LLM-based customized report batch generation method and model, and readable storage medium
Through LLM -based customized reporting methods, the problems of low -generation efficiency, poor flexibility and unstable quality in the existing technology have been solved, and high -efficiency, flexible and high -quality report generation has been realized.
Patent Information
- Application Number
- CN202411362780.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-27
- Publication Date
- 2025-01-28
AI Technical Summary
The prior art has problems of inefficiency, poor flexibility and unstable quality in report generation, especially in handling a large number of similar reports and scenarios requiring rapid response.
Adopt a customized reporting method based on LLM, understand and refine the text content through large language models, obtain the feature rules of the report, and automatically generate reports based on the uploaded data content.
The efficiency of report writing has been improved, the cost and time cost of manpower are reduced, the high quality and consistency of the report are ensured, and the flexibility of the needs of different reporting needs.
Smart Images
Figure SMS_1 
Figure SMS_2
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of information data processing, and specifically relates to a customized report batch generation method, model and readable storage medium based on LLM. Background Art
[0002] There are many ways to conduct research now. Among them, writing reports based on census data is an effective means to obtain basic information for different industries and fields. Taking land survey as an example, writing regional data analysis reports after land survey and land census is a systematic work involving data collection, processing, analysis and interpretation. First, it is necessary to determine the purpose of the report, clarify the target audience and the problems to be solved, and then collect relevant information, including original data, planning documents, laws and regulations, etc. Form a professional team, including GIS experts, statistical analysts, land management experts, etc. In the data collation and analysis stage, the data is cleaned and classified, and statistical analysis and GIS analysis methods are used to deeply analyze the data to reveal the current situation, trends and problems of land use. When writing a report, first design the cover, list the table of contents, and then write an abstract to briefly describe the main content and conclusions. Next, introduce the background, purpose, scope and method of the survey, and explain the data source, processing method, analysis tools and technical route in detail. The result analysis part includes current situation description, problem identification and trend prediction, and finally summarizes the main findings and puts forward suggestions for improving land management. After the report is completed, it will be reviewed internally and reviewed by experts, and modified and improved according to the opinions. Finally, adjust the format, submit the report, and publish and disseminate it if permitted. Throughout the entire process, a rigorous scientific attitude and good professional skills must be maintained to ensure the objectivity, accuracy and practicality of the report.
[0003] Currently, report generation relies on manual writing or simple template-based auto-filling technology. These methods have the following major disadvantages: 1. Efficiency issues: Manual writing is time-consuming and costly, especially when dealing with a large number of similar reports.
[0004] 2. Lack of flexibility: Template-based methods have poor flexibility when dealing with non-standardized data or changing requirements, and are difficult to adapt to different reporting needs.
[0005] 3. Uneven quality: Auto-fill technology may result in inconsistent report quality due to inconsistent data or inaccurate templates.
[0006] These shortcomings limit the breadth and effectiveness of existing technologies in practical applications, especially in scenarios that require fast response and high-quality report output. Summary of the invention
[0007] In order to solve the problems existing in the prior art, the present invention provides a customized report batch generation method, model and readable storage medium based on LLM, the purpose of which is to form an automatic writing method for existing research reports with standardization requirements and high content homogeneity. The existing LLM language large model can be used to better understand the text content and extract keywords, so as to obtain the characteristic rules of the report, and automatically and efficiently write a standard report according to the uploaded data content.
[0008] The technical solution adopted by the present invention is: In a first aspect, the present invention provides a method for batch generating customized reports based on LLM, which is used to automatically fill in and generate data for reports with standardized requirements. The specific steps are as follows: First, determine the type and content requirements of the report, and obtain the standard template of similar reports as a reference. Analyze the content of the standard template through the language model to obtain the requirements of the initial data and the writing information required for report generation; Then, the initial data is collected and organized according to the requirements, and a report is written using the acquired initial data and writing information for manual verification. After the verification is completed, the final report text is obtained.
[0009] In combination with the first aspect, the present invention provides a first implementation of the first aspect, wherein the writing information includes a writing framework and a text library obtained by retrieving relevant data after analyzing the standard template; Among them, the writing framework includes the outline, title content and content order; the text library includes a set of conclusions and a set of inferences.
[0010] In combination with the first implementation of the first aspect, the present invention provides a second implementation of the first aspect, wherein the initial data includes current data and previous year-on-year data.
[0011] In combination with the second implementation of the first aspect, the present invention provides a third implementation of the first aspect, wherein the report text includes a data table and a report body, wherein the data table is formed by a large language model after processing the initial data, and the report body is formed by the large language model by selecting content in a text library in a writing framework through the data in the data table and then filling it in.
[0012] In combination with the second implementation of the first aspect, the present invention provides a fourth implementation of the first aspect. When analyzing a standard template through a large language model, the content of the standard template is first split into a title and content, and a writing framework is formed with the title and the corresponding order of the title. Then, the content is clustered into format content and analysis content, and the format content is filled into the writing framework. After that, the analysis content is processed to form a conclusion set and an inference set of the text library.
[0013] In combination with the fourth implementation of the first aspect, the present invention provides a fifth implementation of the first aspect, when processing the analysis content, first classify it through semantic and context understanding to form conclusions and inferences, then classify several conclusions and split them into several phrases about the conclusions, and use the public database to perform relevance search using the phrases as keywords to expand the content of the conclusions and integrate them to form a conclusion set; The inferences are then classified to determine the substantial content of increase, no change and decrease according to the conclusions obtained from the initial data, and then they are split into several phrases as keywords for relevance search. The inference content in the three categories of increase, no change and decrease is then expanded and integrated to form an inference set.
[0014] In combination with the fifth implementation of the first aspect, the present invention provides a sixth implementation of the first aspect, and the specific method for determining the conclusion and inference in the standard template during the processing of the analysis content is as follows: First, the language model extracts all the conclusive contents from the standard template through semantic and contextual understanding to form a conclusion set. Then, for each conclusion in the conclusion set, the related data in the preceding or succeeding content is found and associated, and then merged into the conclusion set. Then a judgment is made. If the conclusion is used to describe the content and direct changes of related data, it is judged to be a final conclusion. If the conclusion is used to describe the changes of non-related data and does not contain a description of data changes, it is judged to be an inference.
[0015] In combination with the fourth implementation of the first aspect, the present invention provides a seventh implementation of the first aspect, when filling the format content into the writing framework, the variable content including the region, time, and name is formed into a variable type description, when writing the report, the variable filling phrase is formed according to the corresponding variable in the initial data, and when writing the report content, the corresponding type of phrase is selected from the filling phrase according to the variable type description in each part of the content. The steps for filling in the body of the report are as follows: First, analyze the content of the data table to obtain the subject content and scope of the report to be written, and adjust the writing framework according to the scope to form the title of the report text; Then, according to the title, the corresponding data is selected from the data table for analysis. According to the content and direct changes of the data, the corresponding phrases are matched in the conclusion set to form complete content to fill in the title. Then, according to the three types of data determined in the conclusion, namely increase, unchanged and decrease, the relevant phrases of the title in the inference set are matched, and the language macro model is used to retrieve the reasons for the corresponding changes in the data corresponding to the title based on the relevant phrases. According to the reasons, the content with high correlation in the relevant phrases is retained, and several complete inferences are formed by the language macro model. The inference with the highest correlation is selected as the display content to be filled in the electronic version of the report text, and the inference is formed into a hyperlink in the electronic version of the report text that can be clicked to pop up a secondary menu. Other related inferences are arranged in the secondary menu in order of relevance, and are selected and confirmed during manual verification.
[0016] In a second aspect, the present invention provides a model, a report generation model established by any one of the above-mentioned customized report batch generation methods based on LLM.
[0017] In a third aspect, the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements any of the methods described above.
[0018] The beneficial effects of the present invention are: The present invention can obtain the report content of the standard template, understand and process it through a large language model, and summarize the writing rules by itself, thereby forming an automatic writing and generation process for this type of report. The report can be automatically generated by simply inputting the required data content, which improves the writing efficiency of such standardized reports and avoids the problem of low sensitivity to data in manual analysis. The automatic generation of reports significantly reduces labor costs and time costs, and uses professional rule extraction and quality control to ensure that the generated reports maintain both high quality and consistency. DETAILED DESCRIPTION
[0019] The present invention will be further explained below in conjunction with specific embodiments.
[0020] Embodiment 1: It should be noted that this embodiment provides a customized report batch generation method based on LLM, which is used to fill in data and generate reports with standardization requirements. The so-called standardization requirements mean that such reports have a unified framework structure and the same evaluation system. For the same type of data in different regions, time or objects, the same data can be refined and compared with the standard value to obtain the situation at the current stage, and the changes can be obtained by comparing with previous data, so as to obtain conclusions.
[0021] The report in this embodiment is a comprehensive survey report, which is a census survey report with the same standard requirements provided to superiors or designated objects by analyzing the data in the corresponding management area obtained by departments in different administrative regions. The data is based on the existing census data or survey data screened according to needs and can be directly used for processing. The existing processing method is that the corresponding person in charge of the department in the corresponding administrative region writes the report according to the data and then uploads the report.
[0022] However, this method is not only inefficient, but the quality of reports written by authors in different regions varies greatly. Since there is a lot of content and its standardization is inconsistent, it will take a lot of time to modify and feedback the reports in the collection, compilation and screening stages of the same type.
[0023] The method in this embodiment aims to utilize the relatively outstanding content understanding and analysis capabilities of the existing LLM language model to assist people in improving report writing efficiency, and can provide corresponding writing generation capabilities for different types of reports to form a standardized and efficient writing model.
[0024] Specifically, the method in this embodiment is as follows: First, it is necessary to determine the type of report to be written. In this embodiment, taking the land and resources survey report as an example, a standard template for this type of report is obtained. The so-called standard template is a standardized model document written according to the writing requirements of this type of report. The content is substantial and can be written by the report evaluation agency or expert group in actual operation.
[0025] The obtained standard template is input into the LLM language model for learning and analysis, and the writing rules of this type of report are obtained through the content of the standard template. At the same time, multiple standard templates of the same type can be input as training materials to obtain more sample content and optimize the obtained writing rules.
[0026] When learning and analyzing the standard template, the LLM language model will also split it into a text library based on its actual content. This text library is a corpus containing corresponding conclusive sentences. According to the data, the LLM language model can form more corresponding discussion content based on this corpus, which has higher accuracy than a larger range of commonly used corpora.
[0027] Then, the LLM language model extracts the writing rules after learning and analyzing several standard templates of the same type. The user then inputs the initial data of the corresponding area, and the LLM language model analyzes the initial data to first obtain the region, time and object corresponding to the report generated this time. The initial data is then processed to form a data table. The contents of the data table are then used to form the text of the report in sequence according to the writing rules. The data is described in accordance with the conclusive language in the corpus and filled into the text content. Finally, the writing information of the report is generated. After manual verification, the writing information is output as a qualified report text.
[0028] Furthermore, the corpus contains the conclusions extracted from the standard templates through the LLM language model.
[0029] That is, when processing the analysis content, first classify it through semantic and context understanding to form conclusions and inferences, then classify several conclusions and split them into several phrases about the conclusions, and use the public database to search for relevance using the phrases as keywords to expand the content of the conclusions and integrate them into a conclusion set; The inferences are then classified to determine the substantial content of increase, no change and decrease according to the conclusions obtained from the initial data, and then they are split into several phrases as keywords for relevance search. The inference content in the three categories of increase, no change and decrease is then expanded and integrated to form an inference set.
[0030] The specific method for judging the conclusions and inferences in the standard template during the processing of the analysis content is as follows: First, the language model extracts all the conclusive contents from the standard template through semantic and contextual understanding to form a conclusion set. Then, for each conclusion in the conclusion set, the related data in the preceding or succeeding content is found and associated, and then merged into the conclusion set. Then, if the conclusion is used to describe the content and direct changes of related data, it is judged as a conclusion. If the conclusion is used to describe the changes of non-related data and does not contain a description of data changes, it is judged as an inference. The obtained conclusions and inferences are then split into interrelated phrases by the language model to form a conclusion set and an inference set.
[0031] It should be noted that the so-called conclusions and inferences refer to the content with clear results obtained by analyzing the initial data to be collected for the report. For example, for the analysis of tourism conditions in a certain area, the number of tourists in the same period this year has decreased compared with the same period last year, and how much it has decreased. This is a conclusion, that is, a conclusion formed based on a clear comparison of related data. However, to analyze why it has decreased, the standard template may be analyzed based on other data. For example, because the average temperature in the same period this year is higher than the average temperature in the same period last year, and the continuous high temperature warning for many days has led to a decrease in the number of tourists in the area with urban tourism resources as the main factor, this weather data will also appear in the standard template, but it will not be associated with the conclusion, so it can be understood as an inference through this method. Similarly, if there were major events or adjustments to the plan in the region during the same period this year, which were not reflected in the initial data in the form of data, it is also an inference formed based on other information.
[0032] In the prior art, conclusions and inferences are usually made through semantic analysis and context, but the accuracy is not high, especially when it comes to analysis content. This kind of description has no standard format and cannot be distinguished by large language models unless the language such as "analysis, guess, and deduction" is used.
[0033] Furthermore, the method is optimized and the land survey report is taken as the writing object.
[0034] First, for the land survey reports to be collected, the assessment unit organizes experts to write several reports with standard templates. The standard templates are written in accordance with unified format requirements and specifications, and at the same time contain as many different regional types and different conclusions as possible, so as to obtain enough samples to train and obtain the writing standards.
[0035] Then, the LLM language model obtains the writing framework through the standard template and the text library obtained by retrieving relevant data after analyzing the standard template, where the writing framework includes the outline, title content and content order; the text library includes the conclusion set and the inference set.
[0036] Among them, when analyzing the standard template through the language big model, the content of the standard template is first split into title and content, and a writing framework is formed with the title and the corresponding order of the title. Then the content is clustered into format content and analysis content, and the format content is filled into the writing framework. Finally, the analysis content is processed to form the conclusion set and inference set of the text library.
[0037] When processing the analysis content, it is first classified into conclusions and inferences through semantic and contextual understanding, and then several conclusions are classified and split into several phrases about the conclusions. The public database is used to perform relevance search using the phrases as keywords, and the content of the conclusions is expanded and integrated to form a conclusion set; then the inferences are classified, and after the substantive content is classified into increase, no change and decrease, they are split into several phrases as keywords for relevance search, and then the inference content in the three categories of increase, no change and decrease is expanded and integrated to form an inference set.
[0038] When filling the format content into the writing framework, the variable content including region, time, and name will be formed into a variable type description. When writing the report, variable filling phrases are formed according to the corresponding variables in the initial data. When writing the report content, the corresponding type of phrases are selected from the filling phrases according to the variable type description in each part of the content.
[0039] In the process of writing the report body, the content of the data table is first analyzed to obtain the subject content and writing scope of the report to be written, and the writing framework is adjusted according to the writing scope to form the title of the report body; Then, according to the title, the corresponding data is selected from the data table for analysis. According to the content and direct changes of the data, the corresponding phrases are matched in the conclusion set to form complete content to fill in the title. Then, according to the three types of data determined in the conclusion, namely increase, unchanged and decrease, the relevant phrases of the title in the inference set are matched, and the language macro model is used to retrieve the reasons for the corresponding changes in the data corresponding to the title based on the relevant phrases. According to the reasons, the content with high correlation in the relevant phrases is retained, and several complete inferences are formed by the language macro model. The inference with the highest correlation is selected as the display content to be filled in the electronic version of the report text, and the inference is formed into a hyperlink in the electronic version of the report text that can be clicked to pop up a secondary menu. Other related inferences are arranged in the secondary menu in order of relevance, and are selected and confirmed during manual verification.
[0040] After the writing information is formed through the above steps, the report evaluation unit can publish the report generation model containing the writing information to the user, and the report generation model can call the external LLM language model, and use the cloud platform to realize the login and data interaction of multiple users. It can also be a storable medium, on which a computer program is stored, and when the program is executed by a processor, the method described above is implemented.
[0041] Then, users who need to generate land survey reports for corresponding areas can use the collected initial data for screening, and according to the requirements of the report of this type, screen out the corresponding initial data data package, which contains the current land data and the land data that needs to be compared in the past, and input the initial data package into the model or system.
[0042] The model or system is connected to an external LLM language model to analyze the initial data according to the writing information and generate a data table with standardized table content generated according to the writing framework. The data merging and splitting processes involved form a summary form, which can be directly inserted into the final generated report for direct display.
[0043] Then, the LLM language model determines the framework of the report according to the content of the data table and the determined writing framework, and extracts the statements of the format content and analysis content in the corpus. The format content is the standard statement content that conforms to the corresponding content, that is, the homogeneous content that all reports have, which serves as the beginning, the end, and the middle to connect the past and the future. The analysis content is the content with conclusive statements that needs to be generated based on the data in the data table, which contains the definite conclusion content and the inference content obtained by the analysis, all of which are generated by the LLM language model according to the relevance and matched in the corresponding writing framework position to form paragraphs.
[0044] At the same time, in order to expand and enrich the content, after the extraction from the corpus is completed, keywords will be extracted from the content that has been filled in the writing framework, and a matching search will be performed again in the external database. The content of some related solutions obtained through correlation analysis will be filled in the corresponding part of the report.
[0045] Finally, the report text obtained by the LLM language model according to the above processing flow is output to the requester, and the requester manually proofreads and corrects it. The LLM language model automatically modifies it by marking or inputting commands, and finally obtains the report text and submits it to the assessment unit.
[0046] Required Furthermore, an implementation case is provided for the above content.
[0047] 1. Catalog generation (1) Generate the chapter outline for "Soil Types" Main features: In the "Chapter 4 Soil Resources and Evolution", "I. Soil Types and Evolution", "(I) Soil Types" chapter of the "Technical Report", it is necessary to introduce the soil profile of each layer according to the actual situation of local soil types and soil classification principles and levels. In other words, a tool is needed to generate the outline of this chapter according to the soil type situation; Processing flow: Data preparation: Data processing tools automatically generate soil type tables based on the core data tables; Generate outline: The tool automatically generates the chapter outline according to the soil type table and hierarchy rules; Example of a soil type table: The outline generated by the tool is then shown below.
[0048] Sample outline: 1. Coarse bone soil 1.1 Neutral coarse bone soil subclass 1.1.1 Sandy neutral coarse bone soil 1.1.1.1 Crispy stone soil 2. Brown soil 2.1 Typical brown soil subtypes 2.1.1 Sandy brown soil 2.1.1.1 Ridge sandy soil 2.2 Albic brown soil subtype 2.2.1 Sandy white brown soil 2.2.1.1 Brown and white clay 2.2.1.2 Furnace bottom brown and white soil 2.2.1.3 Patina soil 2.2.1.4 Patina yellow purple clay 2.3 Aquic brown soil subtypes 2.3.1 Muddy sandy brown soil 2.3.1.1 Blackboard soil 3. Purple soil 3.1 Neutral purple soil subclass 3.1.1 Soil Types of Zisha Soil 3.1.1.1 Purple clay 3.1.1.2 Thin layer of purple sand soil 3.1.2 Purple clay soil 3.1.2.1 Purple clay 4. Fluvo-aquic soil 4.1 Typical fluvo-aquic soil subtypes 4.1.1 Calcareous fluvo-alkaline loam soil 4.1.1.1 Compound soil 4.1.2 Calcareous tidal clay soil 4.1.2.1 Silt 4.1.3 Fluvo-sand soil 4.1.3.1 Brown sand soil 4.1.3.2 Leakage of sand and soil 4.1.3.3 Clay-bottomed brown sand 4.1.4 Fluvo-aquic loam soil 4.1.4.1 Brownish yellow soil 4.1.4.2 Black brown soil 4.1.4.3 Bottom black brown soil 4.1.4.4 Sandy brown soil 4.1.4.5 Brownish yellow soil with sandy bottom 4.1.5 Tidal clay soil 4.1.5.1 Old loess 4.1.5.2 Black loess 5. Sand ginger black soil 5.1 Typical sub-types of sandy black soil 5.1.1 Black ginger soil 5.1.1.1 Lake Black Soil (2) Generate the chapter outline of "Planning and technical recommendations for soil resource utilization" Key Features: Some chapters in the report require planning suggestions, but the content is relatively routine and lacks regional characteristics Each district and county should present its own thinking and suggestions should be somewhat different. It is not possible to use a unified template across the country. Processing flow: Template making: Use the big model to help generate several sets of templates with key recommendations; Model selection: The tool randomly selects a specified number of points from the corresponding template library; Outline combination: integrate the selected points into an outline of the content of the proposed part of the plan; Template example: The first-level title is "II. Plans and technical recommendations for soil resource utilization" Second level headings (select 3 randomly from the following template library): Serial number Secondary heading 1 Strengthen the standardization and management of rural land transfer to ensure the effective use of arable land resources. 2 Deepen agricultural supply-side structural reform and promote high-quality development of the agricultural industry. 3 Increase investment in agricultural science and technology research and development, and promote agricultural scientific and technological innovation and transformation of results. 4 Promote green agricultural development model, reduce agricultural non-point source pollution, and protect the ecological environment. 5 Improve the agricultural product quality and safety supervision system to ensure the "food safety" of the people. 6 Strengthen the construction and maintenance of agricultural infrastructure and improve the comprehensive agricultural production capacity. 7 Cultivate a new team of professional farmers and improve farmers' agricultural production skills and management level. 8 Optimize the allocation and utilization of agricultural resources to achieve sustainable agricultural development. 9 Promote the integrated development of the primary, secondary and tertiary industries in rural areas and broaden channels for farmers to increase their income. 10 Establish and improve agricultural risk prevention and response mechanisms to reduce agricultural production risks. 11 Optimize soil resource management and improve agricultural utilization efficiency. 12 Enhance the soil's environmental protection function and build a solid ecological safety barrier. 13 Strengthen the farmland protection system to ensure sustainable use of resources. 14 Accelerate industrial structure adjustment and promote the implementation of agricultural planning. Level 3 headings (4 randomly selected from the corresponding template): Then the outline style example after random combination is as follows, and the outline of the original text is as follows.
[0049] 2. Planning and technical suggestions for soil resource utilization (I) Enhance soil environmental protection function and build a solid ecological security barrier 1. Implement soil ecological restoration and management plans (1) Conduct a soil pollution survey and clarify restoration objectives.
[0050] (2) Promote soil remediation technologies such as biological remediation and physical remediation.
[0051] (3) Strengthen monitoring and evaluation of restoration effects to ensure ecological safety.
[0052] 2. Strengthen soil pollution prevention and control and supervision (1) Establish and improve a legal system for soil pollution prevention and control.
[0053] (2) Strengthen supervision of pollution sources such as industrial and mining enterprises and agricultural non-point sources.
[0054] (3) Promote soil pollution prevention and control technologies and equipment and enhance prevention and control capabilities.
[0055] 3. Promote organic agriculture and ecological agriculture models (1) Encourage farmers to use organic fertilizers instead of chemical fertilizers.
[0056] (2) Promote environmentally friendly agricultural technologies such as biological control and physical control.
[0057] (3) Strengthen the construction of ecological agriculture demonstration areas and play a leading role in demonstration.
[0058] 4. Strengthen publicity and education on soil resource protection (1) Carry out activities to popularize knowledge on soil resource protection.
[0059] (2) Use media, the Internet and other channels to strengthen publicity and education efforts.
[0060] (3) Enhance public awareness and participation in soil resource protection.
[0061] (II) Strengthen the farmland protection system to ensure sustainable use of resources 1. Improve the legal and regulatory system for farmland protection (1) Revise existing laws and regulations on farmland protection and strengthen their binding force.
[0062] (2) Formulate detailed implementation rules for farmland protection and clarify the responsibilities of all parties.
[0063] 2. Establish a responsibility system for farmland protection (1) Clarify the responsibilities of governments at all levels, relevant departments and agricultural producers and operators.
[0064] (2) Establish a mechanism for assessing farmland protection targets and strengthen supervision and inspection.
[0065] 3. Implement a strict land occupation and compensation balance system (1) Strictly control the occupation of arable land by non-agricultural construction and implement the balance plan between occupation and compensation.
[0066] (2) Strengthen the quality evaluation and supervision of supplementary cultivated land to ensure the quality of supplementary cultivated land.
[0067] 4. Strengthen the protection and improvement of cultivated land quality (1) Implement actions to protect and improve arable land quality and improve soil conditions.
[0068] (2) Promote scientific fertilization, crop rotation, fallow and other farmland protection measures.
[0069] (3) Strengthen the monitoring and evaluation of arable land quality and improve soil fertility. (III) Accelerate the adjustment of industrial structure and promote the implementation of agricultural planning 1. Formulate a detailed agricultural industry structure adjustment plan (1) Clarify the goals, directions and key tasks of industrial structure adjustment.
[0070] (2) Develop a specific implementation plan and timetable to ensure that the plan is implemented.
[0071] 2. Strengthen planning guidance and support (1) Introduce a series of plans and measures to support the adjustment of agricultural industrial structure.
[0072] (2) Increase support for new agricultural business entities and encourage innovative development.
[0073] 3. Promote agricultural science and technology innovation and transformation of achievements (1) Increase investment in and support for agricultural science and technology innovation.
[0074] (2) Accelerate the transformation and application of agricultural scientific and technological achievements; (3) Improve agricultural science and technology and promote industrial upgrading and development; 4. Strengthen agricultural infrastructure construction (1) Increase investment in agricultural infrastructure and improve production conditions.
[0075] The original text is as follows: 2. Planning and technical suggestions for soil resource utilization 1. Strengthening the planned institutional guarantee for the protection and management of cultivated land 1. Improve farmland infrastructure and establish drought-resistant and flood-drainage systems 2. Balance soil nutrient levels and improve fertility of cultivated land 3. Strengthen the awareness of farmland protection and create a good farmland environment (II) Accelerate the implementation of the plan for agricultural industrial structure adjustment 1. Promote the industrialization and standardization of agricultural products (1) Implementing standardized system specifications (2) Implementing standardized demonstration and guidance (3) Providing standardized public services (4) Support standardized production and operation 2. Innovate agricultural product production and management models (1) Expand the scale of the industry and drive economic development (2) Carry out “agricultural and social connection” to stimulate the vitality of all parties (3) Realize “agricultural and supermarket docking” to solve sales problems (4) Promote subject integration and achieve benefit sharing 3. Give full play to the government's planning guidance and financial support (1) Cultivate and support industrial development (2) Improve the funding mechanism (3) Expanding the agricultural technical talent team (4) Exploiting and utilizing cultural value 3. Promote integrated technology to improve farmland quality 2. Text generation This example uses three generation methods to illustrate.
[0076] (1) Random template + variable assignment (user input) Applicable type: Routine text paragraph Main features: The text content is relatively routine; it has no regional attributes and is applicable nationwide Processing flow: Template making: Use large models to assist in generating several sets of fixed templates; Variable marking: mark the variables in the template; Variable entry: The user enters the value of the corresponding variable; Content assembly: The tool automatically randomly selects one of the templates, inserts variable values, and forms new text to be assembled into the report.
[0077] (2) Prompt word + text / table material (user input) → large model generation Applicable types: excerpt text paragraphs, partial data description text Main features: The text content mainly comes from various literature materials or online public materials; the materials themselves have strong regional attributes and cannot use a national unified template Processing flow: Information entry: The user enters relevant excerpt text or information table; Model call: The tool calls the large model interface, combines the built-in prompt words and the text / table information entered by the user, and then inputs them into the large model; Data recovery: The big model organizes (abbreviates, summarizes or polishes) the relevant content into new text according to the prompt word requirements and then outputs it back to the tool; Text collation: The tool removes paragraph breaks from the output of large models and organizes them into a whole new paragraph of text waiting for report assembly.
[0078] (3) The tool automatically combines prompt words → large model generation Applicable types: texts with expanded key points and partial data description Main features: For plan suggestion texts, after the key points of the plan suggestion are drawn up, the key points can be expanded using the big model; data description texts are mainly text representations of tabular data, without data analysis, mining, etc. The data can be provided to the big model, and the complete data description text can be generated with the help of the big model.
[0079] Processing flow: Model call: The tool calls the large model interface and inputs the built-in prompt words and selected suggested points / specified data information table into the large model at one time; Data recovery: The large model generates text according to the prompt word requirements and then outputs it back to the tool; Text collation: The tool removes paragraph breaks from the output of large models and collate them into a complete text waiting for report assembly.
[0080] 3. Report assembly Generate paragraph texts in sequence according to the text types of each paragraph in the report; set the title and text format according to the title level; assemble the report with the titles and texts of each level in the specified order; before the complete report is output, the user can adjust each paragraph individually (modify the prompt words to regenerate the text / modify the text directly / use the large model to re-polish the text after modifying it, etc.); after fine-tuning, the user can directly start manual review and revision of the first draft.
[0081] 4. Manual review and revision (1) Insert a chart Insert the processed report into the table and related drawings into the specified location of the report.
[0082] (2) Text content review and optimization Fact checking: Carefully check the facts, data, and quotations in the text to ensure the accuracy of the information. For uncertain information, additional verification should be carried out; Eliminate text bias: Identify and eliminate possible gender, racial, cultural or other types of bias in the text and avoid direct negative descriptions; Streamline redundant expressions: Optimize the lengthy and repetitive content in the text, delete unnecessary summary sentences, and ensure clear and concise expression; Check logical consistency: Check the logical flow and structure of the text to ensure that the content is coherent, the views are consistent, and there are no contradictions; Check style and tone: review language expression, correct grammatical errors, and unify language style; Check whether the writing angle and expression are appropriate: Consider the target audience and writing purpose of the report, check the language style and context of the text to ensure that the text is appropriate and suitable in the context; Compliance and sensitivity review: Ensure that the text content complies with relevant laws, regulations and ethical standards, and avoid sensitive or potentially controversial content.
[0083] (3) Insights and regional characteristics analysis Supplementary viewpoints and insights: Based on the survey data, combined with the latest research dynamics and development trends in the fields of agriculture and soil science at home and abroad, supplement the corresponding chapters and paragraphs of the report with viewpoint summaries to add more professional insights to the report. Including: Trend analysis: Based on census data, analyze the long-term changing trends in soil quality, type, pollution status, etc.
[0084] Problem diagnosis: Identify the main problems in soil health, such as soil degradation, pollution, erosion, etc., analyze their causes, and propose targeted solutions.
[0085] Plan suggestions: Based on the actual local conditions, review and adjust the plan suggestions to ensure that the plan suggestions do not contradict the actual local conditions.
[0086] Regional characteristics analysis: Conduct targeted analysis based on the soil characteristics and agricultural development status of different regions, put forward agricultural development suggestions and soil improvement measures tailored to local conditions, and emphasize regional characteristics and advantages; Soil type and distribution: Analyze the soil types and distribution characteristics of different regions and their impact on agricultural production to provide a basis for regional agricultural planning.
[0087] Soil quality and productivity: Assess the quality of soil in each region, including nutrient content, organic matter level, pH, etc., and analyze its impact on crop yields.
[0088] Soil utilization and protection: Analyze the differences in soil utilization and protection measures in different regions, summarize successful soil management experiences and models, and provide reference for other regions.
[0089] Soil ecology and environment: Explore the relationship between soil and ecological environment, analyze the impact of environmental problems such as soil pollution and degradation on the ecosystem, and propose eco-friendly soil management strategies.
[0090] Analysis of local products: It is necessary to describe in detail the ecological needs and geographical characteristics of local products; evaluate production areas based on climate, soil, water resources and other conditions; analyze the current utilization status and potential of production areas, while considering ecological sustainability, and propose specific strategies for optimizing planting layout and industrial development.
[0091] It should be noted that in several embodiments provided in the present application, it should be understood that the disclosed apparatus and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.
[0092] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units; they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.
[0093] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.
[0094] Those skilled in the art can understand that all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.
[0095] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.
[0096] The present invention is not limited to the above optional implementations, and anyone can derive other various forms of products under the enlightenment of the present invention. The above specific implementations should not be understood as limiting the scope of protection of the present invention. The scope of protection of the present invention should be based on the definition in the claims, and the description can be used to interpret the claims.
Claims
1. A customized report batch generation method based on LLM, which is used to automatically fill in and generate data for reports with standardized requirements, and is characterized by: The specific steps are as follows: First, determine the type and content requirements of the report, and obtain the standard template of similar reports as a reference. Analyze the content of the standard template through the language model to obtain the requirements of the initial data and the writing information required for report generation; Then, the initial data is collected and organized according to the requirements, and a report is written using the acquired initial data and writing information for manual verification. After the verification is completed, the final report text is obtained.
2. The method for batch generating customized reports based on LLM according to claim 1, characterized in that: The writing information includes a writing framework and a text library obtained by retrieving relevant data after analyzing the standard template; Among them, the writing framework includes the outline, title content and content order; the text library includes a set of conclusions and a set of inferences.
3. The method for batch generating customized reports based on LLM according to claim 2, characterized in that: The initial data includes current data and previous year-on-year data.
4. The method for batch generating customized reports based on LLM according to claim 3 is characterized in that: The report text includes a data table and a report body. The data table is formed by processing the language macro model according to initial data, and the report body is formed by filling in the language macro model by processing the content in the text library selected from the data in the writing framework through the data in the data table.
5. The method for batch generating customized reports based on LLM according to claim 3, characterized in that: When analyzing standard templates through the language big model, the content of the standard template is first split into title and content, and a writing framework is formed with the title and the corresponding order of the title. Then the content is clustered into format content and analysis content, and the format content is filled into the writing framework. Finally, the analysis content is processed to form the conclusion set and inference set of the text library.
6. The method for batch generating customized reports based on LLM according to claim 5, characterized in that: When processing the analysis content, first classify it through semantic and context understanding to form conclusions and inferences, then classify several conclusions and split them into several phrases about the conclusions, and use the phrases as keywords to perform relevance searches in public databases to expand the content of the conclusions and integrate them into a conclusion set; The inferences are then classified to determine the substantial content of increase, no change and decrease according to the conclusions obtained from the initial data, and then they are split into several phrases as keywords for relevance search. The inference content in the three categories of increase, no change and decrease is then expanded and integrated to form an inference set.
7. The method for batch generating customized reports based on LLM according to claim 6, characterized in that: The specific method for judging the conclusions and inferences in the standard template during the processing of the analysis content is as follows: First, the language model extracts all the conclusive contents from the standard template through semantic and contextual understanding to form a conclusion set. Then, for each conclusion in the conclusion set, the related data in the preceding or succeeding content is found and associated, and then merged into the conclusion set. Then a judgment is made. If the conclusion is used to describe the content and direct changes of related data, it is judged to be a final conclusion. If the conclusion is used to describe the changes of non-related data and does not contain a description of data changes, it is judged to be an inference.
8. The method for batch generating customized reports based on LLM according to claim 7, characterized in that: When filling the format content into the writing framework, the variable content including the region, time, and name is formed into a variable type description. When writing the report, the variable filling phrase is formed according to the corresponding variable in the initial data. When writing the report content, the corresponding type of phrase is selected from the filling phrase according to the variable type description in each part of the content. The steps for filling in the body of the report are as follows: First, analyze the content of the data table to obtain the subject content and scope of the report to be written, and adjust the writing framework according to the scope to form the title of the report text; Then, according to the title, the corresponding data is selected from the data table for analysis. According to the content and direct changes of the data, the corresponding phrases are matched in the conclusion set to form complete content to fill in the title. Then, according to the three types of data determined in the conclusion, namely increase, unchanged and decrease, the relevant phrases of the title in the inference set are matched, and the language macro model is used to retrieve the reasons for the corresponding changes in the data corresponding to the title based on the relevant phrases. According to the reasons, the content with high correlation in the relevant phrases is retained, and several complete inferences are formed by the language macro model. The inference with the highest correlation is selected as the display content to be filled in the electronic version of the report text, and the inference is formed into a hyperlink in the electronic version of the report text that can be clicked to pop up a secondary menu. Other related inferences are arranged in the secondary menu in order of relevance, and are selected and confirmed during manual verification.
9. A model, characterized in that: A report generation model established by a customized report batch generation method based on LLM as described in any one of claims 1-8.
10. A readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.