Industry report generation optimization method and system based on reinforcement learning and knowledge graph

Abstract

The present invention discloses an industry report generation optimization method and system based on reinforcement learning and knowledge graph, which relates to the technical field of industry report generation, and includes the following steps: first collect industry data from multiple sources and pre-process them, determine the collection frequency according to the type and importance, then construct a knowledge graph, use a variety of technologies to identify entities, etc., then design a reinforcement learning model, define the state, action and other spaces and reward functions, train according to a specific algorithm, and then generate a preliminary report based on user requirements by using the model and graph to select data for analysis, then use an optimization algorithm to combine the language model and graph to improve the report, and finally evaluate the feedback to adjust the parameter strategy and improve the system performance. The present invention can improve the efficiency of industry report generation, optimize the report quality, enhance the depth and accuracy, adapt to different needs and can be self-improving, and effectively help the decision-making and development of various industries.

Description

Technical Field

[0001] The present invention relates to the technical field of industry report generation, and in particular to an industry report generation optimization method and system based on reinforcement learning and knowledge graph. Background Art

[0002] In modern business and economic activities, industry reports are an indispensable basis for corporate decision-making, investment analysis, market research, and policy formulation. With the rapid development of information technology and increasingly fierce market competition, the demand for high-quality, timely and in-depth industry reports has increased dramatically.

[0003] Traditional industry report generation mainly relies on manual work, and professionals need to manually collect relevant data from scattered data sources, such as industry news websites, official corporate releases, professional databases, and various statistical reports. This process is not only time-consuming and laborious, but also difficult to ensure the accuracy, completeness, and consistency of the data due to the wide range of data sources and different formats. For example, when collecting corporate financial data, different companies may adopt different accounting standards and report formats, and manual integration of these data is prone to errors and omissions.

[0004] In the data analysis and report writing stage, manual methods are highly dependent on the knowledge reserves and experience judgment of professionals. However, even experienced analysts find it difficult to fully and deeply explore the hidden connections and trends in the face of massive and complex data. Moreover, the language expression and logical structure of manually written reports vary from person to person, making it difficult to ensure a uniform high-quality standard. At the same time, with the rapid development and changes in the industry, new business models, technological innovations and market dynamics continue to emerge, and traditional methods are difficult to follow up and adapt in a timely manner, resulting in a significant reduction in the timeliness and foresight of reports.

[0005] Although some automated tools and technologies have been applied to industry report generation to a certain extent in recent years, there are still many limitations. Existing data collection tools can often only perform simple crawling of specific data sources, and lack the ability to intelligently integrate and preprocess multi-source heterogeneous data. Most data analysis software can only perform basic statistical analysis and cannot perform in-depth reasoning and prediction in combination with industry knowledge systems. In addition, in the report generation process, there is a lack of effective content optimization and quality assessment mechanisms, and the generated reports are often stiff in language, monotonous in content, and lack logic and readability.

[0006] To sum up, the existing industry report generation method can no longer meet the growing demand for high-quality reports. There is an urgent need for an innovative and intelligent industry report generation optimization method and system to improve the accuracy, timeliness, depth and readability of reports, and provide strong support for the decision-making and development of various industries. Summary of the invention

[0007] The present invention proposes an industry report generation optimization method and system based on reinforcement learning and knowledge graph to solve the problems mentioned in the above-mentioned prior art.

[0008] In order to achieve the above purpose, the present invention adopts the following technical solution: an industry report generation optimization method based on reinforcement learning and knowledge graph, comprising:

[0009] Data collection and preprocessing steps: clean, denoise and format the collected data, and store the processed data in the database; the collection frequency f is dynamically adjusted according to the data type t and importance i, and the collection frequency adjustment formula is: , where k1 and k2 are weight coefficients determined based on experience; regular expressions are used in the data cleaning process, and mean filtering, median filtering or semantic analysis is used for denoising;

[0010] Knowledge graph construction steps: Based on the preprocessed data, natural language processing technology and knowledge extraction algorithms are used to identify entities, relationships and attributes in the data and build industry knowledge graphs. The knowledge extraction algorithm uses a combination of rule-based methods, support vector machines, deep learning model machine learning methods, and semi-supervised learning methods;

[0011] Reinforcement learning model training steps: Design a reinforcement learning model, define the state space, action space, reward function and transition probability. The state space consists of nodes n, relations r and attributes a in the knowledge graph, and the current report p generation progress information, which is expressed in vector form as ,in is a feature encoding function. The action space includes operations such as selecting data, analyzing data, writing report fragments, and adjusting report structure. The reward function R is based on the quality evaluation indicators of the generated report, including accuracy A, completeness C, logic L, and readability R. d , the formula is , where α, β, γ, δ are weight coefficients, and ; Model training uses deep Q network DQN or proximal strategy optimization PPO algorithm;

[0012] Report generation steps: Based on the report topic T and requirements set by the user, the trained reinforcement learning model is used, with the support of the knowledge graph, to automatically select relevant data, perform analysis and reasoning, and generate a preliminary industry report according to the predetermined report template.

[0013] Furthermore, the industry report generation optimization method based on reinforcement learning and knowledge graph also includes:

[0014] Report optimization steps: Use the GPT series of models to optimize the initially generated report, and combine the information in the knowledge graph to perform fact checking and knowledge supplement;

[0015] Evaluation and feedback step: Provide the optimized report to users or experts for evaluation, collect evaluation opinions and feedback information, and adjust the parameters and strategies of the reinforcement learning model based on the feedback;

[0016] In the data collection and preprocessing steps, the data source management submodule is responsible for adding, deleting, updating and setting the priority of the data source. s The calculation is based on the reliability of the data source R s , data update frequency F s and data correlation C s , the calculation formula is , where ω1, ω2, and ω3 are weight coefficients. The data processing submodule adopts a distributed computing framework, including Hadoop and Spark, and records data processing logs during the processing.

[0017] Furthermore, in the knowledge graph construction step, the knowledge extraction submodule supports text processing in multiple languages, and the processing accuracy of texts in different languages ​​is A l Through the test statistics of a large number of samples, the calculation formula is: ,in is the number of entities, relations, and attributes extracted correctly, is the total number of extractions, and the knowledge fusion submodule is used to fuse and integrate knowledge from different sources.

[0018] Furthermore, in the reinforcement learning model training step, the model evaluation submodule uses the cross-validation method to evaluate the model performance and adjusts the model parameters and training strategy according to the evaluation results; the cross-validation accuracy A c The calculation formula is , where k is the number of cross-validation folds, a g is the accuracy of the g-th fold. The model visualization submodule is used to visualize the state changes and reward curve information during the model training process;

[0019] Furthermore, in the report generation step, the template management submodule provides a report template library, supports user-defined templates, and meets the needs of different users and industries. The template matching function Based on topic relevance T 、Industry adaptability I and user needs U The formula for calculating compliance is: , where λ1, λ2, λ3 are weight coefficients;

[0020] The model selects a template based on the topic and requirements when generating a report. The probability of selecting a template is P t The calculation formula is , where S is the template matching function, θ is the temperature parameter, m is the total number of templates, I represents industry-related indicators, which are used to measure the industry adaptability of the template, and U represents user demand-related indicators, which are used to measure the conformity of the template with user needs.

[0021] Furthermore, in the report optimization step, the language evaluation submodule integrates multiple language models and selects appropriate evaluation models according to different languages ​​and application scenarios. The selection probability of the language model is P m Based on language type L t Application scenario A s and model performance P p Calculation, the formula is , where G is the model fitness function and μ is the selection parameter.

[0022] Furthermore, in the evaluation and feedback step, the evaluation management submodule is responsible for organizing the evaluation activities, including the management of evaluators, the assignment of evaluation tasks, the collection of evaluation results, and the weight of the evaluators. p Based on the professional level of the assessor p , Experience E p and assessing historical accuracy H p Calculation, the formula is , where ρ1, ρ2, and ρ3 are weight coefficients.

[0023] Furthermore, a system for implementing an industry report generation optimization method based on reinforcement learning and knowledge graph includes:

[0024] Data collection and preprocessing module: responsible for collecting industry-related data from multiple data sources, and performing preprocessing operations such as cleaning, denoising, and format unification, storing the processed data in the database, and providing data support for subsequent analysis and report generation;

[0025] Knowledge graph construction module: Build industry knowledge graphs based on preprocessed data, use natural language processing technology and knowledge extraction algorithms to identify entities, relationships, and attributes, use graph databases to store and manage knowledge graphs, and achieve structured representation and efficient query of industry knowledge;

[0026] Reinforcement learning model training module: Design and train reinforcement learning models, define state space, action space, reward function and transition probability, and optimize model strategies through trial and error learning in a simulated environment, so that it can generate high-quality industry reports based on user needs;

[0027] Report generation module: Based on the report topics and requirements set by the user, with the support of knowledge graph and trained reinforcement learning model, it automatically generates preliminary industry reports, including data selection, analysis, report segment writing and structure organization;

[0028] Report optimization module: optimizes the initially generated report using optimization algorithms, evaluates and improves the report’s language expression, logical structure, and content coherence based on language models and knowledge graphs to ensure report quality;

[0029] Evaluation and feedback module: collects evaluation opinions and feedback information from users or experts on the optimized reports, and uses them to adjust the reinforcement learning model and optimization algorithm to achieve continuous optimization and performance improvement of the system.

[0030] Furthermore, the data source management submodule in the data acquisition and preprocessing module is responsible for adding, deleting, updating and setting priorities of data sources, and the data processing submodule uses Hadoop or Spark distributed computing frameworks to record data processing logs during the processing process to facilitate tracing and troubleshooting;

[0031] The knowledge extraction submodule in the knowledge graph construction module supports text processing in multiple languages; the knowledge fusion submodule is used to fuse and integrate knowledge from different sources;

[0032] The model evaluation submodule in the reinforcement learning model training module uses methods such as cross-validation to evaluate model performance and adjust model parameters and training strategies based on the evaluation results;

[0033] The template management submodule in the report generation module provides a report template library and supports user-defined templates to meet the needs of different users and industries. The report generation engine submodule realizes automatic report generation based on reinforcement learning models and knowledge graphs, and supports real-time preview and adjustment during the generation process.

[0034] Furthermore, the language evaluation submodule in the report optimization module integrates multiple language models and selects appropriate evaluation models according to different languages ​​and application scenarios; the knowledge verification submodule uses authoritative knowledge in the knowledge graph to verify the report content;

[0035] The assessment management submodule in the assessment and feedback module is responsible for organizing assessment activities, including assessment personnel management, assessment task allocation, and assessment result collection; the feedback processing submodule classifies, analyzes, and transforms feedback information.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] First, in terms of efficiency improvement, the automation and intelligence of data collection have been realized, which can quickly integrate multi-source heterogeneous data and significantly shorten the report generation cycle. For example, an industry report that originally took several days to complete can be generated in a few hours with the help of this system, which greatly improves work efficiency.

[0038] Secondly, the quality of the report has made a qualitative leap. The use of knowledge graphs has made the report content more in-depth and accurate, the reinforcement learning model has optimized the report structure and logic, and the optimization algorithm has improved the language expression, making the report clear and easy to understand, and enhancing its practicality and readability. This helps companies and decision makers to grasp industry trends more accurately, make scientific decisions, and reduce decision-making risks.

[0039] Furthermore, the system has good adaptability and scalability. It can be flexibly customized according to the characteristics of different industries and user needs, easily responding to industry changes and new demands, and continuously optimizing itself with use, always maintaining high-quality report output capabilities, and providing long-term and reliable decision-making support for different fields.

[0040] Finally, by optimizing the report generation process and improving the quality, the labor and time costs are reduced and the overall benefits are improved. Whether it is internal market analysis of enterprises, due diligence of investment institutions, or policy research of government departments, efficient and accurate industry reports can be obtained from this patent to promote the efficient development of various industries. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic block diagram of an industry report generation optimization method based on reinforcement learning and knowledge graph proposed by the present invention;

[0042] Figure 2 This is a schematic block diagram of an industry report generation optimization system based on reinforcement learning and knowledge graph proposed in the present invention. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0045] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0046] Reference Figure 1-2 :A method for optimizing industry report generation based on reinforcement learning and knowledge graph, including:

[0047] Data collection and preprocessing steps: Collect industry-related data from multiple data sources, including industry news, corporate financial statements, market research reports, policy and regulatory documents, etc. Perform preprocessing operations such as cleaning, denoising, and formatting of the collected data to extract key information such as time, location, subject, event, data indicators, etc., and store the processed data in the database. The selection of data sources is diverse and authoritative, covering industry authoritative websites, government public data platforms, professional databases, etc. The collection frequency f is dynamically adjusted according to the data type t and importance i to ensure the timeliness and comprehensiveness of the data. The collection frequency adjustment formula is: , where k1 and k2 are weight coefficients determined based on experience. The data cleaning process uses regular expression matching, keyword filtering, duplicate data removal and other technologies, and denoising uses mean filtering, median filtering and other methods (for numerical data) or semantic analysis (for text data). The format is uniformly converted according to the predetermined data format specifications;

[0048] Steps to construct the knowledge graph: Based on the preprocessed data, use natural language processing technology and knowledge extraction algorithms to identify entities, relationships and attributes in the data and construct an industry knowledge graph. Among them, entities include industry enterprises, products, technologies, people, etc., relationships include competitive relationships, cooperative relationships, upstream and downstream relationships, affiliation relationships, etc., attributes include enterprise size, market share, financial indicators, etc., and industry knowledge is structured through knowledge graphs to facilitate subsequent analysis and reasoning. Natural language processing technologies include word segmentation, part-of-speech tagging, named entity recognition, semantic role labeling, etc. The knowledge extraction algorithm adopts a combination of rule-based methods, machine learning methods (such as support vector machines, deep learning models, etc.) and semi-supervised learning methods to improve the accuracy and completeness of knowledge extraction. The constructed knowledge graph uses a graph database (such as Neo4j, etc.) for storage and management, supporting efficient query and reasoning operations;

[0049] Reinforcement learning model training steps: Design a reinforcement learning model, define the state space, action space, reward function and transition probability. The state space consists of information such as nodes n, relations r and attributes a in the knowledge graph, and the current report p generation progress, which is expressed in vector form as ,in is a feature encoding function, and the action space includes operations such as selecting data, analyzing data, writing report fragments, and adjusting report structure. The reward function R is calculated based on the quality evaluation indicators of the generated report (such as accuracy A, completeness C, logic L, readability R d etc.) for design, the formula is (where α, β, γ, δ are weight coefficients, and The model continuously learns through trial and error in a simulated environment and optimizes the strategy to maximize long-term rewards. The model training uses algorithms such as Deep Q Network (DQN) or Proximal Policy Optimization (PPO). During the training process, techniques such as experience replay and target network update are used to improve training efficiency and stability.

[0050] Report generation steps: Based on the report topic T and requirements set by the user, the trained reinforcement learning model, with the support of the knowledge graph, automatically selects relevant data, conducts in-depth analysis and reasoning, and generates a preliminary industry report according to the predetermined report template, including industry status analysis, trend forecast, competition landscape, existing problems and suggestions. The report template is customized according to the industry type I and user needs U, including general templates and personalized templates. When generating a report, the model selects the appropriate template according to the topic and requirements, and the probability of selecting a template P t The calculation formula is , where S is the template matching function, θ is the temperature parameter, m is the total number of templates, I represents industry-related indicators, which are used to measure the industry adaptability of the template, and U represents user demand-related indicators, which are used to measure the conformity of the template with user needs. In the data selection process, the semantic query function of the knowledge graph is used to quickly locate relevant data. The analysis process uses statistical analysis, trend analysis, association analysis and other methods. When writing report fragments, the language generation model is combined to generate natural and fluent text content;

[0051] Report optimization steps: Use an optimization algorithm to optimize the initially generated report. The optimization algorithm evaluates and improves the report's language expression, logical structure, content coherence, etc. based on language models (such as the GPT series models, etc.). At the same time, it combines the information in the knowledge graph to conduct fact verification and knowledge supplementation to ensure that the report content is accurate, rich, and organized. The optimization algorithm evaluates language expression based on indicators such as the perplexity of the language model and the BLEU (bilingual evaluation under study) value. The evaluation of the logical structure is performed by analyzing the logical relationships between paragraphs (such as cause-and-effect relationships, progressive relationships, etc.), and the content coherence is measured by calculating the semantic similarity Ss between sentences. During the optimization process, the report is rewritten, reorganized, and content is added or deleted based on the evaluation results. At the same time, the knowledge in the knowledge graph is used to expand and deepen the report, such as supplementing relevant cases and explaining professional terms.

[0052] Evaluation and feedback step: Provide the optimized report to users or experts for evaluation, collect evaluation opinions and feedback information, adjust the parameters and strategies of the reinforcement learning model and the rules of the optimization algorithm based on the feedback, and continuously improve the quality and effect of report generation. User or expert evaluation adopts a combination of quantitative and qualitative methods. Quantitative evaluation evaluates the various quality indicators of the report by scoring (such as 0-10 points), and qualitative evaluation proposes specific improvement suggestions by writing comments. Feedback information is used to update the reward function of the reinforcement learning model and the parameter adjustment rules of the optimization algorithm. Through continuous iterative optimization, the performance of the report generation system is improved.

[0053] In the present invention, in the data collection and preprocessing steps, the data source management submodule is responsible for adding, deleting, updating and setting the priority of the data source to ensure the efficiency and accuracy of data collection. s The calculation is based on the reliability of the data source R s , data update frequency F s and data correlation C s , the calculation formula is , where ω1, ω2, and ω3 are weight coefficients. The data processing submodule uses a distributed computing framework (such as Hadoop, Spark, etc.) to improve the data processing speed and record data processing logs during the processing process to facilitate tracing and troubleshooting;

[0054] In the present invention, in the knowledge graph construction step, the knowledge extraction submodule supports text processing in multiple languages ​​and can adapt to industry data in different language environments. l Through the test statistics of a large number of samples, the calculation formula is: ,in is the number of entities, relations, and attributes extracted correctly, is the total number of extractions. The knowledge fusion submodule is used to fuse and integrate knowledge from different sources, solve knowledge conflicts and redundancy problems, and improve the quality and integrity of the knowledge graph;

[0055] In the present invention, in the reinforcement learning model training step, the model evaluation submodule uses cross-validation and other methods to evaluate the model performance, and adjusts the model parameters and training strategy according to the evaluation results. c The calculation formula is (where k is the number of cross-validation folds, a g is the accuracy of the g-th fold. The model visualization submodule is used to visualize information such as state changes and reward curves during model training to help users understand and optimize the model;

[0056] In the present invention, in the report generation step, the template management submodule provides a rich report template library, supports user-defined templates, and meets the needs of different users and industries. Template matching function Based on topic relevance T 、Industry adaptability I and user needs U The formula for calculating compliance is: , where λ1, λ2, and λ3 are weight coefficients. The report generation engine submodule realizes automatic report generation based on reinforcement learning model and knowledge graph, and supports real-time preview and adjustment during the generation process;

[0057] In the report optimization step, the language evaluation submodule integrates multiple language models and selects appropriate evaluation models according to different languages ​​and application scenarios. m Based on language type L t Application scenario A s and model performance P p Calculation, the formula is , where G is the model fitness function and μ is the selection parameter. The knowledge verification submodule uses the authoritative knowledge in the knowledge graph to verify the report content to ensure the accuracy and reliability of the report;

[0058] In the present invention, in the evaluation and feedback step, the evaluation management submodule is responsible for organizing the evaluation activities, including the management of evaluators, the allocation of evaluation tasks, the collection of evaluation results, etc. p Based on the professional level of the assessor p , Experience E p and assessing historical accuracy H p Calculation, the formula is , where ρ1, ρ2, ρ3 are weight coefficients. The feedback processing submodule classifies, analyzes and transforms the feedback information to provide effective guidance for the optimization of models and algorithms;

[0059] In the present invention, a system for implementing the industry report generation optimization method based on reinforcement learning and knowledge graph includes:

[0060] Data collection and preprocessing module: responsible for collecting industry-related data from multiple data sources, and performing preprocessing operations such as cleaning, denoising, and format unification, storing the processed data in the database, and providing data support for subsequent analysis and report generation;

[0061] Knowledge graph construction module: Build industry knowledge graphs based on preprocessed data, use natural language processing technology and knowledge extraction algorithms to identify entities, relationships, and attributes, use graph databases to store and manage knowledge graphs, and achieve structured representation and efficient query of industry knowledge;

[0062] Reinforcement learning model training module: Design and train reinforcement learning models, define state space, action space, reward function and transition probability, and optimize model strategies through trial and error learning in a simulated environment, so that it can generate high-quality industry reports based on user needs;

[0063] Report generation module: Based on the report topics and requirements set by the user, with the support of knowledge graph and trained reinforcement learning model, it automatically generates preliminary industry reports, including data selection, analysis, report segment writing and structure organization;

[0064] Report optimization module: optimizes the initially generated report using optimization algorithms, evaluates and improves the report's language expression, logical structure, content coherence, etc. based on language models and knowledge graphs to ensure report quality;

[0065] Evaluation and feedback module: collects evaluation opinions and feedback information from users or experts on the optimized reports, and uses them to adjust the reinforcement learning model and optimization algorithm to achieve continuous optimization and performance improvement of the system.

[0066] In the present invention, the data source management submodule in the data acquisition and preprocessing module is responsible for adding, deleting, updating and setting priorities of data sources to ensure the efficiency and accuracy of data acquisition; the data processing submodule adopts a distributed computing framework (such as Hadoop, Spark, etc.) to improve the data processing speed, and records data processing logs during the processing process to facilitate tracing and troubleshooting;

[0067] The knowledge extraction submodule in the knowledge graph construction module supports text processing in multiple languages ​​and can adapt to industry data in different language environments; the knowledge fusion submodule is used to fuse and integrate knowledge from different sources, solve knowledge conflicts and redundancy problems, and improve the quality and integrity of the knowledge graph;

[0068] The model evaluation submodule in the reinforcement learning model training module uses methods such as cross-validation to evaluate model performance and adjust model parameters and training strategies based on the evaluation results. The model visualization submodule is used to visualize information such as state changes and reward curves during model training to help users understand and optimize the model.

[0069] The template management submodule in the report generation module provides a rich report template library and supports user-defined templates to meet the needs of different users and industries; the report generation engine submodule realizes automatic report generation based on reinforcement learning models and knowledge graphs, and supports real-time preview and adjustment during the generation process.

[0070] In the present invention, the language evaluation submodule in the report optimization module integrates multiple language models and selects appropriate evaluation models according to different languages ​​and application scenarios; the knowledge verification submodule uses authoritative knowledge in the knowledge graph to verify the report content to ensure the accuracy and reliability of the report;

[0071] The evaluation management submodule in the evaluation and feedback module is responsible for organizing evaluation activities, including evaluator management, evaluation task allocation, and evaluation result collection; the feedback processing submodule classifies, analyzes, and transforms feedback information to provide effective guidance for the optimization of models and algorithms.

[0072] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An industry report generation optimization method based on reinforcement learning and knowledge graph, characterized in that: include: Data collection and preprocessing steps: Clean, denoise and format the collected data, and store the processed data in the database; Collection frequency Dynamically adjust according to data type t and importance i. The acquisition frequency adjustment formula is: , where k1 and k2 are weight coefficients determined based on experience; regular expressions are used in the data cleaning process, and mean filtering, median filtering or semantic analysis is used for denoising; Knowledge graph construction steps: Based on the preprocessed data, natural language processing technology and knowledge extraction algorithms are used to identify entities, relationships and attributes in the data and build industry knowledge graphs. The knowledge extraction algorithm uses a combination of rule-based methods, support vector machines, deep learning model machine learning methods, and semi-supervised learning methods; Reinforcement learning model training steps: Design a reinforcement learning model, define the state space, action space, reward function and transition probability. The state space consists of nodes n, relations r and attributes a in the knowledge graph, and the current report p generation progress information, which is expressed in vector form as ,in is a feature encoding function. The action space includes operations such as selecting data, analyzing data, writing report fragments, and adjusting report structure. The reward function R is based on the quality evaluation indicators of the generated report, including accuracy A, completeness C, logic L, and readability R. d , the formula is , where α, β, γ, δ are weight coefficients, and ; Model training uses deep Q network DQN or proximal strategy optimization PPO algorithm; Report generation steps: Based on the report topic T and requirements set by the user, the trained reinforcement learning model is used, with the support of the knowledge graph, to automatically select relevant data, perform analysis and reasoning, and generate a preliminary industry report according to the predetermined report template.

2. The industry report generation optimization method based on reinforcement learning and knowledge graph according to claim 1 is characterized in that: Also includes: Report optimization steps: Use the GPT series of models to optimize the initially generated report, and combine the information in the knowledge graph to perform fact checking and knowledge supplement; Evaluation and feedback step: Provide the optimized report to users or experts for evaluation, collect evaluation opinions and feedback information, and adjust the parameters and strategies of the reinforcement learning model based on the feedback; In the data collection and preprocessing steps, the data source management submodule is responsible for adding, deleting, updating and setting the priority of the data source. s The calculation is based on the reliability of the data source R s , data update frequency F s and data correlation C s , the calculation formula is , where ω1, ω2, and ω3 are weight coefficients. The data processing submodule adopts a distributed computing framework, including Hadoop and Spark, and records data processing logs during the processing.

3. The industry report generation optimization method based on reinforcement learning and knowledge graph according to claim 1 is characterized in that: In the knowledge graph construction step, the knowledge extraction submodule supports text processing in multiple languages, and the processing accuracy of texts in different languages ​​is Through the test statistics of a large number of samples, the calculation formula is: ,in is the number of entities, relations, and attributes extracted correctly, is the total number of extractions, and the knowledge fusion submodule is used to fuse and integrate knowledge from different sources.

4. The industry report generation optimization method based on reinforcement learning and knowledge graph according to claim 1 is characterized in that: In the reinforcement learning model training step, the model evaluation submodule uses the cross-validation method to evaluate the model performance and adjusts the model parameters and training strategy according to the evaluation results; the cross-validation accuracy A c The calculation formula is , where k is the number of cross-validation folds, a g is the accuracy of the g-th fold. The model visualization submodule is used to visualize the state changes and reward curve information during the model training process.

5. The industry report generation optimization method based on reinforcement learning and knowledge graph according to claim 1 is characterized in that: In the report generation step, the template management submodule provides a report template library, supports user-defined templates, and meets the needs of different users and industries. The template matching function Based on topic relevance T 、Industry adaptability I and user needs U The formula for calculating compliance is: , where λ1, λ2, and λ3 are weight coefficients; the model selects a template based on the topic and requirements when generating a report, and the probability of selecting a template is P t The calculation formula is , where S is the template matching function, θ is the temperature parameter, m is the total number of templates, I represents the industry-related index, which is used to measure the industry adaptability of the template, and U represents the user demand-related index, which is used to measure the conformity of the template with the user demand.

6. The method for optimizing industry report generation based on reinforcement learning and knowledge graph according to claim 1 is characterized in that: In the report optimization step, the language evaluation submodule integrates multiple language models and selects appropriate evaluation models according to different languages ​​and application scenarios. The selection probability of the language model is P m Based on language type L t Application scenario A s and model performance P p Calculation, the formula is , where G is the model fitness function and μ is the selection parameter.

7. The method for optimizing industry report generation based on reinforcement learning and knowledge graph according to claim 1 is characterized in that: In the evaluation and feedback step, the evaluation management submodule is responsible for organizing the evaluation activities, including the management of evaluators, the assignment of evaluation tasks, the collection of evaluation results, and the weight of the evaluators. p Based on the professional level of the assessor p , Experience E p and assessing historical accuracy H p Calculation, the formula is , where ρ1, ρ2, and ρ3 are weight coefficients.

8. A system for implementing the industry report generation optimization method based on reinforcement learning and knowledge graph as described in any one of claims 1 to 7, characterized in that: include: Data collection and preprocessing module: responsible for collecting industry-related data from multiple data sources, and performing preprocessing operations such as cleaning, denoising, and format unification, storing the processed data in the database, and providing data support for subsequent analysis and report generation; Knowledge graph construction module: Build industry knowledge graphs based on preprocessed data, use natural language processing technology and knowledge extraction algorithms to identify entities, relationships, and attributes, use graph databases to store and manage knowledge graphs, and achieve structured representation and efficient query of industry knowledge; Reinforcement learning model training module: Design and train reinforcement learning models, define state space, action space, reward function and transition probability, and optimize model strategies through trial and error learning in a simulated environment, so that it can generate high-quality industry reports based on user needs; Report generation module: Based on the report topics and requirements set by the user, with the support of knowledge graph and trained reinforcement learning model, it automatically generates preliminary industry reports, including data selection, analysis, report segment writing and structure organization functions; Report optimization module: optimizes the initially generated report using optimization algorithms, evaluates and improves the report’s language expression, logical structure, and content coherence based on language models and knowledge graphs to ensure report quality; Evaluation and feedback module: collects evaluation opinions and feedback information from users or experts on the optimized reports, and uses them to adjust the reinforcement learning model and optimization algorithm to achieve continuous optimization and performance improvement of the system.

9. The system for optimizing the industry report generation method based on reinforcement learning and knowledge graph according to claim 8 is characterized in that: The data source management submodule in the data collection and preprocessing module is responsible for adding, deleting, updating and setting priorities for data sources. The data processing submodule uses the Hadoop or Spark distributed computing framework and records data processing logs during the processing process to facilitate tracing and troubleshooting. The knowledge extraction submodule in the knowledge graph construction module supports text processing in multiple languages; The knowledge fusion submodule is used to fuse and integrate knowledge from different sources; The model evaluation submodule in the reinforcement learning model training module uses the cross-validation method to evaluate the model performance and adjust the model parameters and training strategy based on the evaluation results; The template management submodule in the report generation module provides a report template library and supports user-defined templates to meet the needs of different users and industries; The report generation engine submodule realizes automatic report generation based on reinforcement learning models and knowledge graphs, and supports real-time preview and adjustment during the generation process.

10. The system for optimizing the industry report generation method based on reinforcement learning and knowledge graph according to claim 8, characterized in that: The language evaluation submodule in the report optimization module integrates multiple language models and selects appropriate evaluation models according to different languages ​​and application scenarios; the knowledge verification submodule uses authoritative knowledge in the knowledge graph to verify the report content; the evaluation management submodule in the evaluation and feedback module is responsible for organizing evaluation activities, including evaluator management, evaluation task allocation, and evaluation result collection; The feedback processing submodule classifies, analyzes and transforms the feedback information.

Images