Method for judging greening probability of listed company
By leveraging text and financial data with advanced natural language processing and ensemble learning, the method effectively addresses the low accuracy of existing greenwashing detection, achieving high precision in predicting greenwashing behavior of listed companies.
Patent Information
- Application Number
- CN202410943163.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-15
- Publication Date
- 2024-12-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing methods for identifying corporate greenwashing have low accuracy in assessing and predicting the behavior of listed companies.
A method using public text and financial data from listed companies, extracting environmental disclosure information through seed words and large language models, and training predictive models with ensemble learning to determine the probability of greenwashing.
The method achieves high accuracy in automatically calculating the probability of greenwashing, with an AUC higher than 0.90 and improved recall and accuracy in identifying greenwashing companies.
Smart Images

Figure CN119090040A_ABST
Abstract
Description
Technical Field
[0001] The present invention specifically relates to a method for judging the probability of greenwashing of listed companies. Background Art
[0002] Currently, existing technologies mainly identify instances of corporate greenwashing by comparing symbolic activities with substantive activities. For example, Walker and Wan collected text information disclosed by companies, manually coded this information into symbolic and substantive categories at different levels, and measured greenwashing by subtracting substantive activities from symbolic activities.
[0003] However, the existing assessment and prediction methods of corporate greenwashing behavior have low accuracy. Summary of the invention
[0004] The present invention provides a method for determining the probability of greenwashing of listed companies to solve the above-mentioned technical problems, and specifically adopts the following technical solutions:
[0005] A method for judging the probability of greenwashing of listed companies, comprising:
[0006] Obtaining public text data and financial data of several listed companies;
[0007] Extract environmental disclosure information from text data;
[0008] Train predictive models based on environmental disclosure information and financial data;
[0009] Input the data of the company to be evaluated into the trained prediction model to obtain the prediction results.
[0010] Furthermore, the text data includes annual reports and corporate social responsibility reports.
[0011] Furthermore, the specific method of extracting environmental disclosure information from text data is:
[0012] Get some seed words about the environment;
[0013] Extract several environmental keywords from the text data based on the seed words, and retrieve corresponding sentences from the annual report and corporate social responsibility report;
[0014] Eliminate reports and sentences that do not contain environmental keywords;
[0015] Annotate the remaining sentences after screening;
[0016] Train a large language model with annotated sentences;
[0017] Several sentences mentioning environmental issues are extracted from the initial text data through the trained large language model.
[0018] Furthermore, the specific method of obtaining several seed words related to the environment is:
[0019] Several seed words related to the environment were extracted from the government work report through the LDA topic analysis method.
[0020] Furthermore, several environmental keywords were extracted from the text data based on the seed words through the Word2Vec algorithm, and the corresponding sentences were retrieved from the annual report and corporate social responsibility report.
[0021] Furthermore, the large language model is a Bert model.
[0022] Furthermore, before annotating the remaining sentences after screening, the remaining sentences after screening are first randomly deleted to retain a predetermined number of sentences.
[0023] Furthermore, the financial data includes cash and cash equivalents, accounts receivable, provision for bad debts, inventory, current assets, fixed assets, total assets, accounts payable, short-term loans, current liabilities, deferred income tax assets, long-term liabilities, surplus reserves, retained earnings, operating income, operating costs, depreciation of fixed assets, amortization of intangible assets, long-term deferred expenses, sales expenses, administrative expenses, financial expenses, investment income, income tax, deferred income tax, operating profit, net profit, equity cash flow, net cash flow from financing activities, year-end closing price, number of outstanding shares, number of employees, net cash flow from investment activities, detailed accounts of construction in progress and detailed accounts of administrative expenses.
[0024] Furthermore, the detailed accounts of the projects under construction and the detailed accounts of the management expenses in the acquired financial data are screened and the items related to environmental protection are retained.
[0025] The invention is beneficial in that the method for determining the probability of greenwashing of listed companies provided uses text information and financial data disclosed by listed companies to automatically calculate the possibility of greenwashing of companies. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, 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 creative labor.
[0027] Figure 1 It is a schematic diagram of a method for judging the probability of greenwashing of listed companies according to the present invention. DETAILED DESCRIPTION
[0028] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present application, and should not be construed as limiting the present application.
[0029] like Figure 1 The method for judging the probability of greenwashing of listed companies of the present application is shown, comprising: S1: obtaining text data and financial data disclosed by a number of listed companies. S2: extracting environmental disclosure information from the text data. S3: training a prediction model based on the environmental disclosure information and financial data. S4: inputting the data of the company to be evaluated into the trained prediction model to obtain a prediction result. The method for judging the probability of greenwashing of listed companies of the present application uses the text information and financial data disclosed by listed companies to automatically calculate the possibility of greenwashing of the company. The above steps are described in detail below.
[0030] For step S1: obtain text data and financial data disclosed by several listed companies.
[0031] In an embodiment of the present application, the text data includes an annual report and a corporate social responsibility (CSR) report.
[0032] Specifically, the annual reports and CSR reports of listed companies were downloaded from the CNINF website (link address: http: / / www.cninfo.com.cn). By searching the names of listed companies, all their public disclosures were obtained, and the annual reports and CSR reports were selected and downloaded.
[0033] Then download the financial data of listed companies from the CSMAR database, including cash and cash equivalents, accounts receivable, bad debt reserves, inventory, current assets, fixed assets, total assets, accounts payable, short-term loans, current liabilities, deferred income tax assets, long-term liabilities, surplus reserves, retained earnings, operating income, operating costs, depreciation of fixed assets, amortization of intangible assets, long-term deferred expenses, sales expenses, administrative expenses, financial expenses, investment income, income tax, deferred income tax, operating profit, net profit, equity cash flow, net cash flow from financing activities, year-end closing price, number of outstanding shares, number of employees, net cash flow from investment activities, detailed accounts of construction in progress and detailed accounts of administrative expenses.
[0034] In the implementation mode of the present application, the detailed accounts of construction in progress and the detailed accounts of management expenses in the acquired financial data are screened and the items related to environmental protection are retained. Specifically, the purpose of each investment listed in the footnote is specially checked. The broad purpose of investment related to the environment is defined, such as environmental protection facilities and equipment and environmental protection technology transformation.
[0035] For step S2: extract environmental disclosure information from text data.
[0036] In the implementation manner of this application, the specific method for extracting environmental disclosure information from text data is:
[0037] Use the RE library to split the text data into sentences.
[0038] Several seed words related to the environment are extracted from the government work reports from 2001 to 2024 through the LDA topic analysis method.
[0039] Through the Word2Vec algorithm, several environmental keywords are extracted from the text data based on the seed words, and the corresponding sentences are retrieved from the annual report and corporate social responsibility report.
[0040] Reports and sentences that do not contain environmental keywords are eliminated to ensure that the subsequent further analysis is based entirely on corporate environmental information disclosure.
[0041] Randomly delete the remaining sentences after screening, and keep a predetermined number of sentences. Specifically, randomly select a predetermined number of sentences
[0042] The remaining sentences after screening are annotated, and the large language model is trained through the annotated sentences. In this application, the large language model is a Bert model.
[0043] Large Language Models (LLMs) are a deep learning-based natural language processing (NLP) technique designed to understand and generate human language. These models are trained on large text datasets to learn the structure, grammar, semantics, and context of a language, and are able to perform a variety of language tasks. LLMs are usually pre-trained on a large amount of text data that covers a wide range of topics and domains, ensuring that the model has good generalization capabilities. These models are usually based on deep neural networks, especially the Transformer architecture, which allows the model to capture long-range dependencies and complex language patterns. LLMs are first pre-trained on a large general corpus to learn a general language representation. The model can then be adapted to specific application scenarios by fine-tuning on smaller datasets for specific tasks. BERT is able to improve the performance of natural language processing (NLP) tasks. The core innovation of the BERT model lies in its bidirectional processing ability for text, which can take into account both the left and right context of a given word, thereby obtaining a richer and more accurate language representation. BERT can take into account the contextual information before and after the word at the same time, which makes the model more accurate in understanding word meaning and sentence structure. The BERT model is first pre-trained on a large-scale text dataset to learn general language representation. Then, by fine-tuning on a task-specific dataset, the model can adapt to a variety of different NLP tasks. With a training set to test set ratio of 9:1, the model's classification accuracy reaches 91.85%.
[0044] Then, the trained large language model is used to extract several sentences that mention environmental issues from the initial text data, and the environmental disclosures of each company are summarized.
[0045] Ingram & Frazier (1980) developed 13 content categories representing corporate environmental disclosures. These categories reflect four testable dimensions that appear in disclosures: evidence, time, specificity, and subject matter. All text data from CSR reports were used to train the Word2Vec model. The semantics and syntax of words were used to compile a list of words most relevant to Ingram and Frazier's (1980) categories. Finally, the TF-IDF weighted word frequency of the subject words of each category in corporate environmental disclosures was counted.
[0046] Among them, the categories obtained according to Ingram & Frazier (1980) are shown as follows:
[0047] Table 113 Representative categories of corporate environmental disclosures
[0048]
[0049] The seed words for each category are established based on existing literature as follows:
[0050] Table 2 Seed words corresponding to the 13 content categories representing corporate environmental disclosures
[0051]
[0052]
[0053] All the text data in the CSR report was used to train the Word2Vec model, and an extended dictionary was generated for each of the above dimensions based on the Word2Vec model. The following are examples of three dimensions:
[0054] Table 3 shows the extended dictionary of three dimensions
[0055]
[0056]
[0057] Using this extended dictionary, the TF-IDF score of the importance of each category of vocabulary in the company's annual environmental disclosures is calculated.
[0058] The TF-IDF weight score calculation process for each category is as follows:
[0059] 1. Calculate term frequency (TF)
[0060]
[0061] Among them, n t,d is the number of times word t appears in document d, N d is the total number of words in document d.
[0062] 2. Calculate Inverse Document Frequency (IDF)
[0063]
[0064] Where |D| is the total number of documents in the collection, and |d:t∈d| is the number of documents containing word t.
[0065] 3. Calculate TF-IDF weight
[0066] TF-IDF weight is the result of multiplying the term frequency TF and the inverse document frequency IDF.
[0067] TF-IDF(t,d)=TF(t,d)×IDF(t)
[0068] After calculating the TF-IDF weighted word frequency of each company's annual environmental disclosure text, we obtained environmental disclosure metrics for 26,081 companies-years. The descriptive statistics are as follows:
[0069] Table 4 Environmental disclosure metrics
[0070]
[0071] For step S3: training a prediction model based on environmental disclosure information and financial data.
[0072] Traditional machine learning methods (such as SVM, decision tree or logistic regression) usually only generate one estimator, while ensemble learning methods combine the predictions of a set of basic estimators (such as decision trees) to improve generalization and robustness. In this application, an ensemble learning method is used to predict greenwashing of Chinese listed companies.
[0073] AdaBoost is one of the most important boosting methods. However, due to the class imbalance problem in many prediction situations, traditional ensemble learning methods usually need to be combined with sampling techniques to balance the class distribution of the training data, by adding examples to the minority class (oversampling) or removing examples from the majority class. RUSBoost is a variant of AdaBoost that uses random undersampling (RUS) to solve the class imbalance learning problem. It works very similarly to AdaBoost, except that RUS is performed in each iteration to solve the imbalance of fraudulent and non-fraudulent companies. Specifically, the RUS algorithm uses a full sample of fraudulent companies during training and a subsample of non-fraudulent companies randomly generated during the same training period when training weak classifiers in each iteration. RUSBoost estimates the ratio between the number of majority class observations (i.e., non-greenwashing samples) and minority class observations (i.e., greenwashing samples) that need to select insufficient samples. In this application, the ratio is set to 1:2 to build the RUSBoost model.
[0074] The input of the RUSBoost model is an unbalanced training dataset where x ij is the jth feature of the i-th sample, y i ∈{0,1} is the greenwashing category variable of the i-th sample. In this application, y i =1 indicates that observation i is a greenwashing company detected by Southern Weekend magazine. i =1 accounts for less than 1% of the data set D.
[0075] The output of the RUSBoost model is a strong classifier F(x).
[0076] The execution steps of the RUSBoost algorithm are as follows:
[0077] 1. Initialize sample weight distribution:
[0078] where i=1,2,...,N
[0079] 2. Iteration process (M rounds in total)
[0080] For each iteration m=1,2,…,M:
[0081] a. Random undersampling: using the current weight distribution D m Randomly undersample the training set to obtain a subset
[0082] b. Training weak classifiers: In the subset Train a weak classifier h on m (x).
[0083] c. Calculate the weighted average error rate:
[0084] Where IfTrue(·) is an indicator function, which takes the value 1 when the condition in the brackets is true, and 0 otherwise.
[0085] d. Calculate the weight of weak classifier:
[0086] e. Update sample weight distribution
[0087] Where Z m is the normalization factor, so that D m+1 becomes a probability distribution:
[0088]
[0089] 3. Output the final strong classifier:
[0090]
[0091] The model training dataset D contains three types of indicators: text measurement indicators, financial indicators, and environmental investment indicators.
[0092] The text measurement index includes the TF-IDF score weight information of 13 categories calculated on the corporate environmental disclosure text in step S2, referring to Ingram and Frazier (1980).
[0093] Among them, the financial measurement indicators include 34 original financial data items and 4 environmental investment data items. The 34 original financial data items are directly obtained from the CSMAR database, referring to the indicators in the financial fraud prediction model of Cecchini et al. (2010) and Bao et al. (2020).
[0094] Environmental investment data items are generated from the construction projects in progress and management expense items in the CSMAR database using environmental keyword matching, and are divided into four categories with reference to Liang and Renneboog (2017): pollutant control, clean energy, ecological restoration, and integrated environmental governance.
[0095] The list of model input data is as follows:
[0096] Table 5 Model input data
[0097]
[0098]
[0099] The model input data is normalized to obtain a dataset D of 26,081 company-year input variables.
[0100] For step S4: the data of the company to be evaluated is input into the trained prediction model to obtain the prediction result.
[0101] After the prediction model is trained, the corresponding public data of a company can be input into the model to automatically output the prediction results.
[0102] Specifically, the model F(x) trained in step S3, where x = (x1, x2, ..., x k ) is the public data of the new company. The model will output: (1) the binary prediction result y∈{0,1}; (2) the probability of being predicted as 1 P(y=1|x).
[0103] Algorithm steps:
[0104] 1. Calculate the binary classification prediction result y:
[0105] y=sign(F(x))
[0106] in,
[0107] 2. Calculate the probability of being predicted as 1 P(y=1|x):
[0108]
[0109] The logistic sigmoid function is used here to convert the model output into a probability value.
[0110] Finally, the binary classification prediction result y and the probability of being predicted as 1 P(y=1|x) are returned.
[0111] The following is the out-of-sample model prediction effect matrix:
[0112] Table 6 Out-of-sample model prediction effect matrix
[0113]
[0114] Area under the receiver operating characteristic curve (AUC). AUC is the area under the receiver operating characteristic (ROC) curve that combines the true positive rate (i.e., recall rate) and the false positive rate (i.e., precision rate) in one graph. An AUC higher than 0.5 indicates that the model is acceptable. Using different test years, the AUC of this model is higher than 0.90.
[0115] The standardized discounted cumulative benefit (NDCG@k) of position k, where k represents the performance indicator, which is the measurement result among the top k% of greenwash probability in the test year, and k∈(1,2,3,4,5). The final reported NDCG@k is the mean of NDCG@k when k is 1%, 2%, 3%, 4% and 5%. DCG@k is defined as: If the i-th observation in the ranked list is a true greenwash, then rel i =1, otherwise 0. NDCG@k is DCG@k normalized by theoretical DCG@k, i.e., NDCG@k = DCG@k / theoretical DCG@k, where theoretical DCG@k is the DCG@k value when all real greenwashing cases are ranked at the top of the ranking list. The NDCG@k obtained by the present invention is higher than that of the similar model Bao (2020).
[0116] The recall value is determined by classifying the top k% of companies with the highest predicted probability of greenwashing as greenwashing companies in the test year. Specifically, recall = TP / TP + FN, where TP is the number of true greenwashing companies included in the top k% of companies predicted to be greenwashing in the test year, and FN is the number of true greenwashing companies that are misclassified as non-greenwashing companies in the bottom 1-k% of companies observed in the test year. The sum of TP and FN is the total number of true greenwashing companies in a test year. The final reported recall is the mean of the recall when k is 1%, 2%, 3%, 4%, and 5%.
[0117] The accuracy is determined by classifying the top k% of companies with the highest predicted greenwashing probability in the test year as greenwashing companies. Specifically, accuracy = TP / TP+FP, where TP is the number of true greenwashes included in the predicted greenwashes in the first k% of the test year, and FP is the number of false greenwashes that are misclassified as greenwashes in the predicted greenwashes in the first k% of the test year. The sum of TP and FP is the number of companies ranked in the top k% in the test year. The final reported accuracy is the mean of the accuracy when k is 1%, 2%, 3%, 4%, and 5%.
[0118] Out-of-sample performance tests show that the present invention has better performance than existing models such as Bao (2020).
[0119] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any form, and any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.
Claims
1. A method for determining the probability of greenwashing of listed companies, characterized in that: Include: Obtaining public text data and financial data of several listed companies; Extract environmental disclosure information from text data; Train predictive models based on environmental disclosure information and financial data; Input the data of the company to be evaluated into the trained prediction model to obtain the prediction results.
2. The method for determining the probability of greenwashing of listed companies according to claim 1, characterized in that: The text data includes annual reports and corporate social responsibility reports.
3. The method for determining the probability of greenwashing of listed companies according to claim 2, characterized in that: The specific method for extracting environmental disclosure information from text data is: Split text data into sentences; Get some seed words about the environment; Extract several environmental keywords from the text data based on the seed words, and retrieve corresponding sentences from the annual report and corporate social responsibility report; Eliminate reports and sentences that do not contain environmental keywords; Annotate the remaining sentences after screening; Train a large language model with annotated sentences; Several sentences mentioning environmental issues are extracted from the initial text data through the trained large language model.
4. The method for determining the probability of greenwashing of listed companies according to claim 3, characterized in that: The specific method of obtaining several seed words related to the environment is: Several seed words related to the environment were extracted from the government work report through the LDA topic analysis method.
5. The method for determining the probability of greenwashing of listed companies according to claim 3, characterized in that: Through the Word2Vec algorithm, several environmental keywords are extracted from the text data based on the seed words, and the corresponding sentences are retrieved from the annual report and corporate social responsibility report.
6. The method for determining the probability of greenwashing of listed companies according to claim 3, characterized in that: The large language model is a Bert model.
7. The method for determining the probability of greenwashing of listed companies according to claim 3, characterized in that: Before annotating the remaining sentences after screening, the remaining sentences after screening are randomly deleted to keep a predetermined number of sentences.
8. The method for determining the probability of greenwashing of listed companies according to claim 1, characterized in that: The financial data includes cash and cash equivalents, accounts receivable, provision for bad debts, inventory, current assets, fixed assets, total assets, accounts payable, short-term loans, current liabilities, deferred income tax assets, long-term liabilities, surplus reserves, retained earnings, operating income, operating costs, depreciation of fixed assets, amortization of intangible assets, long-term deferred expenses, sales expenses, administrative expenses, financial expenses, investment income, income tax, deferred income tax, operating profit, net profit, equity cash flow, net cash flow from financing activities, year-end closing price, number of outstanding shares, number of employees, net cash flow from investing activities, detailed accounts of construction in progress and detailed accounts of administrative expenses.
9. The method for determining the probability of greenwashing of listed companies according to claim 8, characterized in that: The detailed accounts of construction in progress and administrative expenses in the acquired financial data are screened and the items related to environmental protection are retained.
Citation Information
Cited By
Enterprise bleaching detection method for structured information extraction and related equipment
CN121234949A
A structured information extraction method for detecting greenwashing of enterprises and related equipment
CN121234949B
