A large language model-based enterprise optimization task generation method and related device

Abstract

The application relates to the technical field of enterprise behavior optimization, and provides an enterprise optimization task generation method based on a large language model and related equipment, which comprises the following steps: for each substantive issue, generating a structured intermediate variable based on multi-modal ESG data by using a large language model, and calculating a behavior trajectory unit of a target enterprise for the substantive issue in each year according to the structured intermediate variable; for each substantive issue in each year, generating a hierarchical state vector of the target enterprise according to the behavior trajectory unit corresponding to the substantive issue in the year; calculating a behavior deviation value of the target enterprise for each substantive issue in each year according to all the hierarchical state vectors, and performing industry attribute calibration on each behavior deviation value to obtain a calibrated behavior deviation value; and generating an enterprise optimization task set of the target enterprise based on all the calibrated behavior deviation values. The method can improve the executability of the generated enterprise optimization tasks.

Description

Technical Field

[0001] This application relates to the field of enterprise behavior optimization technology, and in particular to an enterprise optimization task generation method and related equipment based on a large language model. Background Technology

[0002] As environmental, social, and governance (ESG) information disclosure gradually becomes an important basis for corporate governance, financial due diligence, investment decisions, and regulatory review, companies generate a large amount of heterogeneous data in ESG reports, official websites, promotional images, videos, news and public opinion, announcements and penalties, as well as financial and operational indicators. This type of data has multiple modalities, including text, images, videos, tables, and numerical values. The data structure is scattered, the semantic granularity varies, and the credibility differs greatly, making it difficult to directly perform unified modeling and collaborative analysis.

[0003] Most existing technologies focus on text mining, static scoring, sentiment analysis, or single-modal recognition of ESG reports. They can usually only answer what companies have disclosed, but it is difficult to further identify what companies have actually done, whether they have achieved results, and whether they have verifiable evidence. Therefore, it is difficult to dynamically, traceably, and reliably characterize companies’ ESG behavior.

[0004] In recent years, multimodal large language models have gained the ability to simultaneously understand text, images, video keyframes, and tables. However, in enterprise ESG scenarios, directly requiring the model to output final conclusions such as "the quality of the enterprise's ESG performance," "whether there is a discrepancy between disclosure and actual behavior," or "evaluation level" can easily lead to the following problems: First, the model output process lacks intermediate variable constraints, resulting in black box and illusion problems; Secondly, the output results make it difficult to pinpoint the source of evidence, failing to meet the traceability requirements of regulatory, auditing, and due diligence scenarios; Third, the model tends to mix promotional statements, actual actions, performance results, and verification materials in its judgment, which is not conducive to attributing the source of the problem.

[0005] Furthermore, existing solutions often remain at the identification or scoring level, typically failing to map the identification results into executable, sortable, and verifiable rectification tasks. Consequently, the system output struggles to directly serve the enterprise's rectification closed-loop process, the due diligence requirements of financial institutions, or the continuous monitoring needs in regulatory technology scenarios. This results in poor executability of the generated enterprise optimization tasks. Summary of the Invention

[0006] This application provides a method and related equipment for generating enterprise optimization tasks based on a large language model, which can solve the problem of poor executability of the generated enterprise optimization tasks.

[0007] In a first aspect, embodiments of this application provide a method for generating enterprise optimization tasks based on a large language model, the method comprising: Obtain multimodal ESG data from the target company; multimodal ESG data includes multimodal data on multiple substantive issues of the target company over multiple years; For each of the aforementioned substantive issues, a large language model is used to generate structured intermediate variables based on multimodal ESG data. Then, behavioral trajectory units of the target enterprise on the substantive issues are calculated based on the structured intermediate variables for each year. The structured intermediate variables include structured information from the multimodal data, and the behavioral trajectory units are used to describe the target enterprise's behavior on the substantive issues for each year. For each substantive issue in each year, a hierarchical state vector of the target company is generated based on the behavioral trajectory unit corresponding to the substantive issue in that year. The hierarchical state vector is used to describe the disclosure layer, action layer, performance layer and verification layer status of the target company on the substantive issue. Based on the state vectors of all levels, calculate the behavioral deviation value of the target enterprise for each substantive issue in each year, and perform industry attribute calibration on each behavioral deviation value to obtain the calibrated behavioral deviation value; A set of enterprise optimization tasks for the target enterprise is generated based on all calibrated behavioral deviation values. The set of enterprise optimization tasks includes multiple optimization tasks that the target enterprise needs to perform on substantive issues, as well as the priority of the optimization tasks.

[0008] Secondly, embodiments of this application provide an enterprise optimization task generation device based on a large language model, comprising: The acquisition module is used to acquire multimodal ESG data of the target company; the multimodal ESG data includes multimodal data of the target company on multiple substantive issues over multiple years; The calculation module is used to generate structured intermediate variables based on multimodal ESG data using a large language model for each substantive issue, and to calculate the target company's behavioral trajectory unit for each year based on the structured intermediate variables. The structured intermediate variables include structured information from the multimodal data, and the behavioral trajectory unit is used to describe the target company's behavior on the substantive issues in each year. The generation module is used to generate a hierarchical state vector for the target company for each substantive issue in each year, based on the behavioral trajectory unit corresponding to the substantive issue in that year. The hierarchical state vector is used to describe the disclosure layer, action layer, performance layer and verification layer status of the target company on the substantive issue. The calibration module is used to calculate the behavioral deviation value of the target enterprise for each substantive issue in each year based on the state vectors of all levels, and to perform industry attribute calibration on each behavioral deviation value to obtain the calibrated behavioral deviation value. The task generation module is used to generate a set of enterprise optimization tasks for the target enterprise based on all calibrated behavioral deviation values. The set of enterprise optimization tasks includes multiple optimization tasks that the target enterprise needs to perform on substantive issues, as well as the priority of the optimization tasks.

[0009] Thirdly, embodiments of this application provide a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described enterprise optimization task generation method based on a large language model.

[0010] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned enterprise optimization task generation method based on a large language model.

[0011] The above-mentioned solution in this application has the following beneficial effects: In the embodiments of this application, multimodal ESG data of the target enterprise is acquired. Then, for each substantive issue, a large language model is used to generate structured intermediate variables based on the multimodal ESG data. Based on these structured intermediate variables, behavioral trajectory units of the target enterprise for each substantive issue in each year are calculated. Then, for each substantive issue in each year, a hierarchical state vector of the target enterprise is generated based on the corresponding behavioral trajectory units. Then, based on all hierarchical state vectors, the behavioral deviation value of the target enterprise for each substantive issue in each year is calculated. Each behavioral deviation value is then calibrated according to industry attributes to obtain calibrated behavioral deviation values. Finally, a set of enterprise optimization tasks for the target enterprise is generated based on all calibrated behavioral deviation values. The generation of behavioral deviation values ​​for each substantive issue based on multimodal ESG data enables detailed and targeted information mining, facilitating the analysis of the target enterprise's behavioral deviations at the substantive issue level and revealing optimizable substantive issues. Calibration of the behavioral deviation values ​​improves their accuracy. Generating enterprise optimization tasks based on accurate behavioral deviation values ​​enables automatic generation of enterprise optimization tasks, avoiding the generation of only conclusions or evaluation indicators, and improving the practicality and executability of the generated optimization tasks.

[0012] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating an embodiment of this application of a method for generating enterprise optimization tasks based on a large language model; Figure 2 A schematic diagram of the structure of an enterprise optimization task generation device based on a large language model provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0015] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0016] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0017] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0018] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0019] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0020] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0021] To address the issue of poor executability of existing generated enterprise optimization tasks, this application provides an enterprise optimization task generation method based on a large language model. This method acquires multimodal ESG data of the target enterprise, then, for each substantive issue, uses a large language model to generate structured intermediate variables based on the multimodal ESG data. Based on these structured intermediate variables, it calculates the target enterprise's behavioral trajectory units for each year's substantive issues. Then, for each substantive issue in each year, it generates a hierarchical state vector of the target enterprise based on the corresponding behavioral trajectory units. Finally, based on all hierarchical state vectors, it calculates the target enterprise's behavioral deviation value for each substantive issue in each year, and calibrates each behavioral deviation value according to industry attributes to obtain calibrated behavioral deviation values. Finally, it generates a set of enterprise optimization tasks for the target enterprise based on all calibrated behavioral deviation values. Specifically, the system generates behavioral deviation values ​​for each substantive issue of the target enterprise based on multimodal ESG data, enabling detailed and targeted information mining. This facilitates the analysis of the target enterprise's behavioral deviations at the substantive issue level, reveals optimizable substantive issues, and calibrates the behavioral deviation values ​​to improve their accuracy. Based on accurate behavioral deviation values, the system generates enterprise optimization tasks, enabling the automatic generation of enterprise optimization tasks and avoiding the generation of only conclusions or evaluation indicators, thus improving the practicality and feasibility of the generated optimization tasks.

[0022] The following is an illustrative example of the enterprise optimization task generation method based on a large language model provided in this application.

[0023] like Figure 1 As shown, the enterprise optimization task generation method based on a large language model provided in this application includes the following steps: Step 11: Obtain multimodal ESG data of the target company.

[0024] The target companies mentioned above are those that need to generate enterprise tasks. The multimodal ESG data mentioned above includes multimodal data (such as images, tables, text, etc.) of the target companies on multiple substantive issues over multiple years.

[0025] The aforementioned substantive issues are related to the three aspects of environment, society and governance, such as climate change, energy use, and waste management, and the multiple years refer to several historical years prior to the current year.

[0026] In some embodiments of this application, multimodal ESG data can be obtained by accessing the target company's official website, management system, etc.

[0027] For example, multimodal ESG data includes ESG report text, report page screenshots, official website promotional materials, corporate promotional images, video keyframes, news and public opinion, announcements and penalty information, financial data, operational indicators, and industry benchmark data, etc. This includes data on enterprise... In the year ,issue The multimodal data set on the above is denoted as: ; in, Represents a collection of text data. This represents a collection of images or page screenshots. Represents the set of video keyframes. This refers to a collection of materials from the official website. This represents a collection of news and public opinion. This represents a set of financial or operational metrics. This represents the set of industry benchmarks.

[0028] After obtaining the above multimodal ESG data, it can be preprocessed: Multi-source data is uniformly mapped into four types of indexes: enterprise identifier, time identifier, ESG issue identifier, and source identifier, in order to form a machine-processable multimodal standardized dataset.

[0029] For numerical indicators, the normalization function for direction correction and interval scaling can be expressed as: ; in, These are the original indicator values. and These represent the minimum and maximum values ​​within the historical range of the issue or within the industry scope, respectively. For smoothing terms, This indicates that the interval is truncated to [0,1].

[0030] Calculate the modal control weights for each input information. Modal control weights are determined by source credibility. (The credibility level can be determined based on the type of evidence source, such as regulatory announcements, third-party verification reports, corporate annual ESG reports, official website news, and general media reports, which can be assigned different levels of credibility.) Timeliness (This can be determined based on the time distance between the evidence release date and the target year; the closer the time distance, the higher the value), completeness of the caliber. (This can be determined based on whether the evidence includes elements such as subject, year, indicator name, unit, statistical boundary, base year, and data source) and subject matching degree. (This can be determined based on the degree of matching between the entities involved in the evidence and the target company, subsidiary, project entity, or supply chain link.) The weighted calculation yields the following result, which can be expressed as: ; in, This indicates that the modal weights are normalized. to These are non-negative weighting coefficients, and all parameters are normalized to the [0,1] interval. Modality control weights are used to control the impact of evidence from different sources and modalities on subsequent calculations, avoiding excessive influence of low-confidence, outdated, incomplete, or subject-mismatched data on the results.

[0031] The reasons for choosing source credibility, time freshness, completeness of definition, and subject matching are that the reliability of evidence in ESG scenarios is mainly affected by the source of evidence, time validity, completeness of indicator definition, and consistency of evidence subjects.

[0032] When a certain mode or index is missing, it is not directly set to zero. Instead, the observed data set is renormalized, and a missing value penalty term is calculated. , can be represented as: ; in, This represents the sum of the weights of the actually observed data items. This represents the sum of the weights of the data items that should theoretically be collected. When When the preset threshold is exceeded, the system generates an optimization task of "supplementing basic data".

[0033] Step 12: For each substantive issue, use a large language model to calculate and generate structured intermediate variables based on multimodal ESG data, and then use the structured intermediate variables to generate behavioral trajectory units of the target enterprise on substantive issues for each year.

[0034] The aforementioned behavioral trajectory units are used to describe the target company's behavior on substantive issues within a given year. They are a multimodal evidence set constructed using the target company, target year, and substantive issues as indices.

[0035] The behavioral trajectory unit comprises multiple sets of disclosure evidence, behavioral evidence, performance evidence, and verification evidence. Disclosure evidence records expressive information such as commitments, goals, strategies, systems, slogans, and explanations; behavioral evidence records substantive actions such as project construction, capital investment, energy-saving renovations, supply chain governance, and system implementation; performance evidence records changes in results such as carbon emissions, energy, water resources, waste, and governance effectiveness; and verification evidence records verifiable materials such as third-party verification, audits, certifications, traceability documents, and data boundary and definition explanations. The behavioral trajectory unit is a four-category evidence set oriented towards "enterprise—year—substantive issues," its function being to separately model what the enterprise said, did, what the results were, and whether they are verifiable.

[0036] In some embodiments of this application, the steps described above—using a large language model to generate structured intermediate variables based on multimodal ESG data, and calculating the target enterprise's behavioral trajectory units on substantive issues for each year based on the structured intermediate variables—include: For each year, perform the following steps: The first step is to use a large language model to understand and extract multimodal data corresponding to substantive issues in the multimodal ESG data for the current year, and obtain structured intermediate variables.

[0037] Specifically, multimodal data corresponding to substantive issues in the annual multimodal ESG data are input into a large language model. The large language model is then used to understand and extract the multimodal data to obtain structured intermediate variables.

[0038] For example, the structured intermediate variables mentioned above include structured information in multimodal data, including at least the following fields: responsible party, year or time range, ESG issue, semantic type, strength value, evidence location, and confidence level.

[0039] The responsible entity is used to identify the enterprise, subsidiary, project entity, supply chain link, or responsible department; the semantic type includes at least one or more of disclosure, publicity, action, performance, verification, and conflict; the evidence location is used to locate page numbers, image areas, video frame numbers, web page paths, news links, or data table cells; the confidence level is used to indicate the model's credibility with the extraction results.

[0040] After obtaining the above structured intermediate variables, multiple behavioral indicators can be calculated, such as: Calculate the intensity of visual rhetoric to address potential propaganda rhetoric in images and videos. It can be represented as: ; ; ; ; ; in, The intensity of natural imagery can be represented by the frequency, salient area, and recognition confidence score of natural elements such as forests, oceans, rivers, grasslands, skies, and animals detected by image recognition models. The intensity of a green hue can be determined by the proportion of green pixels in an image, the proportion of green-themed areas in a layout, or the color label output by a visual model. The intensity of slogan-like expressions can be determined by recognizing promotional slogans, visionary words, and promise words in image or video frames using OCR, and then calculating the intensity by combining word frequency and the salient position of text regions. The strength of real facilities or real production scenarios can be obtained by identifying real action scenarios such as factories, production lines, equipment, environmental protection facilities, project sites, and monitoring equipment. Real facilities or real production scenarios are used as suppression terms in the calculation to avoid misjudging real inputs as propaganda rhetoric. A represents the total area of ​​keyframes in the image or video to be analyzed. Let represent the confidence level for identifying the i-th natural image target. This indicates the area percentage of the target natural image; Indicates the number of green-hue pixels. Indicates the total number of pixels; This indicates the number of promotional slogans recognized by the OCR. This represents the total number of text words recognized by the OCR. This indicates the percentage of the area occupied by the slogan text. This represents the confidence level in identifying the j-th real facility or real production scenario target. This represents the area percentage of the target in the real-world scene. All of the above parameters can be mapped to the [0,1] interval, or can be obtained by normalization after multimodal model recognition. Visual rhetoric intensity is used to identify promotional expressions in images and videos brought about by "natural imagery, green tones, and slogan-like expressions," while using real facilities or real production scenes as deduction items to avoid misjudging real production inputs as propaganda.

[0041] The above parameter formula is one possible calculation method, or it can be obtained by directly identifying and normalizing the multimodal model.

[0042] Calculate the completeness of quantitative disclosure This can be obtained by detecting whether the disclosed content contains key elements such as year, unit, benchmark, boundary, and value, and can be expressed as: ; in, , , , and These are indicator variables for year, unit, baseline, boundary, and value, respectively. They are set to 1 if they exist and 0 if they do not exist.

[0043] Calculate cross-modal consistency This is used to measure whether text, images, videos, news, and business metrics can support each other. The main data source is the comparison results between text, images, videos, news, business metrics, and verification materials, which can be expressed as: ; in, and These represent semantic embedding vectors or index vectors obtained from different modal extractions, respectively. This represents the similarity function, whose result is mapped to the interval [0,1].

[0044] The second step is to construct behavioral trajectory units based on structured intermediate variables.

[0045] For example, the behavioral trajectory unit includes disclosure evidence, behavioral evidence, performance evidence, and verification evidence, and may also include disclosure intensity. Intensity of substantive actions Performance changes Verifiable strength and cross-modal consistency When the same enterprise has behavioral trajectory units on the same issue in consecutive years, they are connected in chronological order to form a cross-period trajectory sequence.

[0046] Each piece of evidence should include at least the company ID, year, issue, source type, original location, extracted fields, model confidence level, manual review status, and task number to facilitate subsequent audit tracking. Elements in the structured intermediate variables can be categorized into four types: disclosure, behavior, performance, and verification, with all elements under each category treated as evidence of that type.

[0047] For example, after obtaining the structured intermediate variables, the system maps them to the disclosure evidence set, action evidence set, performance evidence set, and verification evidence set, respectively, based on the semantic type field, evidence keywords, numerical indicator type, and source type. If the semantic type is disclosure, publicity, goal, or commitment, it enters the disclosure evidence set; if the semantic type is action, project, input, measure, or implementation, it enters the action evidence set; if the semantic type is performance, result, year-on-year change, indicator improvement, or industry benchmarking, it enters the performance evidence set; and if the semantic type is verification, audit, certification, boundary description, definition, or traceable voucher, it enters the verification evidence set.

[0048] Each piece of evidence retains its source type, original location, model confidence level, and manual verification status.

[0049] Step 13: For each substantive issue in each year, generate the hierarchical state vector of the target enterprise based on the behavioral trajectory unit corresponding to the substantive issue in the year.

[0050] The aforementioned hierarchical state vectors are used to describe the target company's disclosure, action, performance, and verification status on substantive issues. These hierarchical state vectors include disclosure intensity, substantive action intensity, performance change intensity, and verifiability intensity.

[0051] Disclosure strength describes the strength of a company’s commitments, goals, strategies, systems, slogans and statements on a given issue. It does not determine whether a company has taken action; it only calculates the strength, completeness and credibility of the disclosure itself. The main data sources are ESG reports, official website materials, promotional images, videos, press releases and announcements. The intensity of substantive actions is used to characterize the substantive actions that an enterprise has implemented or is implementing, including green investment, energy-saving renovation, pollution control, supply chain management, employee rights measures, implementation of governance systems, project construction progress, etc. The main data sources are energy-saving renovation projects, green investment, certification materials, governance measures, supply chain management, and system implementation. The intensity of performance change is used to characterize changes in a company's outcomes on a given issue. The main data sources are indicators such as improvements in carbon emissions, energy, waste, water resources, and governance. Verifiability strength is used to characterize whether a company’s disclosures, actions and performance are supported by verifiable evidence. The main data sources are third-party verification, audit reports, certification materials, traceable vouchers, and explanations of data boundaries and definitions.

[0052] In some embodiments of this application, the steps of generating the hierarchical state vector of the target enterprise based on the behavioral trajectory unit corresponding to the substantive issue in each year include: The first step is to calculate the disclosure intensity corresponding to the substantive issues in the year, based on all disclosure evidence in the behavioral trajectory unit.

[0053] Specifically, disclosure intensity can be expressed as: ; in, Indicates enterprise In the year and issues The intensity of disclosure on the surface Indicates the first The disclosure characteristic value corresponding to each piece of disclosure evidence. This indicates the confidence level of the multimodal large language model in the outcome of the evidence analysis. This represents the evidence quality weights formed by source credibility, time freshness, and subject matching. `clip` indicates that the result is truncated to the [0,1] interval. The disclosure characteristic value is the contribution of a single piece of disclosure evidence to the disclosure intensity D, and is not equivalent to the disclosure intensity of the entire report. It can be calculated based on the strength of commitment, clarity of objective, degree of quantification, salience of location, and frequency of repetition in that piece of disclosure evidence. ; Among them, the strength of the commitment statement Reflecting the strength of expressions such as "committed to," "dedicated to," "continuously promoted," and "fully realized"; the clarity of the goal. This reflects whether it includes clear objectives, deadlines, or responsible parties; and the degree of quantification. This reflects whether specific numerical values, units, base years, or statistical boundaries are included; and the prominence of the disclosure location. This reflects whether the disclosure is located in the title, abstract, key chapter, main visual element of the text or image, or key frame of the video; the frequency of repetition. This reflects the degree to which the same topic appears repeatedly in different materials. to These are the weighting coefficients.

[0054] The second step is to calculate the intensity of substantive actions corresponding to substantive issues under the year, based on all behavioral evidence in the behavioral trajectory unit.

[0055] Specifically, the intensity of substantive action can be expressed as: ; in, Indicates the intensity of actual action. Indicates the first The action characteristic value of the evidence of behavior can be determined by the existence of the project, the amount of investment, the progress of completion, the responsible department, the execution time and the scope of coverage. Indicates the model confidence level. Indicates the weight of the quality of behavioral evidence.

[0056] The third step is to calculate the verifiable strength of the substantive issues for the year based on all the verification evidence in the behavioral trajectory unit.

[0057] Specifically, verifiable strength can be expressed as: ; in, Indicates verifiable strength. This indicates the strength of third-party audit, attestation, or verification materials in the evidence. This indicates the strength of certified, rated, or licensed materials in the verification evidence. This indicates the strength of traceable evidence in the verification evidence. This indicates the strength of the statistical boundaries, accounting standards, base year, and data source explanations in the verification evidence. This indicates the status of manual review or the percentage of review passed in the verification of evidence. to This indicates the corresponding weight.

[0058] The fourth step is to calculate the intensity of performance changes corresponding to substantive issues in the year, based on all performance evidence in the behavioral trajectory unit.

[0059] Specifically, performance changes can be represented as: ; ; in, Indicates the intensity of performance change. Indicates the first The weight of each piece of performance evidence (which can be performance indicators). This represents the change in the index after direction correction and normalization. This represents the direction coefficient of the indicator. A positive value is taken when the direction of improvement is positive, and a negative correction is taken for indicators where a lower value is better. and These represent the index values ​​for the current year and the previous year, respectively, and Norm represents the normalization function.

[0060] The fifth step is to combine the disclosure intensity, substantive action intensity, performance change intensity, and verifiable intensity corresponding to the substantive issues under the year into a vector to obtain the hierarchical state vector.

[0061] The weights within each layer of the hierarchical state vector can be determined by expert rules, sample training, or adaptive combination. To balance ESG domain knowledge and sample data adaptability, the final weight of any layer can be expressed as: ; in, Indicates the first The final weight of each layer Indicates expert weights, This represents the data weights obtained through training based on historical or industry samples. This represents the fusion coefficient.

[0062] Step 14: Based on the state vectors of all levels, calculate the behavioral deviation value of the target enterprise for each substantive issue in each year, and perform industry attribute calibration on each behavioral deviation value to obtain the calibrated behavioral deviation value.

[0063] Specifically, through the formula: ; Calculate the target firm's behavioral deviation value on substantive issues at the end of the year. .

[0064] in, This indicates a function that limits the range of values. Indicates the intertemporal deviation adjustment coefficient, This represents the cross-modal inconsistency adjustment coefficient. Indicates the target company's annual On substantive issues The intensity of symbolic disclosure Indicates the target company's annual On substantive issues The overall substantive intensity of the behavior, Indicates the target company's annual On substantive issues Intertemporal divergence, Indicates the target company's annual On substantive issues Cross-modal consistency: ; ; ; in, , , , This is the non-negative weighting coefficient in the calculation of symbolic disclosure intensity. , , , To incorporate the non-negative weighting coefficients in the calculation of the intensity of substantive behavior, Indicates the year Substantive issues The disclosure intensity in the corresponding hierarchical state vector, Indicates the target company's annual Substantive issues The intensity of visual rhetoric, Indicates the target company's annual Substantive issues The strength of the commitment Indicates the target company's annual Substantive issues The completeness of quantitative disclosure Indicates the year Substantive issues The actual action intensity in the corresponding hierarchical state vector Indicates the year Substantive issues The intensity of performance change in the corresponding hierarchical state vector Indicates the year Substantive issues The verifiable strength in the corresponding hierarchical state vector, Indicates substantive issues under the current year The intensity of disclosure This refers to the substantive issues in the previous year of the current year. The intensity of disclosure Indicates substantive issues under the current year The actual intensity of action, This refers to the substantive issues in the previous year of the current year. The actual intensity of action.

[0065] Through the formula: ; ; Calculate the target company in the annual On substantive issues The calibrated behavior deviation value ; in, , , , , Indicates the calibration coefficient. Indicates the deviation from industry benchmarks. Indicates the level of pressure in the industry. Indicates regulatory sensitivity. This indicates the firm size correction item. This represents the benchmark level of substantive issues in the industry, indicating the level of ESG issues within the target company's industry. The industry benchmark level is determined by industry quantiles based on standardized indicator values ​​of companies in the same industry within the same year or a pre-set historical period. The standard deviation represents the baseline level. This is a smoothing term.

[0066] Step 15: Generate a set of enterprise optimization tasks for the target enterprise based on all calibrated behavioral deviation values.

[0067] The enterprise optimization task set includes optimization tasks that multiple target enterprises need to perform on substantive issues, as well as the priority of these optimization tasks.

[0068] In some embodiments of this application, the step of generating the target enterprise's set of enterprise optimization tasks based on all calibrated behavioral deviation values ​​includes: The first step is to break down each calibrated behavior bias value into multiple bias sources and define the observation amplitude for each bias source.

[0069] For example, sources of bias include at least: excessive disclosure, insufficient behavior, insufficient performance, insufficient validation, cross-modal inconsistency, deviation from industry benchmarks, and missing basic data. The observation magnitude for each source of bias is defined based on the behavioral bias value and its related parameters. Excessive disclosure: = (0, - ); Insufficient behavior: = (0, - ); Insufficient performance: = (0, - ); Insufficient validation: = (0, - ); Cross-modal inconsistency: = (0, - ); Industry benchmark deviation: = ; Basic data is missing: = .

[0070] in, , , and The issue thresholds, representing the intensity of substantive action, performance change, verifiable strength, and cross-modal consistency, can be set by experts or automatically determined based on industry quantiles.

[0071] The second step is to calculate the contribution of each source of bias based on all observed amplitudes.

[0072] Specifically, through the formula: ; Calculate the source of deviation Contribution .

[0073] in, Indicates the source of the deviation The weight, Indicates the source of the deviation The range of observations Indicates the source of the deviation The weight, Indicates the source of the deviation The range of observations This is a smoothing term.

[0074] The third step is to generate multiple structured optimization task objects based on all contributions, and to calculate the task priority of each structured optimization task object.

[0075] The aforementioned structured optimization task objects are used to describe the tasks that enterprises can perform on substantive issues, such as "increasing energy-saving renovations, supply chain governance, or green investment projects." These can be obtained by analyzing and reasoning about the contribution of all sources of deviation using large language models.

[0076] Specifically, through the formula: ; Calculate the task priority of the structured optimization task object. .

[0077] in, , , , , , These are the weighting coefficients used in task priority calculation. This represents the deviation contribution of the structured optimization task object. This indicates the improvement benefits of the structured optimization task object. This indicates the industry and regulatory urgency corresponding to the structured optimization task object. This indicates the verifiability of the results of the structured optimization task object. This represents the implementation cost of the structured optimization task object. This indicates the difficulty of implementing the structured optimization task object.

[0078] For example, large language models can be used, taking all contribution levels, ESG issue labels, industry attributes, and enterprise development stages as input, to output multiple structured optimization task objects. Each structured optimization task object should include at least: task number, associated issue, triggering reason, execution action, responsible party, verification indicators, completion deadline, and evidence requirements. An example of the rules for generating structured optimization task objects is as follows: when ≥0.65 and If the value is ≤0.50, or if the disclosure of excessive contribution exceeds the threshold, a task of "supplementing quantitative disclosure, reducing promotional statements, disclosure boundaries and base year" will be generated. when If the value is ≤0.50 or has not improved for two consecutive years, a task of "adding energy-saving renovation, supply chain governance or green investment projects" will be generated. when When the value is ≤0.50 or lower than the industry benchmark, a task of "developing a plan to improve carbon emissions, energy, water resources, or waste" will be generated. when When the value is ≤0.50, a task of "supplementing third-party verification, audit report, certification materials or traceable vouchers" will be generated. when When the value is ≤0.60, a task of "reviewing and supplementing the differences between promotional images, report texts, news, and operational data" will be generated. when When the industry threshold is reached, a task of "conducting industry benchmarking rectification and disclosing benchmarking criteria" will be generated. when When the value is ≥0.30, a task of "filling in missing data sources and responsible departments" will be generated.

[0079] The fourth step is to determine the set of enterprise optimization tasks for the target enterprise from all structured optimization task objects based on the priority of all tasks.

[0080] For example, the top few structured optimization tasks with the highest task priority values ​​can be used as the enterprise optimization tasks for the target enterprise.

[0081] To provide comprehensive guidance to target companies, a large language model can be used to generate behavior optimization reports based on task priorities and structured optimization of task objects. The report should include at least: the company's multimodal ESG profile, the company's annual issue behavior trajectory, disclosure intensity, substantive action intensity, performance changes, verifiable intensity, cross-modal consistency, behavior deviation index, industry calibration results, deviation source contribution, structured optimization of task objects, task priorities, verification indicators, and traceable materials.

[0082] It is worth mentioning that the behavioral deviation values ​​of target enterprises for each substantive issue are generated based on multimodal ESG data, enabling detailed and targeted information mining. This facilitates the analysis of behavioral deviations of target enterprises at the substantive issue level, reveals substantive issues that can be optimized, and calibrates the behavioral deviation values ​​to improve their accuracy. Based on accurate behavioral deviation values, enterprise optimization tasks are generated automatically, avoiding the generation of only conclusions or evaluation indicators, and improving the practicality and feasibility of the generated optimization tasks.

[0083] The method of this application will be illustrated below with a specific example.

[0084] Taking the ESG data processing of Company A on environmental issues in 2024 as an example, the system first collects Company A's ESG report text, official website pages, pictures, video keyframes, related news, green investment amount, energy-saving renovation projects, carbon emission intensity data and industry benchmark data, and maps them uniformly to the company logo, annual logo and issue logo.

[0085] After standardization, the system inputs the above data into a controlled multimodal large language model. For each piece of evidence, the model outputs only structured intermediate variables, without outputting final judgments such as "the company has excellent environmental performance" or "the company's disclosures are inconsistent with its actual behavior." A preferred JSON output example is as follows: { "subject": "company", "year": "2024", "topic": "carbon emissions", "semanticType": "disclosure", "intensity": 0.82, "evidencePosition": "ESG report page 12", "confidence": 0.91}.

[0086] The above constraints allow the system to retain the location of the original evidence corresponding to each conclusion, providing a basis for subsequent manual review and audit trail.

[0087] After obtaining the structured intermediate variables, the system constructs behavioral trajectory units (Company A, 2024, carbon emissions) according to the enterprise, the year 2024, and the "carbon emissions" issue, and obtains the following example values: Disclosure intensity D=0.778; The intensity of actual action, Aact, is 0.424. Performance change P=0.391; Verifiable strength Ver=0.320; Cross-modal consistency CM=0.580.

[0088] Furthermore, the system calculates symbolic disclosure intensity based on disclosure intensity, visual rhetoric intensity, commitment intensity, and quantitative disclosure completeness; and calculates comprehensive substantive behavior intensity based on substantive action intensity, performance change, verifiable intensity, and cross-modal consistency.

[0089] Subsequently, the system uses the behavioral deviation index formula to obtain, and further combines industry pressure and regulatory sensitivity to obtain the industry-calibrated deviation index.

[0090] The above results indicate that companies' disclosures on this issue are significantly stronger than their overall substantive actions, and that such disclosures have a higher priority for rectification within the industry context.

[0091] Further calculate the contribution of each source of deviation. Assuming the company contributes significantly to "insufficient verification," "deviation from industry benchmarks," and "insufficient quantitative disclosure boundaries," the system can automatically generate the following structured task: Task 1: Supplement the third-party verification report; Triggering reason: insufficient verification contribution is high; Verification indicator: coverage of verifiable evidence.

[0092] Task 2: Conduct industry benchmarking and rectification; Triggering reason: significant deviation from industry benchmarks; Verification indicator: industry benchmark achievement rate.

[0093] Task 3: Supplement the quantification boundary and base year for carbon emissions; Triggering reason: Strong disclosure but insufficient quantification boundary; Verification indicator: Completeness rate of quantification disclosure.

[0094] Task 4: Develop an energy consumption improvement plan; Triggering reason: Insufficient performance improvement; Validation indicator: Improvement rate of core performance indicators.

[0095] The system further prioritizes the above tasks based on deviation contribution, improvement benefits, industry urgency, verifiability, implementation cost, and implementation difficulty, and outputs a task list for enterprise rectification closed loop.

[0096] Taking Company A's publicly available materials on environmental issues in 2024 as an example, the system retrieves its ESG report, official website page, report screenshots, promotional images, video keyframes, green investment amount, energy-saving renovation projects, third-party certification materials, carbon emission intensity, and industry benchmark data.

[0097] After the controlled multimodal large language model outputs intermediate variables, the validation data is shown in Table 1.

[0098] Table 1 Validation Data ; Based on the contribution of the sources of deviation, the system generates optimization tasks as shown in Table 2.

[0099] Table 2 Optimization Tasks ; In green credit, green bonds, responsible investment, or M&A due diligence scenarios, the system can batch access candidate companies' ESG reports, official website materials, news and public opinion, and operating indicators to form the company's annual issue behavior trajectory. For companies with high disclosure intensity but insufficient substantive behavior and performance improvement, the system outputs the source contribution of deviations and task priorities, allowing due diligence personnel to request companies to supplement materials, set post-loan rectification conditions, adjust credit structure, or optimize credit conditions.

[0100] In regulatory technology scenarios, the system can continuously monitor key industry enterprises, calibrate deviation indices based on industry pressure and regulatory sensitivity, and generate lists of key enterprises, key issues, and key rectification tasks. In enterprise SaaS platform scenarios, the system can further break down optimization tasks to responsible departments, completion deadlines, and verification materials, forming a closed-loop management system for ESG rectification.

[0101] Therefore, this application has the following advantages: 1. By restricting the multimodal large language model to a structured intermediate variable outputter, the model avoids directly outputting black-box conclusions, thereby improving the traceability, auditability, and verifiability of the results.

[0102] 2. By constructing behavioral trajectory units based on enterprises, years, and issues, and further connecting them into inter-period trajectory sequences, dynamic representation of enterprise ESG behavior can be achieved, which has a higher time sequence identification capability compared with static text analysis.

[0103] 3. By decoupling computation across four layers—disclosure, behavior, performance, and verification—judgment distortion caused by the mixing of information from different semantic layers is reduced, thereby improving the accuracy of problem localization and the ability to attribute biases.

[0104] 4. Improve the ability to identify mismatches between promotional disclosures and comprehensive substantive behavior by jointly calculating cross-modal consistency and intertemporal deviations.

[0105] 5. By mapping the contribution of deviation sources to task rules, structured optimization task objects are generated, and the rectification order is calculated based on priority to improve the matching degree between rectification suggestions and problem sources, as well as the executability of system output.

[0106] 6. This application can be applied to closed-loop management of enterprise ESG rectification, as well as to due diligence, continuous monitoring of regulatory technology, and enterprise SaaS platform scenarios in financial institutions. It has strong versatility and engineering implementation value.

[0107] The following is an exemplary description of the enterprise optimization task generation device based on a large language model provided in this application.

[0108] like Figure 2 As shown, this application provides an enterprise optimization task generation device based on a large language model. The enterprise optimization task generation device 200 based on a large language model includes: The acquisition module 201 is used to acquire multimodal ESG data of the target company; the multimodal ESG data includes multimodal data of the target company on multiple substantive issues over multiple years; The calculation module 202 is used to calculate and generate structured intermediate variables based on multimodal ESG data using a large language model for each substantive issue, and to generate behavioral trajectory units for the target enterprise on substantive issues for each year based on the structured intermediate variables; the structured intermediate variables include structured information in the multimodal data, and the behavioral trajectory units are used to describe the target enterprise's behavior on substantive issues for each year; The generation module 203 is used to generate a hierarchical state vector for the target company for each substantive issue in each year, based on the behavioral trajectory unit corresponding to the substantive issue in the year. The hierarchical state vector is used to describe the disclosure layer, action layer, performance layer and verification layer status of the target company on the substantive issue. The calibration module 204 is used to calculate the behavioral deviation value of the target enterprise for each substantive issue in each year based on the state vectors of all levels, and to perform industry attribute calibration on each behavioral deviation value to obtain the calibrated behavioral deviation value. The task generation module 205 is used to generate a set of enterprise optimization tasks for the target enterprise based on all calibrated behavioral deviation values. The set of enterprise optimization tasks includes multiple optimization tasks that the target enterprise needs to perform on substantive issues, as well as the priority of the optimization tasks.

[0109] It should be noted that the information interaction and execution process between the above-mentioned devices / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0110] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0111] like Figure 3 As shown, an embodiment of this application provides a terminal device, wherein the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 3 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.

[0112] Specifically, when the processor D100 executes the computer program D102, it acquires the multimodal ESG data of the target enterprise, and then, for each substantive issue, uses a large language model to generate structured intermediate variables based on the multimodal ESG data. Based on the structured intermediate variables, it calculates the behavioral trajectory units of the target enterprise for each substantive issue in each year. Then, for each substantive issue in each year, it generates the hierarchical state vector of the target enterprise based on the behavioral trajectory units corresponding to the substantive issues in that year. Based on all hierarchical state vectors, it calculates the behavioral deviation value of the target enterprise for each substantive issue in each year, and performs industry attribute calibration on each behavioral deviation value to obtain the calibrated behavioral deviation value. Finally, it generates a set of enterprise optimization tasks for the target enterprise based on all calibrated behavioral deviation values. Specifically, the system generates behavioral deviation values ​​for each substantive issue of the target enterprise based on multimodal ESG data, enabling detailed and targeted information mining. This facilitates the analysis of the target enterprise's behavioral deviations at the substantive issue level, reveals optimizable substantive issues, and calibrates the behavioral deviation values ​​to improve their accuracy. Based on accurate behavioral deviation values, the system generates enterprise optimization tasks, enabling the automatic generation of enterprise optimization tasks and avoiding the generation of only conclusions or evaluation indicators, thus improving the practicality and feasibility of the generated optimization tasks.

[0113] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0114] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.

[0115] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.

[0116] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.

[0117] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying the computer program code to the enterprise optimization task generation method apparatus / terminal device based on a large language model, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0118] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0119] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A method for generating enterprise optimization tasks based on a large language model, characterized in that, include: Obtain multimodal ESG data from the target company; The multimodal ESG data includes multimodal data on multiple substantive issues for the target company over multiple years; For each of the substantive issues, a large language model is used to generate structured intermediate variables based on the multimodal ESG data, and behavioral trajectory units of the target enterprise on the substantive issues in each year are calculated based on the structured intermediate variables. The structured intermediate variables include structured information from multimodal data, and the behavioral trajectory units are used to describe the target company's behavior on substantive issues during the year; For each substantive issue in each year, a hierarchical state vector of the target enterprise is generated based on the behavioral trajectory unit corresponding to the substantive issue in that year. The hierarchical state vector is used to describe the target company's disclosure, action, performance, and verification status on substantive issues. Based on all hierarchical state vectors, the behavioral deviation value of the target enterprise for each substantive issue in each year is calculated, and each behavioral deviation value is calibrated for industry attributes to obtain the calibrated behavioral deviation value. A set of enterprise optimization tasks for the target enterprise is generated based on all calibrated behavioral deviation values. The set of enterprise optimization tasks includes optimization tasks that multiple target enterprises need to perform on substantive issues, as well as the priority of the optimization tasks.

2. The enterprise optimization task generation method according to claim 1, characterized in that, The process of utilizing a large language model to generate structured intermediate variables based on the multimodal ESG data, and calculating the behavioral trajectory units of the target enterprise on the substantive issues for each year based on the structured intermediate variables, includes: For each of the aforementioned years, the following steps shall be performed: Using a large language model, the multimodal data corresponding to the substantive issues in the year in the multimodal ESG data are understood and extracted to obtain structured intermediate variables; Behavioral trajectory units are constructed based on the structured intermediate variables.

3. The enterprise optimization task generation method according to claim 2, characterized in that, The behavioral trajectory unit includes multiple disclosure evidence, multiple behavioral evidence, multiple performance evidence, and multiple verification evidence; the hierarchical state vector includes disclosure intensity, substantive action intensity, performance change intensity, and verifiability intensity. The step of generating a hierarchical state vector for the target enterprise based on the behavioral trajectory unit corresponding to each substantive issue in each year includes: Based on all disclosure evidence in the behavioral trajectory unit, calculate the disclosure intensity corresponding to the substantive issue in the year; Based on all behavioral evidence in the behavioral trajectory unit, calculate the substantive action intensity corresponding to the substantive issue in the year; Based on all the verified evidence in the behavioral trajectory unit, calculate the verifiable strength of the substantive issue in the year; Based on all performance evidence in the behavioral trajectory unit, calculate the intensity of performance change corresponding to the substantive issues in the year; The disclosure intensity, substantive action intensity, performance change intensity, and verifiable intensity corresponding to the substantive issues under the given year are combined into a vector to obtain the hierarchical state vector.

4. The enterprise optimization task generation method according to claim 3, characterized in that, The step of calculating the target enterprise's behavioral deviation value for each substantive issue in each year based on all hierarchical state vectors includes: Through the formula: ； Calculate the target company in the annual On substantive issues behavioral deviation value ; in, This indicates a function that limits the range of values. Indicates the intertemporal deviation adjustment coefficient, This represents the cross-modal inconsistency adjustment coefficient. Indicates the target company's annual On substantive issues The intensity of symbolic disclosure Indicates the target company's annual On substantive issues The overall substantive intensity of the behavior, Indicates the target company's annual On substantive issues Intertemporal divergence, Indicates the target company's annual On substantive issues Cross-modal consistency: ； ； ； in, , , , This is the non-negative weighting coefficient in the calculation of symbolic disclosure intensity. , , , To incorporate the non-negative weighting coefficients in the calculation of the intensity of substantive behavior, Indicates the year Substantive issues The disclosure intensity in the corresponding hierarchical state vector, This indicates that the target company in the annual Substantive issues The intensity of visual rhetoric, This indicates that the target company in the annual Substantive issues The strength of the commitment This indicates that the target company in the annual Substantive issues The completeness of quantitative disclosure Indicates the year Substantive issues The actual action intensity in the corresponding hierarchical state vector Indicates the year Substantive issues The intensity of performance change in the corresponding hierarchical state vector Indicates the year Substantive issues The verifiable strength in the corresponding hierarchical state vector, Indicates substantive issues under the current year The intensity of disclosure This refers to the substantive issues in the previous year of the current year. The intensity of disclosure Indicates substantive issues under the current year The actual intensity of action, This refers to the substantive issues in the previous year of the current year. The actual intensity of action.

5. The enterprise optimization task generation method according to claim 4, characterized in that, The process of calibrating each behavioral deviation value according to industry attributes to obtain calibrated behavioral deviation values ​​includes: Through the formula: ； ； Calculate the target company in the annual On substantive issues The calibrated behavior deviation value ; in, , , , , Indicates the calibration coefficient. Indicates the deviation from industry benchmarks. Indicates the level of pressure in the industry. Indicates regulatory sensitivity. This indicates the firm size correction item. This represents the benchmark level for substantive issues in the industry. The standard deviation represents the baseline level. This is a smoothing term.

6. The enterprise optimization task generation method according to claim 1, characterized in that, The process of generating the enterprise optimization task for the target enterprise based on all calibrated behavioral deviation values ​​includes: For each of the calibrated behavior deviation values, the calibrated behavior deviation value is broken down into multiple deviation sources, and the observation amplitude of each deviation source is defined; The contribution of each source of bias is calculated based on all observed amplitudes; Multiple structured optimization task objects are generated based on all contributions, and the task priority of each structured optimization task object is calculated. Based on the priority of all tasks, the set of enterprise optimization tasks for the target enterprise is determined from all structured optimization task objects.

7. The enterprise optimization task generation method according to claim 6, characterized in that, The calculation of the contribution of each source of bias based on all observed amplitudes includes: Through the formula: ； Calculate the source of deviation Contribution ; in, Indicates the source of the deviation The weight, Indicates the source of the deviation The range of observations Indicates the source of the deviation The weight, Indicates the source of the deviation The range of observations This is a smoothing term.

8. The enterprise optimization task generation method according to claim 7, characterized in that, The calculation of the task priority for each structured optimization task object includes: Through the formula: ； Calculate the task priority of the structured optimization task object. ; in, , , , , , These are the weighting coefficients used in task priority calculation. This represents the deviation contribution of the structured optimization task object. This indicates the improvement benefits of the structured optimization task object. This indicates the industry and regulatory urgency corresponding to the structured optimization task object. This indicates the verifiability of the results of the structured optimization task object. This represents the implementation cost of the structured optimization task object. This indicates the difficulty of implementing the structured optimization task object.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the enterprise optimization task generation method based on a large language model as described in any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the enterprise optimization task generation method based on a large language model as described in any one of claims 1 to 8.

Images