ESG risk early warning method and system based on multi-modal data

Abstract

The application provides an ESG risk early warning method and system based on multi-modal data, and belongs to the technical field of risk early warning, and specifically comprises the following steps: based on the deviation of the associated data sources of the target enterprise and other target enterprises in different modal data types, and in combination with the associated modal data types of different enterprises, an ESG risk early warning method for the target enterprise is performed; the risk early warning method is used to determine the deviation of the ESG risk values of the target enterprise between different analysis batches; in combination with the associated data sources of the enterprise in different modal data types, the analysis deviation risk type of the modal data type is determined; based on the analysis deviation risk type in different modal data types and the coincidence of the associated data sources and the target enterprise, the determination of the ESG risk value early warning management method of the enterprise is performed, and the accuracy of the ESG risk early warning processing is improved.

Description

Technical Field

[0001] This invention belongs to the field of risk warning technology, and in particular relates to an ESG risk warning method and system based on multimodal data. Background Technology

[0002] With the deepening of the concept of sustainable development, ESG (Environmental, Social, and Governance) performance has become a key dimension for assessing a company's long-term value and risks. Investors, regulators, and companies themselves have an urgent need for efficient and accurate ESG risk management tools.

[0003] To address the aforementioned technical issues, existing solutions, as described in CN202511452670.1 "A Machine Learning-Based Enterprise ESG Score Prediction Method," first collect multi-dimensional ESG, financial, and risk data. After normalization and handling of missing / outlier values, the dataset is divided. A multi-model system is constructed, using a five-dimensional index to select the optimal LSTM for ESG rating prediction. Combining the LSTM results, with the goal of "maximizing ESG mean and minimizing volatility," a GRU is used to warn of short-term risks, and an LSTM is used to control long-term trends. After adding constraints, the optimal portfolio weights are output through PSO. However, the above technical solutions have the following drawbacks: In the process of ESG risk value assessment and analysis, identification bias often occurs due to the involvement of multiple modalities in the data. Therefore, how to determine the assessment and analysis objectives of ESG risk values ​​based on the modal data of different enterprises and their correlation with other enterprises, thereby identifying the identification bias risk of different modalities, and determining targeted assessment and analysis strategies for different enterprises to improve the reliability of ESG risk value identification and processing, has become an urgent technical problem to be solved.

[0004] Therefore, there is an urgent need for an ESG risk early warning method and system based on multimodal data. Summary of the Invention

[0005] To achieve the objectives of this invention, the following technical solution is adopted: Specifically, this application provides an ESG risk early warning method based on multimodal data, which includes: S1 uses the associated data sources of the enterprise's ESG risk values ​​and the modal data types of the associated data sources to determine the associated modal data types of the enterprise. Based on the degree of overlap of the associated modal data types of different enterprises and the degree of correlation between the associated data sources of the enterprise and the associated modal data types of different enterprises, the target enterprise for the analysis of the ESG risk values ​​of the enterprise is determined. S2, based on the deviation of the analysis target enterprise from other analysis target enterprises in different modal data types of related data sources, and in combination with the related modal data types of different enterprises, conducts an ESG risk warning method for the analysis target enterprise. The risk warning method is used to determine the deviation of the ESG risk value of the analysis target enterprise between different analysis batches, and in combination with the related data sources of the enterprise in different modal data types, the analysis deviation risk type of the modal data type is determined. S3 determines the early warning management method for the ESG risk value of the enterprise based on the analysis deviation risk type of different modal data types and the overlap between the associated data source and the target enterprise of the analysis.

[0006] The beneficial effects of this invention are as follows: Based on the degree of overlap in the related modal data types of different enterprises and the degree of correlation between the related data sources of an enterprise and the related modal data types of different enterprises, the target enterprises for ESG risk value analysis are determined. Through cross-enterprise comparison, the "data types of concern" (i.e., common industry modalities) that are common in the industry are identified. Combining the degree of matching between the target enterprise's own modal type and the common industry modalities, as well as the difficulty of tracing deviations due to the amount of modal data, it is ensured that the ESG data of the finally selected target enterprises is both representative of the industry and has the possibility of accurate analysis within a controllable complexity.

[0007] Based on the types of analytical bias risks across different modal data types and the overlap between related data sources and the target company, a method for early warning management of the company's ESG risk value is determined. This method dynamically decides whether to use a preset number (more rigorous) or a basic number (more economical) of analysis batches for early warning processing, by comprehensively considering the overall modal risk (such as whether there is modal data of severe risk types and the proportion of data sources of severe risk modalities) and its similarity to data sources of known high-volatility companies (the target company). This mechanism aims to optimize computational resources and improve the accuracy and reliability of early warnings while ensuring controllable risk, and to determine the true ESG value assessment results through comparative analysis of multiple batches.

[0008] Furthermore, the associated data source is a data source used for the analysis and processing of the enterprise's ESG risk values.

[0009] Furthermore, the modal data type of the associated data source is determined based on the data type used for ESG risk value analysis and processing in the associated data source.

[0010] Furthermore, the method for determining the associated modal data type of the enterprise is as follows: Based on the modal data type of the associated data source, determine the number of associated data sources in different modal data types; Based on the associated data sources of the company's ESG risk values, determine the number of associated data sources of the company's ESG risk values; The associated modal data types of the enterprise are determined by the number of associated data sources for the enterprise's ESG risk values ​​and the number of associated data sources in different modal data types.

[0011] Furthermore, the method for determining the analysis bias risk type of the modal data type is as follows: Based on the deviation of the ESG risk values ​​of the target enterprise in different analysis batches, determine the analysis and identification process for inconsistencies in ESG risk values ​​between different analysis batches, and use it as the deviation identification process. Based on the proportion of the deviation identification process in the analysis and identification process, enterprises with identification deviation risk among the enterprises are identified; By identifying the associated data sources of enterprises with different identification bias risks in different modal data types, the analytical bias risk type of the modal data type is determined.

[0012] Secondly, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-described ESG risk warning method based on multimodal data when running the computer program.

[0013] Other features and advantages will be set forth in the following description, and the objects and other advantages of the invention are realized and obtained through the structures particularly pointed out in the description and the drawings.

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0015] The above and other features and advantages of the present invention will become more apparent from a detailed description of exemplary embodiments thereof with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart of an ESG risk early warning method based on multimodal data; Figure 2 This is a flowchart illustrating the method for determining the associated modal data types of an enterprise; Figure 3 This is a flowchart illustrating the method for determining the target company in the analysis of ESG risk values ​​within an enterprise. Detailed Implementation

[0017] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0018] Example 1 like Figure 1 As shown, this application provides an ESG risk early warning method based on multimodal data, specifically including: S1 uses the associated data sources of the enterprise's ESG risk values ​​and the modal data types of the associated data sources to determine the associated modal data types of the enterprise. Based on the degree of overlap of the associated modal data types of different enterprises and the degree of correlation between the associated data sources of the enterprise and the associated modal data types of different enterprises, the target enterprise for the analysis of the ESG risk values ​​of the enterprise is determined. Furthermore, the associated data source is a data source used for the analysis and processing of the enterprise's ESG risk values.

[0019] Furthermore, the modal data type of the associated data source is determined based on the data type used for ESG risk value analysis and processing in the associated data source.

[0020] Specifically, such as Figure 2 As shown, the method for determining the associated modal data type of the enterprise is as follows: To accurately identify which modal data types constitute the main information in an enterprise's ESG risk analysis, thus laying the foundation for subsequent targeted risk assessment, the core logic is as follows: the more ESG data sources related to a company, the more complex the information environment, and the higher the risk of contradictions, noise, and identification biases between multiple data sources. In this case, to comprehensively capture various types of information that may contain risk clues and prevent key risk signals from being submerged in the massive amount of data, a lower screening threshold should be used to include more modal data types in the category of "related modalities," so that specific bias identification and calibration can be performed on these modalities in subsequent analysis. Conversely, when the number of data sources is small and the information is relatively concentrated, a higher threshold should be used to focus on the most core modalities for in-depth analysis.

[0021] Based on the modal data type of the associated data source, determine the number of associated data sources in different modal data types; Related data sources: These refer to the data sources used for the analysis and processing of the company's ESG risk values. These data sources form the information foundation for the company's ESG risk analysis, and include, for example, the company's annual sustainability report, environmental penalty records from government environmental protection departments, negative reports from news media, and supply chain audit reports.

[0022] Modal data type: Determined based on the data type used for ESG risk value analysis in the associated data source. Modality refers to the form in which the data is presented, such as text, image, time series, audio, video, etc. An associated data source may contain a single modality or multiple modalities.

[0023] Number of associated data sources in different modal data types: This refers to the number of data sources under each modality after classifying and counting all associated data sources according to their main modalities.

[0024] This step is the foundation and starting point of the entire method. Its purpose is to structure the massive, heterogeneous ESG-related data, classifying them into different modal dimensions and completing preliminary quantitative statistics. Only after completing this basic classification and counting can subsequent proportion analysis and threshold comparisons be based on solid evidence.

[0025] Specific examples: Taking "Green Energy Technology," a listed company, as an example, its associated data sources for ESG risk analysis include: its 2024 Sustainability Report published on its official website (mainly text modality), satellite images of wastewater discharge around its factory taken by an environmental organization (image modality), financial media reports on changes in its board members (text modality), and monthly carbon emission data tables provided by its suppliers (time series modality). In step 1, the system will count: there are 2 associated data sources in the text modality data type (ESG report and financial reports), 1 associated data source in the image modality data type (satellite image), and 1 associated data source in the time series modality data type (carbon emission data table).

[0026] Based on the associated data sources of the company's ESG risk values, determine the number of associated data sources of the company's ESG risk values; Related data sources for the company's ESG risk value: This is a concept of a complete set, referring to the collection of all data sources related to the ESG risk analysis of "Green Energy Technology" that were statistically analyzed and classified in step 1.

[0027] This step aims to obtain a crucial global variable—the total number of data sources. This total number is not only used to calculate the proportion of each modality in subsequent calculations, but more importantly, it directly determines the direction of the "proportion threshold" in subsequent steps, serving as the core anchor of the entire dynamic threshold mechanism.

[0028] Specific examples: Continuing with the previous example, the number of data sources across all modalities of "Green Energy Technology" is added together: 2 (text) + 1 (image) + 1 (time series) = 4. Therefore, the number of associated data sources used for the ESG risk value analysis of "Green Energy Technology" is determined to be 4.

[0029] The associated modal data types of the enterprise are determined by the number of associated data sources for the enterprise's ESG risk values ​​and the number of associated data sources in different modal data types.

[0030] It is understood that the associated modal data types of the enterprise are determined by utilizing the number of associated data sources for the enterprise's ESG risk values ​​and the number of associated data sources in different modal data types, specifically including: The proportion of associated data sources for each modal data type is determined based on the proportion of the number of associated data sources in different modal data types to the total number of associated data sources for the enterprise's ESG risk value. Computational modal data source ratio Percentage of associated data sources: This refers to the percentage of data sources under a specific modality out of the total number of data sources across all modalities. It reflects the relative richness of information in that modality within the overall ESG data composition of the enterprise.

[0031] Absolute numbers alone cannot reflect the relative importance of modalities. A modality with 10 data sources may be less important than a modality with 5 data sources when the total number of data sources is 15. By calculating the proportion, the influence of the total size can be eliminated, and the "density" or "weight" of each modality in the information composition can be fairly measured.

[0032] Percentage is a standardized metric that makes the modal composition comparable among companies of different sizes and with varying data abundance. A high percentage of a modality indicates that the company has a large amount of information output and a wide range of sources in that dimension, and it is a key information window for understanding its ESG risk status.

[0033] Specific examples: For "Green Energy Technology", the number of associated data sources for the text modal data type is 2, and the total number of data sources is 4. Therefore, the proportion of associated data sources is 2 / 4 = 50%. Similarly, the proportion of image modal data is 1 / 4 = 25%, and the proportion of time series modal data is also 25%.

[0034] Based on the number of associated data sources for the company's ESG risk values, the percentage threshold corresponding to the associated modal data types is determined. If the proportion of associated data sources of the modal data type is greater than the proportion threshold, then the modal data type is determined to belong to the associated modal data type of the enterprise.

[0035] Percentage Threshold: A critical value used to screen key modalities. Its core characteristic is that it is dynamic and inversely proportional to the number of data sources associated with the company's ESG risk values. That is, the more data sources associated with a company, the more complex the information environment, and the higher the potential risk of identification bias, the lower this screening threshold becomes. This allows for the inclusion of more modality types in the "associated modalities," laying the foundation for subsequent targeted identification and calibration of potential biases in each modality. Conversely, when the total number of data sources is small, the information is relatively simple, and the risk of bias is low, the threshold increases to screen out the most representative core modalities for in-depth analysis.

[0036] If a fixed, static threshold, such as 30%, is used, when the total number of data sources is small (e.g., only 2), no modality may reach the 30% threshold, making it impossible to identify any related modalities and hindering analysis. Conversely, when the total number of data sources is extremely large (e.g., 1000), the information is highly complex, and many modalities may fail to reach the 30% threshold, thus overlooking many modalities that may contain risk signals and creating blind spots in the identification of deviation risks. Therefore, adopting a dynamic threshold that is inversely proportional to the total number of data sources allows the screening mechanism to adapt to the complexity of the enterprise's information and the level of deviation risk: the more complex the information and the higher the deviation risk, the more extensive the screening should be, identifying more modalities for subsequent specialized risk identification.

[0037] This mechanism is the core innovation of this invention. It endows the ESG risk analysis system with intelligence and risk sensitivity. The system can automatically adjust its judgment criteria based on the complexity of the enterprise's own information ecosystem, directly transforming the "quantity" of data sources into a prediction of "deviation risk," and adjusting the "tolerance" of modality screening accordingly. This dynamic adjustment ensures that regardless of the level of enterprise information disclosure, the system can provide the most comprehensive and appropriate modality set for subsequent targeted risk assessments, laying a solid foundation for accurately locating and identifying deviations and integrating multi-source information.

[0038] Specific examples: For "Green Energy Technology," the number of associated data sources for its ESG risk value is four. According to preset rules, when the number of data sources is small (e.g., 1-5), the information environment is relatively simple, and the risk of deviation is low; therefore, a relatively high threshold is used to ensure focus on the core. The system sets the threshold for the proportion of associated modal data types to be 30% when the number of data sources is four. The following judgment is then made: The text modality accounts for 50%, which is greater than 30%, therefore it is determined that the text modality belongs to the associated modal data type of "Green Energy Technology".

[0039] The proportion of image modalities is 25%, not more than 30%, therefore image modalities do not belong to its associated modal data types.

[0040] The proportion of time series modalities is 25%, not more than 30%, therefore time series modalities are not considered as related modal data types.

[0041] Ultimately, the system determined that the associated modal data type of "Green Energy Technology" is "text modal". This means that in subsequent risk assessments, the focus will be on in-depth analysis of its rich text data sources, with a particular emphasis on identifying potential biases within the text data.

[0042] It should be noted that the more associated data sources a company has, the smaller the percentage threshold corresponding to the associated modal data type.

[0043] Specifically, such as Figure 3 As shown, the method for determining the target companies for ESG risk value analysis in the aforementioned enterprises is as follows: This method precisely selects target companies from a large pool of enterprises suitable for in-depth ESG risk value analysis. The core logic is that not all companies possessing ESG data are suitable as analysis targets. A suitable target company should have a data modality composition that is representative of the industry and analyzable. This method first identifies commonly existing "data types of interest" (i.e., common industry modalities) through cross-enterprise comparison. Then, it combines the abundance of modal data sources for the target company (i.e., the number of data sources for each modality) and the degree of matching between its modality types and common industry modalities to comprehensively determine whether the company has the foundation for reliable ESG risk analysis. In particular, when a company has too many modalities, the complexity of cross-validation of multi-source data can lead to difficulties in tracing biases, making it unsuitable as an analysis target. Through this multi-layered screening mechanism, it ensures that the final selected analysis target companies have ESG data that is both industry-representative and capable of accurate analysis within a controllable complexity.

[0044] S21 determines the proportion of enterprises belonging to the associated modal data type under different modal data types based on the degree of overlap of the associated modal data types of different enterprises, and uses it as the proportion of matching quantities under the modal data type; Related modal data types: These refer to modal data types where, for a specific enterprise, there is at least one related data source. In other words, the enterprise possesses data available for analysis within this modality.

[0045] The degree of overlap of related modal data types among different enterprises: refers to the statistical situation of how many enterprises possess each modal data type within a specific group of enterprises (such as sample enterprises in the same industry, region, or size).

[0046] Matching Quantity Ratio: For a specific modal data type, this represents the proportion of companies within a selected group that possess data of that modality (i.e., that modality belongs to its associated modal data type), out of the total number of companies in that group. This ratio reflects the prevalence or commonality of that modality within the industry.

[0047] This step aims to establish an industry benchmark or reference system. Looking at a company's data modal composition in isolation makes it impossible to determine whether it is mainstream. By calculating the proportion of matches, the prevalence of each modality within the industry can be quantified. A higher proportion of matches for a modality indicates that it is a commonly possessed information dimension by companies in that industry. When conducting ESG risk analysis on such a modality, there are more peer references and experiences to follow, and the analysis results are easier to compare and verify horizontally.

[0048] The proportion of matches is an important industry characteristic indicator, revealing which modalities are key windows for interpreting the ESG risks of that industry. For example, in the mining industry, the proportion of matches for "image modalities" (such as satellite images of mining areas) may be high, because environmental regulation of mining areas generally relies on image evidence; while in the financial industry, the proportion of matches for "time series modalities" (such as financial data and time series of carbon emission intensity) may be even higher. This indicator provides an objective basis for subsequently judging whether a company is "typical".

[0049] Specific examples: The "Smart Evaluation Technology" ESG analysis system selected 20 representative companies in the new energy industry (including "Green Energy Technology") as samples to analyze their data possession across various modalities. The analysis revealed that 18 companies possessed text modal data, resulting in a matching rate of 18 / 20 = 90%; 10 companies possessed image modal data, with a matching rate of 10 / 20 = 50%; 5 companies possessed time-series modal data, with a matching rate of 5 / 20 = 25%; and 2 companies possessed audio / video modal data, with a matching rate of 2 / 20 = 10%.

[0050] S22 determines the number of associated data sources in different modal data types based on the associated data sources of the enterprise in different modal data types; The number of associated data sources across different modal data types: This refers to the number of specific data sources possessed under each modality for the target company being evaluated, according to the previously determined modality classification. For example, how many text data sources and how many image data sources does the company possess?

[0051] This step aims to quantify the information abundance of the target company. While the matching ratio measures the "presence" (whether the modality is prevalent), this step measures the "abundance" (how rich the information is within that modality). The number of data sources for each modality directly reflects the information density and the level of detail in the information disclosure for that dimension. More data sources mean more details to explore, richer cross-validation material, and potentially higher robustness of the analysis conclusions. However, too many data sources can also introduce noise, which needs to be considered in subsequent steps.

[0052] For the "Green Energy Technology" company to be evaluated, the system determined, based on the previous data collection results, that the number of associated data sources in the text modality was 5 (sustainable development report, 3 in-depth reports, 1 industry overview), the number of associated data sources in the image modality was 2 (aerial photos, award ceremony pictures), the number of associated data sources in the time series modality was 1 (carbon emission data), and the number of associated data sources in the audio and video modality was 0.

[0053] S23 uses the proportion of matching data under different modal data types, and the number of associated data sources of the enterprise in different modal data types with different matching proportions, to determine whether the enterprise is a target enterprise for ESG risk value analysis.

[0054] Specifically, in the above steps, modal data types with a matching ratio greater than a preset matching ratio threshold are designated as data types of interest. If no data type of interest exists in the associated data types of the enterprise, it is difficult to determine the reliability of the parsing and processing of the enterprise under the data type of interest based solely on the analysis and processing of the enterprise's ESG risk value. Therefore, it is determined that the enterprise does not belong to the target enterprise for ESG risk value analysis.

[0055] First, the system presets a threshold for the proportion of matches (e.g., 50%). Modal data types with a match proportion greater than this threshold are defined as data types of interest.

[0056] Focus on data types: These refer to modal data types that are highly prevalent within the industry (high matching rate). These modalities represent the mainstream information dimensions of the industry, making risk analysis of them more comparable and valuable for industry analysis.

[0057] Setting the data types of interest is to filter out modalities that are representative of the industry. If a company lacks even the modalities commonly found in the industry, its data composition deviates from industry norms. ESG risk analysis based on its existing data may lead to conclusions that are difficult to align with industry standards, significantly reducing the reliability and universality of the analysis. The data types of interest constitute the "standard modality set" for industry ESG analysis. It serves both as the basis for selecting companies and as the priority direction for subsequent multimodal fusion analysis.

[0058] Specific examples: The system sets the matching ratio threshold to 50%. Based on the calculation results of step S21, the matching ratios of text modality (90%) and image modality (50%) are both greater than or equal to 50%, and therefore are identified as data types of interest. Time series modality (25%) and audio / video modality (10%) are not considered data types of interest.

[0059] Case A: No data type of interest exists. If no data types of interest are found in the company's associated data types, it indicates that the company's data modality deviates significantly from industry norms. In this case, relying solely on the company's existing ESG data makes it difficult to determine the information the company should present under those industry-wide modes of interest (i.e., data types of interest), thus making it impossible to assess its relative ESG performance within the industry, and compromising the reliability of the analytical conclusions. Therefore, the company is determined not to be a target company for ESG risk value analysis.

[0060] Suppose another "old energy company" has only time-series modal data sources (such as historical production data), while the industry's focus data types are text and images. Since this company does not have any data types of interest, the system will directly determine that it does not belong to the target enterprise of the analysis.

[0061] It should be noted that the associated data type of the enterprise is a modal data type with associated data sources.

[0062] Additionally, it should be noted that if the associated data types of the enterprise include data types of interest, the following content is also included: Scenario 1: If the total number of related data types of the enterprise is obtained and it is determined that the total number of related data types of the enterprise is greater than the preset threshold for the number of related data types, then when performing early warning analysis of ESG risk values, if the same data is analyzed and the ESG risk values ​​of different batches are inconsistent, it is difficult to determine under which related data types there is identification bias risk. Therefore, it is determined that the enterprise is not a target enterprise for ESG risk value analysis. Scenario B: Data type of interest exists: If at least one data type of interest exists among the enterprise's associated data types, further segmentation is required. At this point, the total number of associated data types for the enterprise needs to be determined, i.e., the number of modal types that the enterprise actually owns in its data sources (regardless of whether the modality belongs to a data type of interest).

[0063] Case B1: Total number of associated data types > Preset threshold for the number of associated data types When the total number of related data types for a company is excessively large (exceeding a preset threshold, such as 3), it indicates that the company's ESG information modalities are extremely rich and the data dimensions are numerous. In this case, conducting ESG risk value early warning analysis on the company often requires cross-validation by integrating data from multiple sources. However, if inconsistencies in ESG risk values ​​appear between different batches of analysis results, the large number of modalities involved and the complex relationships between data sources make it difficult to quickly and accurately pinpoint which one or more related data types' identification bias led to the difference in the final results. This high difficulty in tracing the source of bias reduces the controllability and reliability of the analysis process. Therefore, in such cases, it is determined that the company is not a target company for ESG risk value analysis.

[0064] Preset threshold for the number of associated data types: A pre-defined empirical value used to determine whether the modality richness of an enterprise is within a controllable range. Exceeding this value indicates that too many modalities may lead to unacceptable levels of analytical complexity and difficulty in tracing the source of biases.

[0065] Identification bias risk: refers to the risk that, during multi-source data analysis, errors may occur in the interpretation of information about a particular modality due to quality issues, acquisition errors, or insufficient model applicability, thereby affecting the overall risk assessment.

[0066] This situation reflects the consideration of the controllability of the analysis process. ESG risk analysis pursues accuracy and interpretability. When the number of modalities is too large, although the information is comprehensive, the system complexity increases dramatically, and any small deviation may be amplified and difficult to locate, thus reducing the reliability of the analysis results. Setting this threshold is to achieve a balance between the "comprehensiveness" and "controllability" of information.

[0067] This mechanism prevents the analysis system from falling into the "data overload" trap, ensuring that the data modal complexity of the selected target companies is within the range that existing analysis techniques and deviation tracing capabilities can handle, thereby guaranteeing the stability and interpretability of the analysis conclusions.

[0068] Assume that "Wanguo Mining" has four types of associated data: text, image, time series, and audio / video. The system's preset threshold for the number of associated data types is 3. Since 4 > 3, the system determines that the number of modalities is too large. If the analysis results of different batches are inconsistent, it will be difficult to pinpoint which modality the deviation originates from. Therefore, the system determines that "Wanguo Mining" is not a target company for analysis.

[0069] Case 2: If the total number of related data types of the enterprise is not greater than the preset threshold for the number of related data types, then it is determined that all related data types of the enterprise belong to the data types of interest, and therefore the enterprise is determined to be the target enterprise for ESG risk value analysis. Case B2: The total number of related data types is less than or equal to the preset threshold for the number of related data types, and all related data types of the enterprise are data types of interest. If the total number of related data types for a company does not exceed a threshold, and all of these types are data types of interest, it indicates that the company's data modality composition is neither overly complex nor does it fully conform to industry norms. In this case, the company possesses the ideal foundation for ESG analysis: a moderate number of modalities, and each modality is representative of the industry. Therefore, this company can be directly identified as the target company for ESG risk value analysis.

[0070] "Green Energy Technology" has two associated data types: text and images, both of which are less than or equal to a threshold of 3, and both types are considered data types of interest (text 90%, images 50%). The system determines that it meets the ideal conditions and identifies "Green Energy Technology" as a target company for ESG risk value analysis.

[0071] Scenario 3: If the associated data types of the enterprise are not all of the data types of interest, the proportion of matching numbers in the data types of interest will be used as the weight value of the data types of interest. Based on the sum of the weight values ​​of the proportion of the enterprise in different data types of interest, a weighting ratio will be determined. It will be determined whether the weighting ratio is greater than a preset weighting ratio threshold. If it is, the enterprise will be determined as a target enterprise for ESG risk value analysis. If not, the enterprise will be determined as not a target enterprise for ESG risk value analysis.

[0072] Scenario B3: The total number of related data types is less than or equal to the preset threshold for the number of related data types, but not all of the enterprise's related data types belong to the data types of interest (i.e., it includes at least one non-interested data type): This is the most common scenario: a company possesses some mainstream industry modalities, as well as some non-mainstream, personalized modalities. In this case, a more refined assessment of its overall analytical value is needed. The specific method is as follows: the proportion of matches within the data type of interest is used as the weight value for that data type. Then, a weighted ratio is determined based on the sum of the weight values ​​of the company's proportions across different data types of interest. Finally, it is determined whether this weighted ratio is greater than a preset weighted ratio threshold. If yes, the company is identified as the target company for analysis; otherwise, it is determined not to be one.

[0073] Weighting: Here, the proportion of matches for the data type itself is used directly as the weight. The higher the proportion of matches, the more prevalent the modality is in the industry, and the greater its value for industry benchmarking and commonality analysis. Therefore, it should occupy a higher weight in the comprehensive evaluation.

[0074] Weighted Ratio: This refers to the sum of the weight values ​​of all data types of interest to the company. It quantifies the company's overall coverage of common industry information dimensions. Since non-interest data types are not included (weight is 0), this ratio actually measures the degree of alignment between the company's data modality composition and the core industry modality.

[0075] Preset weighting threshold: A critical value used for final decision-making. If this value is exceeded, the company is considered to have a sufficiently high level of coverage of common industry modalities, and even if some non-mainstream modalities exist, it will not affect their value as analytical targets.

[0076] When a company has both mainstream and non-mainstream modalities, a simplistic "one-size-fits-all" approach is inappropriate. Mainstream modalities provide a basis for benchmarking against the industry, while non-mainstream modalities may contain unique risk information. Weighted summation can quantify the overall contribution of the mainstream modality. If the weighted sum of the mainstream modalities is sufficiently high, it indicates that the majority of the company's data is still industry-wide commonalities, and the analysis results have broad reference value. Conversely, if the weighted sum of the mainstream modalities is very low, it indicates that the data is dominated by personalized modalities, has poor industry comparability, and is unsuitable as a regular analytical target.

[0077] This mechanism enables a "grayscale" assessment of enterprise data composition, avoiding black-and-white judgments and allowing more enterprises to be included in the analysis while ensuring the quality of the analysis. It cleverly combines industry universality with enterprise specificity, providing a quantitative and interpretable basis for the selection of analysis targets.

[0078] Specific examples: Assume that "Laneng Technology" has two associated data types (≤threshold 3): text modality (a data type of interest, matching rate 90%, 8 data sources) and audio / video modality (a data type of non-interest, matching rate 10%, 3 data sources). Not all of its associated data types belong to the data type of interest. The system calculates the weighted ratio as follows: 90% weight of the text modality multiplied by 8 / 11 = 65.4% (audio / video modality is not included). With a preset weighted ratio threshold of 80%, the system determines that "Laneng Technology" does not belong to the target company for analysis.

[0079] S2, based on the deviation of the analysis target enterprise from other analysis target enterprises in different modal data types of related data sources, and in combination with the related modal data types of different enterprises, conducts an ESG risk warning method for the analysis target enterprise. The risk warning method is used to determine the deviation of the ESG risk value of the analysis target enterprise between different analysis batches, and in combination with the related data sources of the enterprise in different modal data types, the analysis deviation risk type of the modal data type is determined. Specifically, the method for determining the ESG risk warning method for the target enterprise is as follows: For the selected target companies, a differentiated ESG risk early warning method is tailored to their specific needs. The core logic lies in the fact that different companies have varying data modalities, different rankings in data abundance within their industry, and different degrees of alignment with mainstream industry modalities. Therefore, their sensitivity to changes in ESG risks and the urgency of monitoring also differ. This method dynamically determines the most suitable early warning trigger conditions for each company by comprehensively considering its ranking in the number of data sources for each type of data of interest (i.e., its relative position in information richness within the industry) and the degree of overlap between its own data modality and the industry's data types of interest. For companies with abundant information and a high degree of alignment with mainstream industry data, the most sensitive early warning strategy is adopted (analysis is triggered upon any data update) to ensure timely detection of anomalies. For companies with relatively marginal information or a single modality, a more robust early warning strategy is adopted (triggered only by batch updates or updates to key modalities) to balance analytical resources and early warning accuracy. Through this refined and adaptive early warning mechanism, the efficiency and effectiveness of ESG risk monitoring are maximized.

[0080] S31 uses the associated modal data types of different enterprises to determine the data types of interest among the modal data types; Focus on data types: These refer to modal data types that are highly prevalent within a specific group of enterprises (such as those in the same industry or within the same assessment pool). They are typically composed of modalities whose matching rate exceeds a preset threshold. These modalities represent the mainstream forms of ESG information expression within that group of enterprises.

[0081] Focusing on data types serves as the benchmark for all subsequent analyses. Before customizing early warning methods, it is essential to first identify which modalities are recognized as key information dimensions within the industry or assessment group. This step elevates the analytical framework from the individual level to the group level, ensuring that subsequent ranking and overlap analyses have a unified reference point.

[0082] Focusing on data types defines the "core modal set" of ESG risk assessment, making it possible to compare different companies on the same dimension, and providing a benchmark for judging the "typicality" of a company's data composition.

[0083] Specific examples: "Smart Evaluation Technology's" ESG system calculates the matching ratio of each modality based on historical data from its target enterprise pool (including 10 companies such as Green Energy Technology, Blue Energy Technology, New Energy Technology, Trina Solar, and Ruiguang Energy), and determines the data types of interest as text modality and image modality according to steps S21 to S23.

[0084] S32 determines the number of associated data sources for the target enterprise under different data types of interest based on the deviation between the target enterprise and other target enterprises in different modal data types, and uses the number of associated data sources to determine the ranking result of the target enterprise under different data types of interest. Other target companies in the analysis: This refers to all other companies in the same target pool as the company being evaluated, forming the benchmark group for comparison.

[0085] Bias in data sources: This refers to quantifying the relative information abundance of a target company by comparing the number of data sources used by the target company with those of other companies across different data types of interest. This comparison is ultimately reflected in a ranking.

[0086] Ranking Results: This refers to the ranking of all target companies within a specific data type of interest, based on the number of related data sources they possess, from most to least. A higher ranking indicates greater information richness for that company within that modality, meaning it has broader data sources and a larger volume of information.

[0087] Knowing only the number of a company's own data sources is insufficient; it must be placed within a group of peers to determine whether its information richness is leading, average, or lagging. Ranking results directly reflect a company's "data position" within that modality, and this relative position is crucial for determining the sensitivity of early warnings: the more information-rich a company is, the more significant any subtle changes may contain important signals, requiring more sensitive monitoring.

[0088] The ranking results transform absolute quantities into relative positions, eliminating the impact of orders-of-magnitude differences between different modalities and making the rankings across different modalities comparable. This provides a crucial, standardized input dimension for the development of subsequent early warning strategies.

[0089] The system analyzes five representative companies in the target enterprise pool (Green Energy Technology, Blue Energy Technology, New Energy Technology, Trina Solar, and Ruiguang Energy), and counts and sorts the number of data sources for each company under the data type of interest. Number of text modal data sources: New Energy Technology (12), Blue Energy Technology (8), Green Energy Technology (5), Ruiguang Energy (4), Trina Solar (3). Ranking results: New Energy Technology 1st, Blue Energy Technology 2nd, Green Energy Technology 3rd, Ruiguang Energy 4th, Trina Solar 5th.

[0090] Number of image modal data sources: New Energy Technology (4), Green Energy Technology (2), Trina Solar (1), Blue Energy Technology (0), Ruiguang Energy (0). Ranking results: New Energy Technology 1st, Green Energy Technology 2nd, Trina Solar 3rd, Blue Energy Technology 4th (tied), Ruiguang Energy 4th (tied).

[0091] S33 uses the sorting results under different data types of interest, as well as the degree of overlap between the related data types and the data types of interest of the target enterprise, to determine the ESG risk warning method for the target enterprise.

[0092] It should be noted that the sorting results are obtained by sorting the number of associated data sources of the target enterprise under the data type of interest from most to least.

[0093] Specifically, the degree of overlap between the related data types and the data types of interest of the target enterprise is determined based on the proportion of the data types of interest among the related data types of the target enterprise.

[0094] Furthermore, by utilizing the ranking results under different data types of interest, and the degree of overlap between the related data types and the data types of interest of the target enterprise, the ESG risk warning method for the target enterprise is determined, specifically including: Scenario 1: Based on the degree of overlap between the associated data types and the data types of interest of the target enterprise, the proportion of the data types of interest in the associated data types is determined and used as the proportion of the data types of interest. When the proportion of the data types of interest is greater than the preset proportion threshold, in order to improve the timeliness of the identification and handling of anomalies in the data types of interest, the ESG risk warning method for the target enterprise is determined to be to perform ESG risk value analysis and identification processing whenever the target enterprise updates any associated data source, and to determine whether to issue a warning message based on the analysis and identification processing results.

[0095] When a company's data type of concern accounts for more than 60% of its data, it indicates that its data modality is highly concentrated in the mainstream dimensions of the industry. At this point, to improve the timeliness of identifying and handling potential anomalies in the data types of concern, any subtle changes warrant attention. Therefore, the method for determining the company's ESG risk warning is as follows: whenever the target company updates any related data source, ESG risk value analysis and identification are performed, and a warning message is issued based on the analysis and identification results.

[0096] Green Energy Technology: Its associated data types are text, image, and time series, totaling three. Among them, text and image are two data types of concern, with the proportion of concern types being approximately 66.7% (2 / 3) > 60%, triggering condition 1. Therefore, the most sensitive early warning is adopted for Green Energy Technology: any update to any associated data source (including non-concerned modalities such as time series) will immediately trigger ESG risk analysis.

[0097] Xinyuan Technology: The associated data types are text, image, and time series, totaling 3. The data types of focus are text and image, accounting for approximately 66.7% > 60%. For the same trigger condition 1, the most sensitive early warning is adopted.

[0098] Trina Energy: The associated data types are text and image, both of which are data types of interest. The ratio is 100% > 60%. Trigger condition 1, using the most sensitive early warning.

[0099] Scenario 2: If the proportion of the attention type is not greater than the preset attention proportion threshold, then the ranking result of the target enterprise under different attention data types is determined. When the ranking result of the target enterprise under any attention data type is before the target order, the ESG risk warning method of the target enterprise is determined to be that as long as the target enterprise is updated in any related data source, the ESG risk value analysis and identification processing is performed, and the warning information is determined based on the analysis and identification processing result.

[0100] When a company's focus type ratio is no higher than 60% (i.e., its modal composition is not typical), but its data source quantity ranking is very high (e.g., top 2) under a certain focus type, it indicates that the company has a significant information advantage in a key dimension. Any updates it makes in this dimension may have a significant impact on the judgment of the entire industry. Therefore, the most sensitive early warning strategy is also required.

[0101] LanNeng Technology: The associated data types are text and audio / video, totaling 2. Among them, text is the data type of interest, with the interest type ratio being 1 / 2 = 50% ≤ 60%. Its text modality ranks 2nd, within the target order (top 2), therefore trigger condition 2, which also adopts the most sensitive alert: any update to the associated data source will trigger the analysis.

[0102] Scenario 3: When the ranking result of the target enterprise under any data type of interest is not before the target order, the average ranking ratio under different data types of interest is determined based on the ranking results under different data types of interest. It is then determined whether the average ranking ratio under different data types of interest is less than a preset ranking ratio threshold. If so, the ESG risk warning method for the target enterprise is to perform ESG risk value analysis and identification processing whenever the target enterprise has more than a preset number of related data sources updated or when the related data sources of the data type of interest are updated. Based on the analysis and identification processing results, it is determined whether to issue a warning message. If not, proceed to the next step. Based on the ranking results under different data types of interest, determine the average ranking proportion under each data type of interest. The ranking proportion is defined as the company's ranking under a certain data type of interest divided by the total number of target companies analyzed; a higher value indicates a lower ranking. Determine whether this average value exceeds a preset ranking proportion threshold.

[0103] Ranking ratio: The standardized ranking eliminates the influence of the total number of companies. The value range is (0,1], and the larger the value, the lower the ranking.

[0104] Average Ranking Proportion: The arithmetic mean of the ranking proportions across all data types of interest to the enterprise, comprehensively reflecting the enterprise's average data position across all core dimensions.

[0105] When a company's overall ranking is low, its information richness is at a low level within the industry. If its average ranking percentage is greater than a threshold, it indicates that its ranking is indeed very low (e.g., all at the bottom). In this case, a more conservative early warning strategy should be adopted to avoid frequently triggering invalid analyses due to sporadic updates. If the average ranking percentage is not greater than the threshold, it means that although the ranking is not high, it is not extremely low overall, and there is still value in further evaluation.

[0106] The average ranking percentage provides a comprehensive positional indicator, helping the system to further differentiate between relatively good and relatively bad companies within the "overall lower" group, thereby enabling more granular strategy stratification.

[0107] Consider Ruiguang Energy. Its associated data types are text and time series, totaling two. One of the data types of interest is text, with a proportion of 1 / 2 = 50% ≤ 60%. Its text modality ranks 4th, not in the top 2; the image modality has no data source, therefore no ranking under any data type of interest. Its ranking proportion under its only data type of interest is 4 / 5 = 0.8. The average is 0.8. The preset ranking proportion threshold is 0.4; 0.8 > 0.4, satisfying the threshold, thus directly triggering the conclusion of situation 3: the ESG risk warning method is a compromise strategy (triggered by updates to associated data sources or data types of interest exceeding a preset number (e.g., more than 2)). In this case, sub-step 3.2 is unnecessary.

[0108] Based on the average ranking ratio under different data types of interest and the average ratio of interest types, the analysis and identification matching value of the target enterprise is determined. It is then determined whether the analysis and identification matching value of the target enterprise is greater than a preset matching threshold. If so, the ESG risk warning method for the target enterprise is to perform ESG risk value analysis and identification processing whenever any related data source of the target enterprise is updated, and determine whether to issue a warning message based on the analysis and identification processing result. If not, the ESG risk warning method for the target enterprise is to perform ESG risk value analysis and identification processing whenever the target enterprise has more than a preset number of related data sources updated or when related data sources of the data type of interest are updated, and determine whether to issue a warning message based on the analysis and identification processing result.

[0109] If the average ranking ratio in sub-step 3.1 is greater than the preset ranking ratio threshold, the ESG risk warning method for the enterprise is directly determined as follows: whenever the target enterprise is updated in more than a preset number of related data sources or when the related data sources of the data type of interest are updated, the ESG risk value analysis and identification process is carried out, and the warning information is determined based on the analysis and identification process results.

[0110] If the matched value identified in sub-step 3.2 is greater than the preset matched threshold, then the ESG risk warning method for the enterprise is determined as follows: whenever the target enterprise is updated in any related data source, the ESG risk value is analyzed and identified (i.e., upgraded to the most sensitive warning).

[0111] If the matched value identified in sub-step 3.2 is not greater than the preset matching threshold, then the ESG risk warning method for the enterprise is still the above-mentioned compromise strategy (updated when the preset quantity is exceeded or when the data type of the enterprise is updated).

[0112] Preset Quantity: A pre-defined integer threshold representing the minimum number of data source updates required to trigger an analysis. When there are sporadic updates to non-critical modalities, they may be considered noise and need to accumulate to a certain number before triggering, in order to filter out invalid fluctuations.

[0113] Updates to associated data sources for the data type being monitored: This is a special case of an "OR" condition, meaning that as long as there is an update to any data source under the monitored data type, even if there is only one, the analysis will be triggered immediately, reflecting the priority focus on the core modality.

[0114] For this type of enterprise, we cannot analyze every slight movement like the most sensitive enterprises (which would be a waste of resources), nor can we completely ignore them. Setting a "batch update" condition can filter out noise, while the "focus on data type updates" condition ensures that core information is not missed. This compromise strategy achieves optimal monitoring of peripheral enterprises with limited resources.

[0115] This strategy balances the timeliness of early warnings with the effectiveness of computing resources, ensuring that the most noteworthy changes (batch changes or core modal changes) are processed in a timely manner, while avoiding the waste of resources caused by frequent analysis starts due to sporadic non-core data updates.

[0116] Ruiguang Energy (ranking average 0.8 > 0.4) directly adopts a compromise strategy: when two or more related data sources are updated, or when any one data source under the text modality (its only data type of interest) is updated, ESG risk value analysis and identification processing is performed.

[0117] Huaguang Energy (assuming its average ranking ratio is 0.3 ≤ 0.4, and the analysis identifies a matching value of 0.35 > 0.2) will use the most sensitive alert: any update to the associated data source will trigger the analysis.

[0118] It should be noted that the analysis batch is determined based on different analysis batches within the same ESG analysis and identification process.

[0119] Specifically, the method for determining the analysis bias risk type of the modal data type is as follows: In multi-batch ESG risk analysis, identifying which modal data types are more likely to cause bias in the analysis results provides precise targeting for subsequent model optimization and data governance. The core logic is as follows: by comparing the ESG risk values ​​of the same company across different analysis batches, companies with frequently inconsistent risk values ​​are identified—these are the "bias risk companies." Then, using these companies as clues, the common modal data types are traced, and the concentration of bias risk companies under each modality is statistically analyzed to determine the analysis bias risk level (no risk, moderate risk, severe risk) for that modality data type. This mechanism helps the system pinpoint the root cause of the bias, achieving closed-loop optimization of the multimodal analysis process.

[0120] S41 uses the deviation of the ESG risk values ​​of the target enterprise in different analysis batches to determine the analysis identification and processing process for inconsistent ESG risk values ​​between different analysis batches, and uses it as the deviation identification and processing process. Analysis batch: Refers to different analysis batches within the same ESG analysis and identification process. For example, the system may conduct a complete ESG risk assessment on the same batch of companies weekly or monthly. As an analysis and identification process, each assessment includes more than 3 analysis batches, that is, the assessment is repeated more than 3 times to determine the risk of identification bias.

[0121] ESG risk value deviation refers to the difference between the ESG risk values ​​calculated in different batches of analysis for the same target company. This difference can only be caused by the deviation of the model identification results in the data source, that is, some modal data itself has unstable or biased characteristics.

[0122] Deviation identification and processing: This refers to the analytical process for specific batches where ESG risk values ​​differ between batches. Specifically, if the results of a particular analytical batch are inconsistent with the results of a previous batch, then that batch can be marked as a deviation identification and processing step relative to the previous batch.

[0123] This step is fundamental to identifying the source of deviations. By tracking the fluctuations in the ESG risk values ​​of the same company over time, we can pinpoint analytical events that exhibit "abnormal fluctuations." These events serve as primary evidence for subsequent assessments of whether the company faces systemic identification bias risks.

[0124] By concretizing the abstract concept of "deviation" into a statistically verifiable "deviation identification and processing procedure," an operational indicator is provided for quantifying the stability of a company's risk analysis. By statistically analyzing the frequency of these processes, a company's consistency across multiple batches of analysis can be objectively assessed.

[0125] Over the past year, "Smart Assessment Technology's" ESG system conducted a total of 10 ESG risk assessments on five target companies (Green Energy Technology, Blue Energy Technology, New Energy Technology, Trina Solar, and Ruiguang Energy). The system compared the risk values ​​of each company across adjacent batches and recorded any inconsistencies. For example, for Green Energy Technology, significant differences were observed in the results of the 3rd, 6th, and 9th batches of the assessment analysis process; therefore, Green Energy Technology was marked as having 3 deviation identification processes. Similarly, the number of deviation identification processes for other companies was also counted based on their actual fluctuations.

[0126] S42 determines the enterprises with identification deviation risk among the enterprises based on the proportion of the deviation identification process in the analysis identification process; Specifically, if the proportion of the deviation identification process in the analysis and identification process of the enterprise is greater than the preset deviation process proportion threshold, then the enterprise is determined to be an enterprise with deviation risk.

[0127] Analysis and identification process: This refers to the total number of analysis batches performed on this company. For example, if the system performs 10 analyses on a company, then the total number of analysis and identification processes is 10.

[0128] Deviation identification processing percentage: This refers to the percentage of deviation identification processing steps performed by the company out of the total number of analysis identification processing steps. This percentage reflects the degree of instability in the company's results across multiple batch analyses.

[0129] Preset deviation percentage threshold: A pre-defined critical percentage. When a company's deviation percentage exceeds this threshold, it indicates that the company's analysis results are frequently inconsistent, its risk value has low reliability, and there is a high probability of systematic identification bias. Therefore, it is identified as a "company with identification bias risk".

[0130] Companies at high risk of identification bias: These are companies whose results in multiple ESG risk analyses exhibit volatility exceeding the normal range, indicating a high risk of identification bias. These companies are the focus of subsequent modal deviation analysis.

[0131] The mere presence of deviation events is insufficient to determine if a company has systemic problems; the frequency of these events must also be considered. By setting a percentage threshold, "random deviations" and "systematic deviations" can be distinguished. Only companies exhibiting systematic deviations deserve to be flagged as high-risk companies and used for subsequent modal attribution analysis.

[0132] Identifying companies at risk of deviation forms a bridge connecting "outcome deviation" and "causal modality." These companies act like "probes," and their common characteristics (i.e., shared modal data types) guide us to find the modalities that may lead to deviation.

[0133] S43 determines the analytical bias risk type of the modal data type by identifying the associated data sources of enterprises with different identification bias risks in different modal data types.

[0134] Specifically, if there are no enterprises with identification bias risk among the enterprises of the associated data type, then the analysis bias risk type of the modal data type is determined to be a risk-free type.

[0135] Additionally, it is understood that if there are enterprises at risk of identification bias among the enterprises whose associated data type exists in the modal data type, this includes the following: Enterprises whose associated data types exist in the modal data type are considered as associated enterprises. It is determined whether the proportion of enterprises with identification bias risk among the associated enterprises is greater than a preset risk enterprise proportion threshold. If so, the analysis bias risk type of the modal data type is determined to be a severe risk type. If not, the analysis bias risk type of the modal data type is determined to be a general risk type.

[0136] Related companies: refers to all target companies that possess a specific modality data type (i.e., that modality belongs to its related data type).

[0137] The percentage of firms at risk of identification bias: This refers to the percentage of firms in a given modality that are marked as having an identification bias risk, out of all firms in that modality. This percentage reflects the strength of the association between that modality and identification bias.

[0138] Preset risk enterprise ratio threshold: A pre-set critical percentage used to determine whether modal risk is "severe" or "moderate".

[0139] Risk-free type: This means that under this modality, no related company is at risk of identification bias. This indicates that the data source for this modality is stable across multiple batches of analysis and is unlikely to cause identification bias.

[0140] General risk type: This refers to a mode where there are companies at risk of identification bias, but the proportion does not exceed a preset threshold. This indicates that the mode may be related to bias, but the correlation is weak, or only a few companies have problems in this mode.

[0141] Severe Risk Type: This refers to a modality where the proportion of companies at risk of identification bias exceeds a preset threshold. This indicates a high correlation between this modality and identification bias, and it is likely a key source of inconsistencies in results across multiple analysis batches, requiring priority for remediation and model optimization.

[0142] By statistically analyzing the proportion of firms exhibiting "identification bias" in each modality, the "identification bias risk" of that modality can be quantified. If all firms in a given modality are risk-free, then that modality is undoubtedly risk-free. If identification bias exists, the concentration of the bias needs to be examined: high concentration (exceeding a threshold) indicates that the modality is a high-risk type; low concentration indicates that the risk is relatively dispersed and may be influenced by other factors.

[0143] This step completes the attribution analysis from "enterprise-level deviations" to "modal-level risks," providing a clear priority for system optimization: first, focus should be placed on modalities identified as "severe risk types," thoroughly examining their data source quality, feature extraction methods, or model suitability; second, focus should be placed on modalities of "general risk types"; and for "risk-free types," the existing processing procedures can be maintained. This hierarchical management greatly improves the efficiency of model iteration and governance.

[0144] Based on the previous data types of enterprise associations (using the previous settings) and the results of identifying enterprises with risk of bias: The text modality is associated with five companies: Green Energy Technology (risk), Blue Energy Technology (non-risk), New Energy Technology (risk), Trina Solar (non-risk), and Ruiguang Energy (risk). Of these, three are at risk (Green Energy, New Energy, and Ruiguang), representing 3 / 5 = 60%.

[0145] The companies associated with the image modality are: Green Energy Technology (risk), Xinyuan Technology (risk), and Trina Energy (non-risk), totaling 3 companies. Among them, 2 are risky companies (Green Energy and Xinyuan), accounting for approximately 66.7%.

[0146] The related companies in the time series modality are: Green Energy Technology (risk), Xinyuan Technology (risk), and Ruiguang Energy (risk), totaling 3 companies. Among them, 3 companies are at risk (Green Energy, Xinyuan Technology, and Ruiguang Energy), representing 100% of the total.

[0147] Related companies in the audio and video modality sector: Laneng Technology (non-risk), 1 company in total. There are 0 risky companies, representing 0% of the total.

[0148] The preset risk enterprise ratio threshold is set to 50%. Then: Text modality: 60% > 50%, belongs to the severe risk type; Image modality: 66.7% > 50%, belongs to the severe risk type; Time series modality: 100% > 50%, belongs to the severe risk type; Audio and video modality: 0% ≤ 50%, and there are no risky enterprises, belongs to the no-risk type (because there are no enterprises with identification bias risk).

[0149] S3 determines the early warning management method for the ESG risk value of the enterprise based on the analysis deviation risk type of different modal data types and the overlap between the associated data source and the target enterprise of the analysis.

[0150] Specifically, the method for determining the early warning management method for the enterprise's ESG risk value is as follows: For different types of enterprises (regardless of whether they are the target of the analysis), a scientific and efficient ESG risk early warning management method is developed, particularly determining the number of analysis batches to be executed in the same analysis and identification process. The core logic is that the rigor of the early warning should match the enterprise's potential risk level and data characteristics. By comprehensively considering the overall modal risk (such as whether there is modal data of severe risk types, and the proportion of data sources for severe risk modalities) and its similarity to the data sources of known high-volatility enterprises (the target of the analysis), the system dynamically decides whether to use a preset number (more rigorous) or a basic number (more economical) of analysis batches for early warning processing. This mechanism aims to ensure that, under the premise of controllable risk, computational resources are optimized, the accuracy and reliability of early warnings are improved, and the evaluation analysis results of the true ESG value are determined through comparative analysis of multiple batches.

[0151] S51 determines the modal data type of severe risk based on the analysis of deviation risk types in different modal data types; Analysis bias risk type: This refers to the risk level assigned to each modal data type based on bias tracing from previous batch analyses, including no-risk, moderate-risk, and severe-risk types. It reflects the likelihood that the modal data will cause inconsistencies in results across multiple batch analyses.

[0152] High-risk modal data types: These are modalities that are identified as having the highest level of deviation risk and require special attention and priority management.

[0153] This step is fundamental to the entire early warning management methodology. Only by identifying which modalities are "high-risk" can it be determined whether a company is exposed to these risks. It extracts the most critical risk dimensions from a macro-level modal risk map, and the data types of modalities with severe risks constitute the "risk sources" for subsequent risk assessments. Any company that possesses data on these modalities needs to consider the analytical uncertainties that these modalities may bring to its early warning management methodology.

[0154] Specific examples: Based on the results of the analysis of the risk type of modal data type in the previous steps of "Intelligent Evaluation Technology" (see the previous example for details), it is determined that: the modal data types with serious risks include: text modality, image modality, and time series modality, and the risk-free type is: audio and video modality.

[0155] S52 determines the number of overlaps between the associated data source of the enterprise and the target enterprise of the analysis based on the overlap between the associated data source and different target enterprises of the analysis, and determines the data source similarity coefficient between the associated data source and different target enterprises of the analysis based on the number of overlaps; Related data sources: refers to all data sources owned by the company that are used for ESG risk analysis.

[0156] Overlapping data: This refers to the number of data sources that a company's related data sources share with the related data sources of a target company in the analysis.

[0157] Data source similarity coefficient: A metric that quantifies the degree of similarity between a given company and a target company in an analysis at the data source level. It is calculated as follows: Data source similarity coefficient = Number of overlapping data sources / (Total number of related data sources for the given company + Total number of related data sources for the target company - Number of overlapping data sources) (i.e., Jaccard similarity coefficient). The coefficient value ranges from 0 to 1; a higher value indicates that the two companies share more data sources and have a more similar foundation of ESG information.

[0158] If a company shares a large amount of the same data sources as companies identified as "identification bias risk companies" (i.e., companies with high volatility among the target companies in the analysis), then that company may also suffer from the same data source quality issues, thus facing similar identification bias risks. The likelihood of this "risk contagion" can be quantified by calculating a similarity coefficient.

[0159] Data source similarity coefficients bridge the gap between a company and a known risk pool. High similarity means that the company may have inherited some risk characteristics from the target company being analyzed, thus requiring greater caution in early warning management.

[0160] Consider a new company, "Huaguang Energy," which is not currently the target company for "Zhiping Technology's" analysis. Its related data sources include: a text-based report titled "2024 Sustainable Development Report," a "financial media in-depth report," and an "industry association analysis report"; an image-based aerial photograph of the factory; and a time-series data set of quarterly carbon emissions. There are a total of five data sources.

[0161] The similarity coefficients between Huaguang Energy and the data sources of each target company in the analysis are now calculated (the list of data sources for the target companies in the analysis follows the previous settings; the specific list is omitted, but the composition of the data sources for each company is known, and the number of overlaps can be calculated): With Green Energy Technology: There are 2 overlaps (sustainable development reports, carbon emission data). Green Energy Technology has a total of 8 data sources, while Huaguang Energy has 5. The union of these data sources is 8 + 5 - 2 = 11. The similarity coefficient is 2 / 11 ≈ 0.18.

[0162] The calculated similarity coefficients are as follows: with Green Energy Technology: 0.18, with Blue Energy Technology: 0.12, with New Energy Technology: 0.25, with Trina Solar: 0.10, with Ruiguang Energy: 0.20.

[0163] S53 utilizes the modal data type of the severe risk type and the data source similarity coefficient between the enterprise and different analysis target enterprises to determine the early warning management method of the enterprise's ESG risk value.

[0164] Furthermore, if the enterprise belongs to the target enterprise of the analysis or the proportion of related data sources of the modal data type with severe risk in the related data sources of the enterprise is greater than the preset risk data source proportion threshold, then the early warning management method for the enterprise's ESG risk value is to use a preset number of batches for analysis and processing in the same analysis and identification process when performing early warning processing of ESG risk value.

[0165] The percentage of associated data sources for high-risk modal data types: This refers to the percentage of all associated data sources owned by the enterprise that belong to the high-risk modal type. The higher this percentage, the greater the proportion of high-risk modalities in the enterprise's data.

[0166] Preset risk data source ratio threshold: Used to determine whether an enterprise is exposed to an excessively high proportion of high-risk modal data.

[0167] For companies that are the target of the analysis, their data has already been shown to have some volatility in previous steps, thus requiring a more rigorous early warning process. For non-target companies, if most of their data sources belong to high-risk modes, they may also face significant identification biases, thus requiring more rigorous multi-batch analysis to ensure the reliability of the early warnings.

[0168] This condition ensures that companies with high risk exposure (whether in terms of identity or data composition) are subject to more prudent monitoring to offset the impact of potential biases on the results of a single analysis.

[0169] Green Energy Technology (belonging to the target company of the analysis): Trigger 1, its early warning management method is to conduct analysis using a preset number of 5 batches.

[0170] Assuming another company, "Mingguang Energy" (not the target company in the analysis), has a total of 4 associated data sources, 3 of which belong to the high-risk modality (such as one text, one image, and one time series). Then the proportion of high-risk modality data sources is 3 / 4 = 75% > 50%. Similarly, triggering condition 1, 5 batches of analysis are used.

[0171] Furthermore, if the enterprise is not a target enterprise in the analysis and the proportion of associated data sources of the severe risk type in the associated data sources of the enterprise is not greater than the preset risk data source proportion threshold, then the deviation weight value of different modal data types is determined according to the analysis deviation risk type of different modal data types. If the sum of the deviation weight values ​​of different modal data types is greater than the preset deviation weight threshold, then the early warning management method for the ESG risk values ​​of all enterprises is determined to be that when performing early warning processing of ESG risk values, the same analysis and identification process uses a preset number of batches for analysis and processing, thereby ensuring the reliability of the analysis and processing.

[0172] Deviation weight value: A quantitative score assigned to each modal risk type to measure its risk severity.

[0173] Sum of Deviation Weights: For a specific firm, the weights of each modality it possesses are summed. This value reflects the overall risk level of the modalities involved in that firm.

[0174] Even if a non-target enterprise has a low proportion of high-risk modal data sources, the overall identification bias risk is determined. This risk is calculated by summing the bias weights of different modal data types. If the bias risk is significant, the entire system's analysis and processing must be more vigilant, employing more rigorous batch processing to ensure the reliability of early warnings for all enterprises (not just the target enterprise). This is a comprehensive risk control measure.

[0175] Specific examples: There are 3 data sources belonging to the severe risk mode, plus audio and video (no risk). The sum of the deviation weights = 2 + 2 + 2 = 6, which is not greater than 10. Therefore, sub-step 2.1 is not triggered, which means that the early warning management method for all enterprises (including Zhiguang Energy itself and all other target enterprises for analysis) is upgraded to use a preset number of 5 batches for analysis.

[0176] Furthermore, if the sum of the deviation weight values ​​of different modal data types is not greater than the preset deviation weight threshold, and if there are multiple modal data types with severe risk in the associated data types of the enterprise, then the early warning management method for the enterprise's ESG risk value is determined to be that when performing early warning processing of ESG risk value, the same analysis and identification process uses a preset number of batches for analysis and processing, thereby ensuring the reliability of the analysis and processing.

[0177] If the sum of the deviation weight values ​​is less than or equal to the preset deviation weight threshold, then it is further determined that if there are multiple modal data types of serious risk types (i.e., at least two) in the associated data types of the enterprise, then the early warning management method of the enterprise is to adopt a preset number of batches.

[0178] When an enterprise has two or more severe risk modalities, even if there are not many data sources under each modality, the combined risks of multiple modalities may lead to unstable analysis results. Therefore, more rigorous batch cross-validation is required.

[0179] This criterion identifies companies that are "high-risk in multimodal data" but "not particularly high in data volume," and grants them the same level of prudent treatment as companies with a high proportion of high-risk data sources.

[0180] Additionally, it should be noted that if the associated data types of the enterprise do not contain multiple modal data types with severe risk types, the analysis target enterprises whose data source similarity coefficients are greater than a preset similarity coefficient threshold are identified based on the data source similarity coefficients between them and different analysis target enterprises. These are then considered as similar enterprises. It is then determined whether the number of similar enterprises exceeds a preset similar enterprise number threshold. If so, then the overall identification risk is low, and the identification deviation of the ESG risk value analysis results of similar enterprises is also low. Therefore, under low risk conditions, the early warning management method for the enterprise's ESG risk values ​​is determined to be that when performing early warning processing of ESG risk values, the same analysis and identification process uses a basic number of batches for analysis and processing. If not, the early warning management method for the enterprise's ESG risk values ​​is determined to be that when performing early warning processing of ESG risk values, the same analysis and identification process uses a preset number of batches for analysis and processing, thereby ensuring the reliability of the analysis and processing.

[0181] Similar companies: Target companies whose data source similarity coefficient exceeds a preset threshold.

[0182] Preset threshold for the number of similar companies: This is used to determine whether there are enough target companies with similar data sources to the company to form a "group protection effect", implying that the company's data composition is mainstream and stable.

[0183] If a company's risk characteristics are not prominent (no multiple severity modalities), but it shares data sources with a large number of target companies in the analysis, it indicates that its data source is commonly used and validated in the industry, and the reliability of its analysis results is high. Therefore, a more economical base batch size can be used. Conversely, if there are few similar companies, its data source composition is relatively unique and lacks reference. To be on the safe side, a preset batch size should still be used.

[0184] This step utilizes the principle of "similar reference" to associate unknown companies with known companies through the similarity of data sources. By inferring the risk level of unknown companies based on the stability of known companies, the risk level of early warning strategies can be accurately downgraded.

[0185] Specific examples: "Xuguang Energy" has a total of 10 data sources, of which 4 are high-risk modalities (40% ≤ 50%). It only possesses one high-risk modality (text), while the rest are moderate or no-risk. Xuguang Energy only has text (high-risk) and audio / video (none). Now, we calculate its similarity coefficient with each target company in the analysis, and obtain: With Green Energy Technology: 0.20 (>0.15), with Blue Energy Technology: 0.10 (≤0.15), with New Energy Technology: 0.22 (>0.15), with Trina Solar: 0.18 (>0.15), with Ruiguang Energy: 0.12 (≤0.15); There are three similar companies (coefficient > 0.15): Green Energy Technology, Xinyuan Technology, and Trina Solar. The preset threshold for the number of similar companies is set to 2, therefore 3 > 2. Thus, Xuguang Energy's early warning management method uses a base number of 3 batches for analysis.

[0186] If there is only one similar enterprise (≤2), then the preset quantity of 5 batches should be used.

[0187] It should be noted that the preset quantity is greater than the basic quantity.

[0188] Example 2 Secondly, the present invention provides a computer system comprising: a memory and a processor connected in communication, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the above-described ESG risk warning method based on multimodal data when running the computer program.

[0189] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0190] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0191] The above description is merely one or more embodiments of this specification and is not intended to limit this specification. Various modifications and variations can be made to the one or more embodiments of this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of one or more embodiments of this specification should be included within the scope of the claims of this specification.

Claims

1. An ESG risk early warning method based on multimodal data, characterized in that, Specifically, it includes: Using the associated data sources of an enterprise's ESG risk values ​​and the modal data types of those data sources, the associated modal data types of the enterprise are determined. Based on the degree of overlap of the associated modal data types of different enterprises and the degree of correlation between the associated data sources of the enterprise and the associated modal data types of different enterprises, the target enterprises for ESG risk value analysis in the enterprise are determined. Based on the deviation of the target enterprise in different modal data types from other target enterprises in the associated data sources, and in combination with the associated modal data types of different enterprises, an ESG risk warning method is used to determine the deviation of the target enterprise's ESG risk values ​​between different analysis batches. In combination with the associated data sources of the enterprise in different modal data types, the analysis deviation risk type of the modal data type is determined. Based on the types of analytical bias risks in different modal data types and the overlap between the associated data sources and the target enterprise, the early warning management method for the enterprise's ESG risk value is determined.

2. The ESG risk early warning method based on multimodal data as described in claim 1, characterized in that, The associated data source is the data source used for the analysis and processing of the company's ESG risk values.

3. The ESG risk early warning method based on multimodal data as described in claim 1, characterized in that, The modal data type of the associated data source is determined based on the data type used for ESG risk value analysis and processing in the associated data source.

4. The ESG risk early warning method based on multimodal data as described in claim 1, characterized in that, The method for determining the associated modal data type of the enterprise is as follows: Based on the modal data type of the associated data source, determine the number of associated data sources in different modal data types; Based on the associated data sources of the company's ESG risk values, determine the number of associated data sources of the company's ESG risk values; The associated modal data types of the enterprise are determined by the number of associated data sources for the enterprise's ESG risk values ​​and the number of associated data sources in different modal data types.

5. The ESG risk early warning method based on multimodal data as described in claim 4, characterized in that, By utilizing the number of associated data sources for the company's ESG risk values ​​and the number of associated data sources across different modal data types, the associated modal data types of the company are determined, specifically including: The proportion of associated data sources for each modal data type is determined based on the proportion of the number of associated data sources in different modal data types to the total number of associated data sources for the enterprise's ESG risk value. Based on the number of associated data sources for the company's ESG risk values, the percentage threshold corresponding to the associated modal data types is determined. If the proportion of associated data sources of the modal data type is greater than the proportion threshold, then the modal data type is determined to belong to the associated modal data type of the enterprise.

6. The ESG risk early warning method based on multimodal data as described in claim 5, characterized in that, The more associated data sources an enterprise has, the smaller the percentage threshold corresponding to the associated modal data type.

7. The ESG risk early warning method based on multimodal data as described in claim 1, characterized in that, The method for determining the target companies for ESG risk value analysis in the aforementioned enterprises is as follows: Based on the degree of overlap of the associated modal data types of different enterprises, the proportion of enterprises belonging to the associated modal data type under different modal data types is determined, and this proportion is used as the matching proportion under the modal data type. Based on the associated data sources of the enterprise in different modal data types, determine the number of associated data sources in different modal data types; By using the proportion of matching data under different modal data types, and the number of associated data sources for the enterprise in the modal data types with different matching proportions, it can be determined whether the enterprise is a target enterprise for ESG risk value analysis.

8. The ESG risk early warning method based on multimodal data as described in claim 7, characterized in that, The associated data type of the enterprise is a modal data type that has associated data sources.

9. The ESG risk early warning method based on multimodal data as described in claim 1, characterized in that, The method for determining the early warning management method for the ESG risk value of the enterprise is as follows: Based on the analysis of the risk types of deviations in different modal data types, determine the modal data types of severe risk types; Based on the overlap between the associated data source of the enterprise and the target enterprise of the analysis, the number of overlaps between the associated data source and different target enterprises of the analysis is determined, and the data source similarity coefficient between the associated data source and different target enterprises of the analysis is determined based on the number of overlaps. An early warning management method for determining the ESG risk value of an enterprise is established by utilizing the modal data type of the severe risk type and the data source similarity coefficient between the enterprise and different target enterprises for analysis.

10. A computer system, comprising: A memory and processor connected by communication, and a computer program stored in the memory and capable of running on the processor, characterized in that, when the processor runs the computer program, it executes an ESG risk warning method based on multimodal data as described in any one of claims 1-9.

Images