes according to programs stored in a read-only memory 1002 or loaded from a storage device 1003 into a random access memory 1004. Various programs and data required for operation of the electronic device are also stored in the random access memory 1004. The processing device 1001, the read-only memory 1002, and the random access memory 1004 are connected to each other through a bus 1005. An input / output interface 1006 is also connected to the bus. Generally, the following systems can be connected to the input / output interface 1006: input devices 1007 including, for example, a touch screen, a touch pad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; the storage device 1003 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 1009. The communication device 1009 can allow the electronic device to communicate wirelessly or wired with other devices to exchange data. Although the electronic device with various systems is shown in the figure, it should be understood that all the shown systems are not required to be implemented or possessed. More or less systems can be alternatively implemented or possessed.

[0101] In particular, according to the embodiments disclosed in the present application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program codes for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network through the communication device, or installed from the storage device 1003, or installed from the read-only memory 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the methods of the embodiments disclosed in the present application are performed.

[0102] The electronic device provided by the present application adopts the report generation method based on large models in the above embodiments, which can solve the technical problem of low efficiency of the current report generation method. Compared with the prior art, the electronic device provided by the present application has the same beneficial effects as the report generation method based on large models provided by the above embodiments, and other technical features in the electronic device are the same as the features disclosed in the previous embodiment method, which will not be repeated here.

[0103] It should be understood that parts of the present application can be realized by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0104] The above merely provides a specific implementation of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0105] The present application provides a computer readable storage medium having stored thereon computer readable program instructions (i.e., computer programs) for performing the large model based report generation method in the above embodiments.

[0106] The computer readable storage medium provided by the present application may, for example, be a U disk, but is not limited to an electric, magnetic, optical, electromagnetic, infrared, or semiconductor system or device, or any combination of the above. More specific examples of the computer readable storage medium can include, but are not limited to, an electric connection with one or more conductive wires, a portable computer disk, a hard disk, a random access memory (RAM), a read only memory (ROM), an erasable programmable read only memory (EPROM or flash memory), an optical fiber, a portable compact disk read only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present embodiment, the computer readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system or device. The program code contained on the computer readable storage medium can be transmitted by any suitable medium, including but not limited to an electric wire, an optical cable, an RF (Radio Frequency), etc., or any suitable combination of the above.

[0107] The above computer readable storage medium can be contained in an electronic device; or can exist separately without being assembled into an electronic device.

[0108] The above computer readable storage medium carries one or more programs, which, when executed by an electronic device, cause the electronic device to implement each embodiment of the large model based report generation method as described above.

[0109] Computer program code for carrying out operations of the present application can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0110] The flow diagrams and the block diagrams in the drawings are meant as methodological and functional description of implementations of systems, methods, and computer program products according to various embodiments of the present application. In this regard, each block in the flow diagrams or block diagrams can represent a module, a segment, or a portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may

[0111] The modules involved in the embodiments of the present application can be implemented in software or hardware. In some cases, the names of the modules do not constitute a limitation on the modules themselves.

[0112] The readable storage medium provided by the present application is a computer readable storage medium, which stores computer readable program instructions (i.e., computer programs) for executing the above-mentioned large model-based report generation method, and can solve the technical problem of low efficiency of the current report generation method. Compared with the prior art, the computer readable storage medium provided by the present application has the same beneficial effects as the large model-based report generation method provided by the above-mentioned embodiments, and will not be described here.

[0113] The application further provides a computer program product comprising a computer program which, when executed by a processor, implements the steps of the large model-based report generation method as described above.

[0114] The computer program product provided by the application can solve the technical problem of low efficiency of the current report generation method. Compared with the prior art, the beneficial effects of the computer program product provided by the application are the same as those of the large model-based report generation method provided by the above-mentioned embodiments, and are not described here.

[0115] The above only describes some embodiments of the application, and does not limit the patent scope of the application. Any equivalent structural transformation, direct / indirect application in other related technical fields, or direct / indirect application in other related technical fields based on the technical concept of the application and the content of the specification and drawings are included in the patent protection scope of the application.

Claims

1. A report generation method based on a large model, characterized in that, The method includes the following steps: Identify the logical structure, placeholders, and task dependencies in the preset report template, and initialize the context state carrier based on the logical structure, placeholders, and task dependencies; Obtain the report generation requirements, convert the report generation requirements into SQL query data, and generate prompt words based on the SQL query data; The prompt words are input into a preset generation model to obtain fill content. The fill content is then updated to the context state carrier, and the placeholders in the report template are filled based on the context state carrier to obtain the final report.

2. The report generation method based on a large model as described in claim 1, characterized in that, The steps for identifying the logical structure, placeholders, and task dependencies in the preset report template include: The logical structure of the report template is obtained by parsing the chapters and titles in the report template, and the placeholders in the report template are also parsed. Task dependencies are determined based on the chapters and titles, wherein the task dependencies are used to characterize the contextual dependencies between the chapters.

3. The report generation method based on a large model as described in claim 1, characterized in that, Before the step of generating prompt words based on the SQL query data, the method further includes: The cleaned text is obtained by cleaning industry white papers and policy documents. A pre-trained language model is used to generate text vector embeddings based on the cleaned text, and the text vector embeddings are stored in a vector database. The step of generating prompt words based on the SQL query data includes: Based on the report generation requirements, the target text vector embedding is searched in the vector database, and prompt words are generated based on the target text vector embedding and the SQL query data.

4. The report generation method based on a large model as described in claim 3, characterized in that, The step of generating prompt words based on the SQL query data includes: Generate language vector embeddings for each of the multiple languages, and store each language vector embedding in the vector database; The step of searching for the target text vector embedding in the vector database based on the report generation requirements, and generating prompt words based on the target text vector embedding and the SQL query data, includes: Based on the report generation requirements, the target text vector embedding is searched in the vector database; based on the target language of the report generation requirements, the target language vector embedding is searched in the vector database. Prompt words are generated based on the target text vector embedding, the target language vector embedding, and the SQL query data.

5. The report generation method based on a large model as described in claim 1, characterized in that, The generative model includes a data analysis model and a text generation model; The step of inputting the prompt words into a preset generation model to obtain the filled content includes: Input the numerical prompts from the prompts into the data analysis model to obtain statistical calculation results; Input the text-type prompts from the prompts into the text generation model to obtain the analyzed text; The populated content is generated based on the statistical calculation results and the analyzed text.

6. The report generation method based on a large model as described in claim 1, characterized in that, The step of updating the populated content to the context state carrier and populating the placeholders in the report template based on the context state carrier to obtain the final report includes: The report title is determined based on the dependency-free starting task, and the report title is written into the context state carrier to obtain the target carrier; The chapter list is traversed, the target carrier is updated with the corresponding fill content for each chapter, and the updated target carrier is filled into the placeholder of the report template to obtain the final report.

7. The report generation method based on a large model as described in claim 1, characterized in that, The step of updating the populated content to the context state carrier and populating the placeholders in the report template based on the context state carrier to obtain the final report includes: Input the placeholder command into the business database to obtain business data; Input the business data and the corresponding placeholder type into a preset generator to obtain multimodal content; The final report is obtained by filling the placeholders in the report template with the multimodal content based on the context state carrier.

8. The report generation method based on a large model as described in any one of claims 1 to 7, characterized in that, After the steps of updating the populated content to the context state carrier and populating the placeholders in the report template based on the context state carrier to obtain the final report, the method further includes: The final report is input into the sensitive information detection module to obtain the sensitive content detection results; If the sensitive content detection results indicate that there is sensitive content in the final report, then the final report is de-sensitized using regular expressions to obtain a de-sensitized report.

9. An electronic device, characterized in that, The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the report generation method based on a large model as described in any one of claims 1 to 8.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the report generation method based on a large model as described in any one of claims 1 to 8.

Images