ESG index evaluation method, device and medium
By calculating the similarity and balanced sampling between the enterprise's text and ESG indicators, and optimizing the encoder, the problem of data imbalance in the ESG evaluation method is solved, and the accuracy and rationality of ESG indicator evaluation is improved.
Patent Information
- Application Number
- CN202411390835.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-08
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-10-08
AI Technical Summary
The existing ESG indicator evaluation method has the problem of low rationality, which is mainly reflected in the diverse information sources of enterprises and the uneven data between and within indicators, resulting in inaccurate evaluation results.
By obtaining multiple first-level evaluation indicators and secondary evaluation indicators of the ESG evaluation system, we calculate the similarity between the text and indicators of the target enterprise, divide the label-free text and label text, perform balanced sampling and encoding, and build comprehensive loss function optimization encoder, and then calculate the enterprise's ESG indicator evaluation results.
The accuracy and rationality of the ESG index evaluation results are improved, and the balance and practicality of the evaluation results are enhanced by optimizing the performance of the encoder.
Smart Images

Figure CN118898424B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of enterprise evaluation, and in particular to an ESG indicator evaluation method, device and medium. Background Art
[0002] The Environmental, Social and Governance (ESG) rating system, also known as the ESG rating, evaluates companies using three indicators: environment (E), society (S), and governance (G). It was created by business and non-profit organizations to assess the consistency of companies with sustainable development goals in terms of commitment, performance, business model, and structure. Initially, ESG ratings were used by investment companies to screen and evaluate companies in their various funds and portfolios. Today, job seekers, customers, and others also refer to these ratings when evaluating business relationships, and rated companies are able to better understand their strengths, weaknesses, risks, and opportunities.
[0003] However, after in-depth research, it is found that the differences in ESG rating methods between different companies are still large, mainly reflected in the diversity of corporate information sources, imbalanced data between indicators and within indicators, etc. In addition, the existing methods for building corporate ESG indicator systems have not yet proposed specific and feasible implementation plans for the problem of data imbalance. Therefore, the existing ESG indicator evaluation methods have the problem of low rationality of ESG indicator evaluation. Summary of the invention
[0004] The present application provides an ESG indicator evaluation method, device and medium, which can solve the problem of low rationality of ESG indicator evaluation.
[0005] In a first aspect, an embodiment of the present application provides an ESG indicator evaluation method, the ESG indicator evaluation method comprising:
[0006] Obtain multiple first-level evaluation indicators of the ESG evaluation system and multiple second-level evaluation indicators under each first-level evaluation indicator, and obtain multiple texts of each target enterprise; the texts are used to describe the corporate information of the target enterprise;
[0007] For each target enterprise, the similarity between each text of the target enterprise and each secondary evaluation indicator is calculated respectively, and multiple key texts are determined from all texts according to all similarities, and all key texts are divided into multiple unlabeled texts and multiple to-be-determined texts, and for each to-be-determined text, the true label corresponding to the to-be-determined text is obtained to obtain the labeled text, and all unlabeled texts and all labeled texts are sampled evenly to obtain an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts; at least one similarity among all similarities corresponding to the key texts is greater than the similarity threshold, and the true label is one of the multiple primary evaluation indicators;
[0008] Encode each unlabeled subset using an encoder to obtain an unlabeled data representation, and classify each unlabeled data representation using a classifier to obtain a predicted label for each unlabeled text in each unlabeled subset;
[0009] According to all predicted labels, the true labels of each label text in all label subsets, and all unlabeled data representations, a comprehensive loss function is constructed, and the encoder is optimized based on the comprehensive loss function to obtain an optimized encoder; the comprehensive loss function is used to describe the accuracy of all predicted labels and the accuracy of all unlabeled data representations;
[0010] Obtain multiple text sets of the enterprise to be evaluated, and use the optimized encoder to encode each text set to obtain multiple coded representations of the enterprise to be evaluated; the labels of all texts in the text set are the same first-level evaluation indicator;
[0011] Calculate the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators based on all coding representations.
[0012] Optionally, calculate the similarity between each text and each secondary evaluation indicator, including:
[0013] By formula:
[0014]
[0015] Calculate the The text and The similarity between the secondary evaluation indicators ;
[0016] in, represents the cosine similarity calculation, , The total number of texts representing the target enterprise, , Represents the total number of secondary evaluation indicators.
[0017] Optionally, all unlabeled texts and all labeled texts are sampled evenly to obtain an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts, including:
[0018] Calculate the class probability of each first-level evaluation indicator;
[0019] Determine the unlabeled subset from all unlabeled texts based on all category probabilities;
[0020] Determine a label subset from all label texts based on all category probabilities.
[0021] Optionally, calculate the class probabilities for each first-level evaluation metric, including:
[0022] By formula:
[0023]
[0024] Calculate the Class probability of the first-level evaluation index ;
[0025] in, Indicates The number of label texts corresponding to the first-level evaluation indicators, Indicates The number of label texts corresponding to the first-level evaluation indicators, Indicates the parameters, , , , Represents the number of first-level evaluation indicators.
[0026] Optionally, the comprehensive loss function is:
[0027]
[0028] in, represents the comprehensive loss function value, , , All represent hyperparameters, represents the inter-class loss function, represents the intra-class loss function, Represents the ESG loss function:
[0029]
[0030]
[0031]
[0032] in, represents the number of first-level evaluation indicators, Represents all the corresponding predicted labels and all the true labels of the label subset The probability distribution of the first-level evaluation indicators, represents the overall probability distribution of all predicted labels and all true labels for a subset of labels, express and The distance between the bulldozers, Indicates The entropy of the first-level evaluation index, Indicates the number of data sources. Indicates that in all unlabeled subsets and labeled subsets, the corresponding The dataset of the first-level evaluation indicators comes from The percentage of all key texts from the data sources in the dataset. and All represent hyperparameters, represents the number of scoring features, Indicates The weight of the scoring feature, Indicates The root mean square error of the scoring features, represents the number of target enterprises, Indicates The first The credibility of a feature, Indicates the use of The unlabeled data corresponding to each enterprise is represented by the The first The predicted feature values, Indicates The first The true eigenvalues.
[0033] Optionally, the encoder is optimized based on the comprehensive loss function to obtain an optimized encoder, including:
[0034] Determine whether the comprehensive loss function value is less than the preset loss function value;
[0035] If so, the encoder is taken as the optimized encoder;
[0036] Otherwise, adjust the parameters of the encoder, and return to the steps of calculating the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise respectively, and determining multiple key texts from all texts based on all similarities, dividing all key texts into multiple unlabeled texts and multiple to-be-determined texts, and obtaining the true label corresponding to the to-be-determined text for each to-be-determined text respectively, obtaining the labeled text, performing balanced sampling on all unlabeled texts and all labeled texts, and obtaining the steps of an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts.
[0037] Optionally, the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators are calculated based on all coding representations, including:
[0038] Use LSTM to predict each encoding representation separately to obtain the sentiment classification corresponding to each encoding representation; the sentiment classification is one of positive sentiment, neutral sentiment, and negative sentiment;
[0039] Calculate the ESG indicator evaluation results of the company to be evaluated under all primary evaluation indicators based on all sentiment classifications.
[0040] Optionally, the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators are calculated based on all sentiment categories, including:
[0041] By formula:
[0042]
[0043] Calculate the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators ;
[0044] in, Represents the first evaluation indicator The weight of Represents the first evaluation indicator The weight of Represents the first evaluation indicator The weight of Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, It represents the sum of the number of coding representations with negative emotion in all coding representations of the enterprise to be evaluated and the number of coding representations with positive emotion in all coding representations.
[0045] In a second aspect, an embodiment of the present application provides an ESG indicator evaluation device, including:
[0046] An acquisition module is used to acquire multiple first-level evaluation indicators of the ESG indicator evaluation system and multiple second-level evaluation indicators under each first-level evaluation indicator, and to acquire multiple texts of each target enterprise; the texts are used to describe the corporate information of the target enterprise;
[0047] A balanced sampling module calculates the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise, and determines multiple key texts from all texts based on all similarities, divides all key texts into multiple unlabeled texts and multiple to-be-determined texts, and obtains the true label corresponding to each to-be-determined text for each to-be-determined text to obtain the labeled text, performs balanced sampling on all unlabeled texts and all labeled texts, and obtains an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts; at least one similarity among all similarities corresponding to the key texts is greater than the similarity threshold, and the true label is one of the multiple primary evaluation indicators;
[0048] The classification module encodes each unlabeled subset using an encoder to obtain an unlabeled data representation, and classifies each unlabeled data representation using a classifier to obtain a predicted label for each unlabeled text in each unlabeled subset;
[0049] A construction module is used to construct a comprehensive loss function based on all predicted labels, the true label of each label text in all label subsets, and all unlabeled data representations, and optimize the encoder based on the comprehensive loss function to obtain an optimized encoder; the comprehensive loss function is used to describe the accuracy of all predicted labels and the accuracy of all unlabeled data representations;
[0050] The encoding module obtains multiple text sets of the enterprise to be evaluated, and uses the optimized encoder to encode each text set to obtain multiple coded representations of the enterprise to be evaluated; the labels of all texts in the text set are the same first-level evaluation indicator;
[0051] The calculation module calculates the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators based on all coding representations.
[0052] In a third aspect, an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned ESG indicator evaluation method when executing the above-mentioned computer program.
[0053] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the above-mentioned ESG indicator evaluation method is implemented.
[0054] The above solution of the present application has the following beneficial effects:
[0055] In an embodiment of the present application, by calculating the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise respectively, and determining multiple key texts from all texts based on all similarities, all key texts are divided into multiple unlabeled texts and multiple texts to be determined, and for each text to be determined respectively, the true label corresponding to the text to be determined is obtained to obtain the labeled text, and all unlabeled texts and all labeled texts are sampled evenly, which can improve the balance and reality of the data in the obtained unlabeled subsets and labeled subsets; based on all predicted labels with balance and reality, the true label of each labeled text in all labeled subsets, and all unlabeled data representations, a constructed The comprehensive loss function is highly reasonable, and can be used to describe the accuracy of the predicted label and unlabeled data representation, and then to describe the performance of the encoder; the encoder is optimized based on a reasonable comprehensive loss function to obtain an optimized encoder, so that the performance of the encoder is improved, and the accuracy of the encoder's output results is improved, and then multiple text sets of the company to be evaluated are obtained, and each text set is encoded using the optimized encoder to obtain multiple encoded representations of the company to be evaluated, and finally the ESG indicator evaluation results of the company to be evaluated are calculated based on all the encoded representations, and the accuracy of the encoded representation obtained using the encoder with improved performance is improved, thereby effectively improving the rationality of the ESG indicator evaluation results.
[0056] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0058] Figure 1 A flowchart of an ESG indicator evaluation method provided in one embodiment of the present application;
[0059] Figure 2 A schematic diagram of the workflow of an ESG indicator evaluation system provided in one embodiment of the present application;
[0060] Figure 3 A schematic diagram of the structure of an ESG indicator evaluation device provided in one embodiment of the present application;
[0061] Figure 4 A schematic diagram of the structure of a terminal device provided in one embodiment of the present application. DETAILED DESCRIPTION
[0062] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.
[0063] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.
[0064] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0065] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.
[0066] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.
[0067] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.
[0068] In response to the problem of low rationality of existing ESG indicator evaluation, an embodiment of the present application provides an ESG indicator evaluation method, which calculates the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise respectively, and determines multiple key texts from all texts based on all similarities, divides all key texts into multiple unlabeled texts and multiple to-be-determined texts, and obtains the true label corresponding to each to-be-determined text for each to-be-determined text respectively, obtains the labeled text, and performs balanced sampling on all unlabeled texts and all labeled texts, thereby improving the balance and reality of the data in the obtained unlabeled subsets and labeled subsets; based on all predicted labels with balance and reality, each label in all labeled subsets The real label of the text and the representation of all unlabeled data are used to construct a comprehensive loss function with high rationality. The accuracy of the predicted label and the representation of unlabeled data can be described through the comprehensive loss function, and then the performance of the encoder can be described; the encoder is optimized based on a reasonable comprehensive loss function to obtain an optimized encoder, so that the performance of the encoder is improved, and the accuracy of the output result of the encoder is improved, and then multiple text sets of the company to be evaluated are obtained, and each text set is encoded using the optimized encoder to obtain multiple encoded representations of the company to be evaluated, and finally the ESG indicator evaluation results of the company to be evaluated are calculated according to all the encoded representations, and the accuracy of the encoded representation obtained by the encoder with improved performance is improved, thereby effectively improving the rationality of the ESG indicator evaluation results.
[0069] Next, an illustrative explanation of the ESG indicator evaluation method provided in this application is given.
[0070] like Figure 1 As shown, the ESG indicator evaluation method provided in this application includes the following steps:
[0071] Step 11, obtain multiple first-level evaluation indicators of the ESG evaluation system and multiple second-level evaluation indicators under each first-level evaluation indicator, and obtain multiple texts of each target enterprise.
[0072] The above text is used to describe the corporate information of the target company, such as "the company's carbon emissions seriously exceeded the standard" and "the company's revenue and profit for the whole of last year."
[0073] In some embodiments of the present application, the above-mentioned multiple first-level evaluation indicators are the three first-level evaluation indicators of environment (E), society (S), and corporate governance (G) in the ESG evaluation system, and the second-level evaluation indicators are the factors under the first-level evaluation indicators. For example, the first-level evaluation indicator E corresponds to the second-level evaluation indicators such as carbon emissions, pollution and waste, and climate change; the first-level evaluation indicator S corresponds to the second-level evaluation indicators such as social reputation and media relations; the first-level evaluation indicator G corresponds to the second-level evaluation indicators such as internal control, management mechanism, and finance.
[0074] It should be noted that multiple texts of the target enterprise can be obtained by performing text extraction on relevant news, financial reports, prospectuses and other documents of the target enterprise. The above-mentioned relevant news, financial reports and prospectuses can be encoded into multiple text sequences using software such as Tokenizer, and then the text sequences can be padded using software such as pad_sequences to obtain multiple texts of the same length.
[0075] Step 12, for each target enterprise, calculate the similarity between each text of the target enterprise and each secondary evaluation indicator, and determine multiple key texts from all texts based on all similarities, divide all key texts into multiple unlabeled texts and multiple to-be-determined texts, and for each to-be-determined text, obtain the true label corresponding to the to-be-determined text to obtain the labeled text, perform balanced sampling on all unlabeled texts and all labeled texts, and obtain an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts.
[0076] At least one of all similarities corresponding to the key text is greater than the similarity threshold, and the true label is one of the multiple first-level evaluation indicators. The labeled text is the key text that has obtained the true label, and the unlabeled text is the key text that has not obtained the true label.
[0077] In some embodiments of the present application, the above-mentioned steps of calculating the similarity between each text of the target enterprise and each secondary evaluation indicator, and determining multiple key texts from all texts according to all similarities, dividing all key texts into multiple unlabeled texts and multiple to-be-determined texts, and respectively obtaining the true label corresponding to each to-be-determined text for each to-be-determined text, obtaining the labeled text, performing balanced sampling on all unlabeled texts and all labeled texts, and obtaining the steps of an unlabeled subset including multiple unlabeled texts and a labeled subset including multiple labeled texts include:
[0078] In the first step, the similarity between each text and each secondary evaluation indicator is calculated.
[0079] By formula:
[0080]
[0081] Calculate the The text and The similarity between the secondary evaluation indicators .
[0082] in, represents the cosine similarity calculation, , The total number of texts representing the target enterprise, , Represents the total number of secondary evaluation indicators.
[0083] In the second step, multiple key texts are determined from all texts based on all similarities.
[0084] Specifically, for each text, if there is a similarity greater than a similarity threshold among all similarities corresponding to the text, the text is regarded as a key text; if all similarities corresponding to the text are less than the similarity threshold, the text is not marked.
[0085] The third step is to divide all key texts into multiple unlabeled texts and multiple to-be-determined texts, and obtain the true label corresponding to each to-be-determined text to obtain the labeled text.
[0086] It should be noted that when dividing the key text, it is necessary to ensure that the number of unlabeled text and the number of text to be determined are both greater than 0.
[0087] Exemplarily, each text to be determined can be marked manually to obtain the real label corresponding to the text to be determined, such as the text "the company's carbon emissions seriously exceed the standard", the real label corresponding to the text is the first-level evaluation indicator E.
[0088] The fourth step is to perform balanced sampling on all unlabeled texts and all labeled texts to obtain an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts.
[0089] The labels of the multiple unlabeled texts included in the unlabeled subset may be different first-level evaluation indicators, and the labels of the multiple labeled texts included in the labeled subset may be different first-level evaluation indicators. For example, the labeled subset includes 3 labeled texts, the labels of the first labeled text and the second labeled text are first-level evaluation indicator E, and the label of the second labeled text is first-level evaluation indicator S. In addition, the unlabeled subset only includes multiple unlabeled texts selected from all unlabeled texts, and the labeled subset only includes multiple labeled texts selected from all labeled texts.
[0090] First, the class probabilities for each first-level evaluation metric are calculated.
[0091] Specifically, through the formula:
[0092]
[0093] Calculate the Class probability of the first-level evaluation index .
[0094] in, Indicates The number of label texts corresponding to the first-level evaluation indicators, Indicates The number of label texts corresponding to the first-level evaluation indicators, Indicates the parameters, , , , Represents the number of first-level evaluation indicators.
[0095] It should be noted that the above category probabilities are used to describe the probability that each unlabeled text or labeled text is labeled as a first-level evaluation indicator, such as The probability of the category of the first-level evaluation indicator is 20%, then the label of each unlabeled text or labeled text is The probability of a first-level evaluation indicator is 20%.
[0096] Then, the unlabeled subset is determined from all the unlabeled texts according to all the category probabilities.
[0097] Specifically, according to each category probability, multiple unlabeled texts corresponding to each first evaluation indicator are selected from all unlabeled texts in turn, and all the determined unlabeled texts are integrated into a set to obtain an unlabeled subset.
[0098] Exemplarily, the category probabilities of the three first-level evaluation indicators are 20%, 40%, and 25%, respectively. For each unlabeled text, when determining the multiple unlabeled texts corresponding to the first first-level evaluation indicator in the unlabeled text, each unlabeled text has a 20% probability of being selected. If the unlabeled text is not selected, then when determining the multiple unlabeled texts corresponding to the second first-level evaluation indicator in the unlabeled text, the unlabeled text has a 40% probability of being selected.
[0099] Finally, a label subset is determined from all label texts based on all category probabilities.
[0100] It should be noted that the steps for determining the labeled subset are the same as the steps for determining the unlabeled subset, that is, according to each category probability, multiple label texts corresponding to each first evaluation indicator are selected from all label texts in turn, and all the determined label texts are integrated into a set to obtain a label subset. Since the category probability represents the probability that each unlabeled text or the label of the label text is a first-level evaluation indicator, when determining the unlabeled subset, the probability of each unlabeled text being selected is the same, and when determining the label subset, the probability of each label text being selected is the same, thereby achieving balanced sampling.
[0101] It is worth mentioning that by evenly sampling all key texts to obtain labeled subsets and unlabeled subsets, the data in the labeled subsets and the unlabeled subsets are balanced, thereby improving the rationality of the labeled subsets and the unlabeled subsets.
[0102] Step 13: Encode each unlabeled subset using an encoder to obtain an unlabeled data representation, and classify each unlabeled data representation using a classifier to obtain a predicted label for each unlabeled text in each unlabeled subset.
[0103] In some embodiments of the present application, the encoder may be a recurrent neural network, a multi-layer perceptron or other model, the classifier may be a fully connected neural network or other model, and the predicted label may be one of multiple primary evaluation indicators.
[0104] Step 14, construct a comprehensive loss function based on all predicted labels, the true label of each label text in all label subsets, and all unlabeled data representations, and optimize the encoder based on the comprehensive loss function to obtain an optimized encoder.
[0105] The above comprehensive loss function is used to describe the accuracy of all predicted labels and the accuracy of all unlabeled data representations.
[0106] The above comprehensive loss function is:
[0107]
[0108] in, represents the comprehensive loss function value, , , All represent hyperparameters, represents the inter-class loss function, represents the intra-class loss function, Represents the ESG loss function:
[0109]
[0110]
[0111]
[0112] in, represents the number of first-level evaluation indicators, Represents all the corresponding predicted labels and all the true labels of the label subset The probability distribution of the first-level evaluation indicators, represents the overall probability distribution of all predicted labels and all true labels for a subset of labels, express and The distance between the bulldozers, Indicates The entropy of the first-level evaluation index, Indicates the number of data sources. Indicates that in all unlabeled subsets and labeled subsets, the corresponding The dataset of the first-level evaluation indicators comes from The percentage of all key texts in the data source in the dataset. and All represent hyperparameters, represents the number of scoring features, Indicates The weight of the scoring feature, Indicates The root mean square error of the scoring features, represents the number of target enterprises, Indicates The first The credibility of a feature, Indicates the use of The unlabeled data corresponding to each enterprise is represented by the The first The predicted feature values, Indicates The first The true eigenvalue.
[0113] Illustratively, the above-mentioned scoring features can be features such as the company's environmental policy, employee benefit plan, corporate governance structure, etc. that have a significant impact on the company's ESG indicator evaluation results. The above-mentioned data source is the source of the text, i.e. the relevant news, financial statements, prospectus, etc. mentioned above. The predicted feature values obtained by using the unlabeled data representation in the above text can be processed by models such as convolutional neural networks to obtain predicted feature values.
[0114] In some embodiments of the present application, the above-mentioned step of optimizing the encoder based on the comprehensive loss function to obtain the optimized encoder is:
[0115] Determine whether the comprehensive loss function value is less than the preset value of the loss function.
[0116] If so, the encoder is taken as the optimized encoder.
[0117] Otherwise, adjust the parameters of the encoder, and return to the steps of calculating the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise respectively, and determining multiple key texts from all texts based on all similarities, dividing all key texts into multiple unlabeled texts and multiple to-be-determined texts, and obtaining the true label corresponding to the to-be-determined text for each to-be-determined text respectively, obtaining the labeled text, performing balanced sampling on all unlabeled texts and all labeled texts, and obtaining the steps of an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts.
[0118] After each adjustment of the encoder parameters, the unlabeled subset and the labeled subset are obtained again, so that the encoder processes a different unlabeled subset each time and obtains a different unlabeled data representation, thereby improving the accuracy of the encoder for different inputs and effectively improving the robustness and versatility of the encoder.
[0119] It should be noted that the encoder can be updated using the exponential moving average (EMA). When the encoder is updated, the classifier can be updated at the same time, using the formula:
[0120]
[0121] Calculate the updated parameters of the classifier .
[0122] in, Indicates that the classifier is The parameters of the first-level evaluation indicators are: represents the two-norm, represents a hyperparameter, when When , the formula degenerates into the standard L2 regularization. No scaling is done.
[0123] It is worth mentioning that constructing a comprehensive loss function can describe the accuracy of the predicted label and unlabeled data representation through the comprehensive loss function, and then describe the performance of the encoder. Optimizing the encoder using the comprehensive loss function can improve the performance of the encoder.
[0124] Step 15, obtaining multiple text sets of the enterprise to be evaluated, and encoding each text set using the optimized encoder to obtain multiple coded representations of the enterprise to be evaluated.
[0125] All texts in the above text set are labeled with the same first-level assessment indicator, and the above-mentioned companies to be assessed are companies that need to undergo ESG indicator assessment.
[0126] Specifically, multiple texts of the company to be evaluated are obtained from relevant news, financial statements, prospectuses and other documents of the company to be evaluated, and multiple key texts are determined from all the texts. Each key text can be marked manually or by other means to obtain a label for each key text (one of multiple first-level evaluation indicators), and all key texts with the same label are divided into a set, and each set is evenly divided into multiple subsets, each subset is regarded as a text set, and then each text set is encoded using an optimized encoder to obtain multiple encoded representations.
[0127] For example, the optimized encoder can be run using computer software such as matlab and python to encode the text collection and obtain the encoded representation.
[0128] It is worth mentioning that using the optimized encoder to encode the text collection can improve the accuracy of the encoding, thereby improving the accuracy of the encoding representation in expressing the information of the text collection.
[0129] Step 16: Calculate the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators based on all coding representations.
[0130] In some embodiments of the present application, the step of calculating the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators according to all coding representations includes:
[0131] In the first step, the long short-term memory network (LSTM) is used to predict each encoding representation separately to obtain the sentiment classification corresponding to each encoding representation.
[0132] The above-mentioned sentiment classification is one of positive sentiment, neutral sentiment and negative sentiment, which is used to describe whether the information described by the corresponding coding representation is positive information, negative information or neutral information for the enterprise to be evaluated.
[0133] Specifically, through the formula:
[0134]
[0135]
[0136]
[0137]
[0138]
[0139]
[0140] Calculate the sentiment classification corresponding to the encoded representation .
[0141] in, , , , , , , , , and All represent parameters. Indicates the hidden state, represents the output gate, Indicates the unit status, represents the hidden representation of the previous time step, represents the encoded representation of the input, represents the forget gate, represents the cell state at the previous time step, Represents the input gate.
[0142] The second step is to calculate the ESG indicator evaluation results of the company to be evaluated under all primary evaluation indicators based on all sentiment classifications.
[0143] Specifically, through the formula:
[0144]
[0145] Calculate the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators .
[0146] in, Represents the first evaluation indicator The weight of Represents the first evaluation indicator The weight of Represents the first evaluation indicator The weight of Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, It represents the sum of the number of coding representations with negative emotion in all coding representations of the enterprise to be evaluated and the number of coding representations with positive emotion in all coding representations.
[0147] For example, the ESG indicator evaluation result of the enterprise to be evaluated can be a numerical value. The larger the value, the better the ESG indicator evaluation result of the enterprise to be evaluated. It can also be a rating obtained based on the value and the preset level range, such as the preset level A is The value is greater than 80, and the grade is B. The value is less than or equal to 80 and greater than 50, and the grade is C. The value is less than or equal to 50.
[0148] It is worth mentioning that the ESG indicator evaluation results obtained based on the coding representation with improved accuracy are highly accurate and reasonable.
[0149] The ESG indicator evaluation system is illustrated below with reference to a specific example.
[0150] like Figure 2As shown, the ESG index evaluation system includes a preprocessing module, a balancing module and a classification rating module. In the preprocessing module, similarity calculation is performed on the enterprise information (i.e., the text mentioned above) and the index (i.e., the secondary evaluation index mentioned above), and multiple key texts are determined to obtain an unbalanced data set. In the balancing module, the unbalanced data set is subjected to balanced sampling, encoder, and classifier processing to obtain a balanced encoding representation. In the classification rating module, the encoding representation is processed using LSTM to finally obtain the ESG index evaluation result.
[0151] The following is an exemplary description of the ESG indicator evaluation device provided in this application.
[0152] like Figure 3 As shown, the embodiment of the present application provides an ESG indicator evaluation device, and the ESG indicator evaluation device 300 includes:
[0153] The acquisition module 301 acquires multiple first-level evaluation indicators of the ESG evaluation system and multiple second-level evaluation indicators under each first-level evaluation indicator, and acquires multiple texts of each target enterprise; the texts are used to describe the enterprise information of the target enterprise;
[0154] The balanced sampling module 302 calculates the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise, and determines multiple key texts from all texts based on all similarities, divides all key texts into multiple unlabeled texts and multiple to-be-determined texts, and obtains the true label corresponding to each to-be-determined text for each to-be-determined text to obtain the labeled text, performs balanced sampling on all unlabeled texts and all labeled texts, and obtains an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts; at least one similarity among all similarities corresponding to the key texts is greater than the similarity threshold, and the true label is one of the multiple primary evaluation indicators;
[0155] The classification module 303 encodes each unlabeled subset using an encoder to obtain an unlabeled data representation, and classifies each unlabeled data representation using a classifier to obtain a predicted label for each unlabeled text in each unlabeled subset;
[0156] A construction module 304 constructs a comprehensive loss function according to all predicted labels, the true label of each label text in all label subsets, and all unlabeled data representations, and optimizes the encoder based on the comprehensive loss function to obtain an optimized encoder; the comprehensive loss function is used to describe the accuracy of all predicted labels and the accuracy of all unlabeled data representations;
[0157] The encoding module 305 obtains multiple text sets of the enterprise to be evaluated, and uses the optimized encoder to encode each text set to obtain multiple encoding representations of the enterprise to be evaluated; the labels of all texts in the text set are the same first-level evaluation indicator;
[0158] The calculation module 306 calculates the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators based on all coding representations.
[0159] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.
[0160] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.
[0161] like Figure 4 As shown, an embodiment of the present application provides a terminal device. The terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 4 Only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 implements the steps of any of the above-mentioned method embodiments when executing the computer program D102.
[0162] Specifically, when the processor D100 executes the computer program D102, for each target enterprise, the similarity between each text of the target enterprise and each secondary evaluation indicator is calculated, and based on all similarities, multiple key texts are determined from all texts, all key texts are divided into multiple unlabeled texts and multiple to-be-determined texts, and for each to-be-determined text, the true label corresponding to the to-be-determined text is obtained to obtain the labeled text, and all unlabeled texts and all labeled texts are sampled in a balanced manner, so as to improve the balance and reality of the data in the obtained unlabeled subsets and labeled subsets; based on all predicted labels with balance and reality, the true label of each labeled text in all labeled subsets, and all unlabeled texts are sampled in a balanced manner. The comprehensive loss function constructed by the company is highly reasonable and can describe the accuracy of the predicted label and unlabeled data representation through the comprehensive loss function, thereby describing the performance of the encoder; the encoder is optimized based on a reasonable comprehensive loss function to obtain an optimized encoder, so that the performance of the encoder is improved, and the accuracy of the output result of the encoder is improved, and then multiple text sets of the company to be evaluated are obtained, and each text set is encoded by using the optimized encoder to obtain multiple encoded representations of the company to be evaluated, and finally the ESG indicator evaluation results of the company to be evaluated are calculated according to all the encoded representations, and the accuracy of the encoded representation obtained by the encoder with improved performance is improved, thereby effectively improving the rationality of the ESG indicator evaluation results.
[0163] The processor D100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.
[0164] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart memory card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card (Flash Card), etc. equipped on the terminal device D10. Further, the memory D101 may also include both an internal storage unit of the terminal device D10 and an external storage device. The memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data, and other programs, such as the program code of the computer program, etc. The memory D101 may also be used to temporarily store data that has been output or is to be output.
[0165] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.
[0166] An embodiment of the present application provides a computer program product. When the computer program product runs on a terminal device, the terminal device can implement the steps in the above-mentioned method embodiments when executing the computer program product.
[0167] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, which can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device that can carry the computer program code to the ESG index evaluation method device / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.
[0168] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.
[0169] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.
[0170] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.
Claims
1. An ESG indicator evaluation method, characterized in that: include: Obtain multiple first-level evaluation indicators of the ESG evaluation system and multiple second-level evaluation indicators under each of the first-level evaluation indicators, and obtain multiple texts of each target enterprise; The text is used to describe the corporate information of the target enterprise; For each of the target enterprises, the similarity between each text of the target enterprise and each secondary evaluation indicator is calculated respectively, and based on all the similarities, multiple key texts are determined from all the texts, and all the key texts are divided into multiple unlabeled texts and multiple to-be-determined texts, and for each of the to-be-determined texts, the true label corresponding to the to-be-determined text is obtained to obtain the labeled text, and all the unlabeled texts and all the labeled texts are evenly sampled to obtain an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts; at least one similarity among all the similarities corresponding to the key texts is greater than a similarity threshold, and the true label is one of the multiple primary evaluation indicators; Encode each of the unlabeled subsets using an encoder to obtain an unlabeled data representation, and classify each of the unlabeled data representations using a classifier to obtain a predicted label for each unlabeled text in each of the unlabeled subsets; According to all predicted labels, the true label of each label text in all label subsets, and all unlabeled data representations, a comprehensive loss function is constructed, and the encoder is optimized based on the comprehensive loss function to obtain an optimized encoder; the comprehensive loss function is used to describe the accuracy of all predicted labels and the accuracy of all unlabeled data representations; Acquire multiple text sets of the enterprise to be evaluated, and use the optimized encoder to encode each of the text sets to obtain multiple coded representations of the enterprise to be evaluated; the labels of all texts in the text sets are the same first-level evaluation indicator; Calculate the ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators based on all coding representations.
2. The ESG indicator evaluation method according to claim 1, characterized in that: The calculation of the similarity between each text and each secondary evaluation index includes: By formula: Calculate the The text and The similarity between the secondary evaluation indicators ; in, represents the cosine similarity calculation, , The total number of texts representing the target enterprise, , Represents the total number of secondary evaluation indicators.
3. The ESG indicator evaluation method according to claim 1, characterized in that: The balanced sampling of all unlabeled texts and all labeled texts to obtain an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts includes: Calculate the class probability of each first-level evaluation indicator; Determine the unlabeled subset from all unlabeled texts based on all category probabilities; Determine a label subset from all label texts based on all category probabilities.
4. The ESG indicator evaluation method according to claim 3, characterized in that: The calculation of the category probability of each first-level evaluation indicator includes: By formula: Calculate the Class probability of the first-level evaluation index ; in, Indicates the The number of label texts corresponding to the first-level evaluation indicators, Indicates The number of label texts corresponding to the first-level evaluation indicators, Indicates the parameters, , , , Represents the number of first-level evaluation indicators.
5. The ESG indicator evaluation method according to claim 1, characterized in that: The comprehensive loss function is: in, represents the comprehensive loss function value, , , All represent hyperparameters, represents the inter-class loss function, represents the intra-class loss function, Represents the ESG loss function: in, represents the number of first-level evaluation indicators, Represents all the corresponding predicted labels and all the true labels of the label subset The probability distribution of the first-level evaluation indicators, represents the overall probability distribution of all predicted labels and all true labels for a subset of labels, express and The distance between the bulldozers, Indicates The entropy of the first-level evaluation index, Indicates the number of data sources. Indicates that in all unlabeled subsets and labeled subsets, the corresponding The data set of the first-level evaluation indicators comes from The percentage of all key texts from the data sources in the dataset. and All represent hyperparameters, represents the number of scoring features, Indicates The weight of the scoring feature, Indicates The root mean square error of the scoring features, represents the number of target enterprises, Indicates The first The credibility of a feature, Indicates that the use of the The first The first Predicted feature values, Indicates the The first The true eigenvalue.
6. The ESG indicator evaluation method according to claim 1, characterized in that: The step of optimizing the encoder based on the comprehensive loss function to obtain an optimized encoder includes: Determine whether the comprehensive loss function value is less than a preset loss function value; If yes, the encoder is used as the optimized encoder; Otherwise, adjust the parameters of the encoder, and return to the step of calculating the similarity between each text of the target enterprise and each secondary evaluation indicator for each target enterprise respectively, and determining multiple key texts from all texts based on all similarities, dividing all key texts into multiple unlabeled texts and multiple texts to be determined, and obtaining the true label corresponding to each text to be determined respectively, obtaining the labeled text, performing balanced sampling on all unlabeled texts and all labeled texts, and obtaining the steps of an unlabeled subset containing multiple unlabeled texts and a labeled subset containing multiple labeled texts.
7. The ESG indicator evaluation method according to claim 1, characterized in that: The ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators are calculated based on all coding representations, including: Using LSTM to predict each of the encoded representations, respectively, to obtain a sentiment classification corresponding to each of the encoded representations; the sentiment classification is one of positive sentiment, neutral sentiment, and negative sentiment; The ESG indicator evaluation results of the company to be evaluated under all primary evaluation indicators are calculated based on all sentiment classifications.
8. The ESG indicator evaluation method according to claim 7, characterized in that: The ESG indicator evaluation results of the enterprise to be evaluated under all primary evaluation indicators are calculated according to all sentiment classifications, including: By formula: Calculate the ESG indicator evaluation results of the company to be evaluated under all primary evaluation indicators ; in, Represents the first evaluation indicator The weight of Represents the first evaluation indicator The weight of Represents the first evaluation indicator The weight of Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is positive emotion among all corresponding encoding representations, Represents the first evaluation indicator The number of encoding representations whose emotion category is negative emotion among all corresponding encoding representations, It represents the sum of the number of coding representations with negative emotion category in all coding representations of the enterprise to be evaluated and the number of coding representations with positive emotion category in all coding representations.
9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the ESG indicator evaluation method according to any one of claims 1 to 8 is implemented.
10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the ESG indicator evaluation method according to any one of claims 1 to 8 is implemented.
Citation Information
Patent Citations
Mixing proportion estimation-based enterprise purchase and sales item tax risk prediction method
CN116011623A
Generation of explanatory summaries
US20130325440A1
ESG Supply Chain Forecasting
US20220327439A1
