Machine learning-based quantitative evaluation method and system for investor's attention to ESG (Enterprise Service Gateway)

Abstract

The invention belongs to the technical field of natural language processing, and particularly relates to a quantitative evaluation method and system for investors' attention to ESG based on machine learning, and the method comprises the steps: collecting and analyzing ESG report text content, carrying out the preprocessing, and carrying out the word embedding through a word2vec model, so as to obtain the vector representation of ESG related words. On the basis, seed words representing three dimensions of ESG are defined, and a word closest to the seed words is identified by using a word2vec model to expand the ESG related word dictionary. Finally, based on the dictionary, the attention degree of the institution investor on the ESG problem in field investigation and survey is calculated, and quantitative evaluation of the ESG attention degree of the institution investor is achieved. According to the method, the subjectivity of measurement is reduced, the accuracy is improved, and a reliable, efficient and economical technical means is provided for measurement of the ESG attention of the institution investor.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and specifically relates to a method and system for quantitatively evaluating investors' attention to ESG based on machine learning. Background Art

[0002] As global awareness of environmental protection and pollution reduction and carbon reduction increases, institutional investors are paying more and more attention to the environmental, social and governance (ESG) performance of companies, especially in heavily polluting industries. There are currently four main methods to measure the degree of attention paid by institutional investors to ESG. The first method is to directly understand the ESG attention of institutional investors through questionnaires. The second method is to use the weighted average of the ESG ratings of companies in the institutional investor portfolio as a measurement indicator. The third method is to evaluate the ESG attention of institutional investors by analyzing shareholder proposals, which is more common abroad. The fourth method is to measure their attention to ESG by analyzing the content of institutional investors' field research.

[0003] However, these existing measurement methods all have limitations and shortcomings to varying degrees. Although the questionnaire survey method is direct, it has the problems of inaccurate measurement and high cost, and the coverage is limited, the sample is not representative enough, and the quality of the answers is difficult to guarantee. The method of using the weighted average of ESG ratings is affected by the inconsistency of ratings. The standards and results of different rating agencies vary greatly, resulting in inconsistent measurement results. The shareholder proposal analysis method is more feasible abroad, but it is difficult to implement in China because data is difficult to obtain. The method of analyzing the content of institutional investors' field research has the problem of strong subjectivity and low efficiency in determining ESG-related terms, because the ESG-related terms manually determined by experts may not fully and accurately reflect the ESG information in the research content. Summary of the invention

[0004] The purpose of this invention is to provide a quantitative assessment method and system for investors' attention to ESG based on machine learning, so as to solve the problems that traditional measurement methods often rely on ESG-related words manually determined by experts, and have strong subjectivity and low efficiency.

[0005] The present invention achieves the above-mentioned purpose through the following technical solutions: In the first aspect, the present invention proposes a quantitative assessment method for investors' attention to ESG based on machine learning, the method comprising: S1. Collect ESG reports and parse their text content to obtain text data of ESG reports; S2, preprocessing the text data to obtain a corpus to be processed; S3, embedding the corpus to be processed based on the word2vec model to obtain vector representations of all words in the corpus to be processed to form a vector data set; wherein the word2vec model is trained using a skip-gram model based on negative sampling; S4. define seed words representing corresponding dimensions in the ESG report, use a pre-trained word2vec model to identify words similar to the seed words to expand the ESG-related words, and obtain an ESG-related word dictionary; S5. Determine investors’ attention to ESG based on the frequency of ESG-related words appearing in investors’ set question-and-answer scenarios.

[0006] Further, step S2 includes: S2.1, performing data cleaning on the text data; S2.2. Use Jieba word segmentation tool to segment the cleaned text data and delete special characters, numbers and punctuation marks to form a corpus to be processed.

[0007] Further, step S3 includes: S3.1. Call the python library to confirm the input and output, initialize the word2vec model parameters, set the context window size, word vector dimension, and the number of iterations of the corpus, and select the skip-gram model. S3.2, define the skip-gram model structure as follows: (1) Input layer: Each word in the corpus to be processed is represented as a word with a dimension equal to the vocabulary size One-hot row vector of dimension ; (2) Hidden layer: through Dimensional parameter matrix , projecting the one-hot vector of the input layer into Spatial ,in Corresponding to W The corresponding row vector; (3) Output layer: Dimensional parameter matrix , receiving the output of the hidden layer as input, predicting each word in the corpus to be processed Appear in The original probability of the context is obtained through the softmax activation function to obtain the true probability distribution. Zhongyu The corresponding column vector is expressed as ; S3.3, training model: negative sampling technology is used to optimize the skip-gram model, and the optimization objective function is: ; ; in is the positive sample set, is the negative sample set, Neighboring words of the central word From the positive sample set , Neighboring words of the central word From the negative sample set ; Using the gensim library, according to the parameters set in step S3.1 and the optimization objective function, traverse the corpus to be processed for iterative training until the set number of iterations is reached; S3.4. Obtain word vector representation and save the model: After training is completed, each word in the corpus to be processed is converted into a vector representation of a set dimension to form a vector data set, and the trained model is saved.

[0008] Further, step S4 includes: S4.1. Define seed words representing the environmental, social, and governance dimensions of ESG, and use the pre-trained word2vec model in step S3.4 to generate word vectors for any seed word in each ESG dimension; S4.2. Calculate the average vector of each ESG dimension, calculate the cosine similarity between the vectors of each word in the vector data set, and select the most similar set number of words for each dimension as ESG-related words; S4.3. Add the ESG-related words to the vector data set to obtain an ESG-related word dictionary.

[0009] Further, step S5 includes: Based on the ESG-related word dictionary, the investor's attention to ESG in a set question-and-answer scenario is calculated as follows: ; in Indicated in Year Company of investors pay attention to ESG. Indicates the company No. The frequency of ESG-related terms in ESG reports in 2018.

[0010] In a second aspect, the present invention proposes a quantitative evaluation system for investors' ESG attention based on machine learning, which is applied to perform any of the above-mentioned evaluation methods, and the system includes: A data collection module is used to collect ESG reports and parse their text content to obtain text data of the ESG reports; A preprocessing module, used for preprocessing the text data to obtain a corpus to be processed; A word embedding module, used for embedding the corpus to be processed based on a word2vec model to obtain vector representations of all words in the corpus to be processed and form a vector data set; wherein the word2vec model is trained using a skip-gram model based on negative sampling; A dictionary building module is used to define seed words representing corresponding dimensions in the ESG report, and use a pre-trained word2vec model to identify words similar to the seed words to expand the ESG-related words and obtain an ESG-related word dictionary; The calculation module is used to determine investors' attention to ESG based on the frequency of ESG-related words appearing in investors' set question-and-answer scenarios.

[0011] Furthermore, the preprocessing module also includes a word segmentation unit and a stop word deletion unit. The word segmentation unit is used to perform word segmentation using the Jieba word segmentation tool, and the stop word deletion unit is used to delete special characters, numbers and punctuation marks to form a corpus to be processed.

[0012] Furthermore, the word embedding module includes: Initialization unit, used to call the Python library to confirm input and output, to initialize the word2vec model parameters, and to set the context window size, word vector dimension, and the number of iterations of the corpus, and to select the skip-gram model; A model building unit is used to build a skip-gram model structure, including an input layer, a hidden layer, and an output layer, wherein the input layer represents each word in the corpus to be processed as a one-hot row vector with a dimension equal to the vocabulary size, the hidden layer projects the one-hot vector of the input layer into a space vector through a parameter matrix, and the output layer receives the output of the hidden layer as input through another parameter matrix, predicts the probability of each word in the corpus to be processed appearing in the context of a given central word, and obtains the true probability distribution through a softmax activation function; The model training unit uses the gensim library to iterate the corpus to be processed according to the parameters set by the initialization unit and the optimization objective function defined by the negative sampling optimization unit until the set number of iterations is reached; The word vector acquisition unit is used to convert each word in the corpus to be processed into a vector representation of a set dimension after training is completed, and form a vector data set.

[0013] Furthermore, the dictionary construction module includes: Seed word definition unit, used to define seed words representing the environmental, social, and governance dimensions of ESG; The word vector generation unit is used to generate a word vector for any seed word in each ESG dimension by combining the pre-trained word2vec model in the word embedding module; A similar word selection unit is used to calculate the average vector of each ESG dimension, and calculate the cosine similarity with each word vector in the vector data set, and select the most similar set number of words for each dimension as ESG related words; The dictionary construction unit is used to add the selected ESG related words to the vector data set to obtain the ESG related word dictionary.

[0014] The beneficial effects of the present invention are: The present invention introduces the machine learning model word2vec to determine ESG-related words, which significantly improves the existing method of measuring the ESG attention of institutional investors. By training the word2vec model, the present invention can automatically learn ESG-related vocabulary from a large number of ESG report texts, greatly reducing subjectivity and improving the accuracy and efficiency of measurement. At the same time, because the word2vec model can capture the semantic and syntactic similarities between words, it can more comprehensively reflect the attention of institutional investors to ESG issues in field research, providing investors, companies and regulators with more reliable and scientific ESG attention quantification indicators. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A flowchart of a method for quantitatively evaluating investors' ESG concerns based on machine learning proposed in the present invention; Figure 2 A flow chart of step S2 in the method for quantitatively evaluating investors' ESG attention based on machine learning proposed in the present invention; Figure 3 A flow chart of step S3 in the method for quantitatively evaluating investors' ESG attention based on machine learning proposed in the present invention; Figure 4 It is a schematic diagram of the structure of the skip-gram model in the present invention; Figure 5 A flow chart of step S4 in the method for quantitatively evaluating investors' ESG attention based on machine learning proposed in the present invention; Figure 6 This is a schematic diagram of the structure of the quantitative evaluation system for investors' ESG attention based on machine learning proposed in the present invention. DETAILED DESCRIPTION

[0016] The present application is further described in detail below in conjunction with the accompanying drawings. It is necessary to point out here that the following specific implementation methods are only used to further illustrate the present application and cannot be understood as limiting the scope of protection of the present application. Technical personnel in this field can make some non-essential improvements and adjustments to the present application based on the above application content.

[0017] Example 1

[0018] like Figure 1 As shown, this embodiment proposes a quantitative assessment method for investors' ESG attention based on machine learning, and the method includes: S1. Collect ESG reports and parse their text content to obtain text data of ESG reports; S2, preprocessing the text data to obtain a corpus to be processed; S3, embedding the corpus to be processed based on the word2vec model to obtain vector representations of all words in the corpus to be processed to form a vector data set; wherein the word2vec model is trained using the skip-gram model based on negative sampling; S4. Define seed words representing the corresponding dimensions in the ESG report, use the pre-trained word2vec model to identify words similar to the seed words to expand the ESG-related words, and obtain the ESG-related word dictionary; S5. Determine investors’ attention to ESG based on the frequency of ESG-related words appearing in investors’ set question-and-answer scenarios.

[0019] Combination Figure 2 , further preferably, step S2 comprises: S2.1, perform data cleaning on text data; S2.2. Use Jieba word segmentation tool to segment the cleaned text data and delete special characters, numbers and punctuation marks to form a corpus to be processed.

[0020] Combination Figure 3 , further preferably, step S3 comprises: S3.1. Call Python libraries: including gensim, pathlib and other libraries required for the entire word2vec model training process; based on gensim, use PathLineSentences to build an iterable input corpus and preliminarily build a word2vec model; confirm input and output, initialize model parameters, call the word2vec model, confirm the input corpus path and model output path, and set parameters such as context window size, word vector dimension, and number of corpus iterations, and select the skip-gram model; S3.2. Define the skip-gram model structure as: (1) Input layer: Each word in the corpus to be processed is represented as a word with a dimension equal to the vocabulary size One-hot row vector of dimension ; (2) Hidden layer: through Dimensional parameter matrix , projecting the one-hot vector of the input layer into Spatial ,in correspond Zhongyu The corresponding row vector; (3) Output layer: Dimensional parameter matrix , receiving the output of the hidden layer as input, predicting each word in the corpus to be processed Appear in The original probability of the context is obtained through the softmax activation function to obtain the true probability distribution. Zhongyu The corresponding column vector is expressed as .

[0021] like Figure 4 As shown in the figure, the skip-gram model training process of Gensim is optimized using negative sampling technology. The optimization objective function is: ; ; in is the positive sample set, is the negative sample set, Neighboring words of the central word From the positive sample set , Neighboring words of the central word From the negative sample set .

[0022] It should be noted that in this embodiment, word pairs that appear in the same window are set to be positive samples, and those that do not appear in the same window are set to be negative samples. During model training, the weight is updated once after each sample training. The more vocabulary samples there are, the greater the training workload. Since the number of negative samples is generally much larger than the number of positive samples, in order to solve the problem of training calculation speed, the present invention will perform sampling (random selection) calculations in negative samples, and only update part of the weights instead of calculating all of them, thereby reducing the calculation time.

[0023] S3.3, training model: using the gensim library, according to the parameters set in step S3.3 and the optimization objective function, traverse the corpus to be processed for iterative training until the set number of iterations is reached; S3.4. Obtain word vector representation and save the model: After training is completed, each word in the corpus to be processed is converted into a vector representation of a set dimension to form a vector data set, and the trained model is saved.

[0024] Combination Figure 4 , further preferably, step S4 comprises: S4.1. Define seed words representing the environmental, social, and governance dimensions of ESG, and generate word vectors for any seed word in each ESG dimension by combining the pre-trained word2vec model in step S3.4. S4.2, calculate the average vector of each ESG dimension, calculate the cosine similarity between the vectors of each word in the vector dataset, and select the most similar set number of words for each dimension as ESG-related words; S4.3. Add ESG related words to the vector dataset to obtain an ESG related word dictionary.

[0025] Further preferably, step S5 comprises: Based on the ESG-related word dictionary, we calculated investors’ concerns about ESG issues in a set question-and-answer scenario as follows: ; in Indicated in Year Company of investors pay attention to ESG. Indicates the company No. The frequency of ESG-related terms in ESG reports in 2018.

[0026] According to the above embodiment of the present invention, the above quantitative evaluation method comprises the following steps when implemented: 1. Data Collection and Preprocessing Data collection: ESG reports were manually downloaded from the public information website Juchao Information Network (http: / / www.cninfo.com.cn / new / index) in pdf format. These reports mainly disclose the performance and progress of companies in the three aspects of environment, society, and governance. The records of questions and answers from on-site investigations by institutional investors were downloaded from the China Research Data Service Platform (CNRDS) in txt format.

[0027] Data preprocessing: First, use Python to parse the ESG report in pdf format to obtain the text of the ESG report. Then, preprocess the parsed ESG text data, including data cleaning and word segmentation. Before inputting the ESG corpus into the word2vec model, perform a standard text preprocessing process to improve the efficiency of the machine learning training process. Specifically, use Jieba for word segmentation and remove stop words that contribute little to the semantic content of the sentence, such as "de", special characters, numbers, punctuation marks, etc.

[0028] Similarly, preprocess the text content of the Q&A section of on-site investigations by institutional investors, including data cleaning and word segmentation. Before inputting the ESG corpus into the word2vec model, perform a standard text preprocessing process to improve the efficiency of the machine learning training process. Specifically, use Jieba for word segmentation and remove stop words that contribute little to the semantic content of the sentence, such as "de", special characters, numbers, punctuation marks, etc.

[0029] II. Construction of an ESG-related word dictionary Determine seed words: Refer to domestic and foreign ESG research literature and rating standards, and combine the actual situation in China to select seed words representing the three dimensions of environment, society, and governance. For example, for the environmental dimension, words such as "carbon emissions" and "energy conservation and emission reduction" can be selected; for the social dimension, words such as "employee rights and interests" and "social responsibility" can be selected; for the governance dimension, words such as "board structure" and "internal control" can be selected.

[0030] Train the word2vec model: Use the preprocessed ESG report text content as input to train the word2vec model. This model can map each word in the text to a high-dimensional vector space, so that words with similar semantics are closer in the vector space.

[0031] Expand the ESG-related word dictionary: Use the trained word2vec model to calculate the similarity between each seed word and other words in the text, and select the N words with the highest similarity as the expansion words of the seed word. Combine these expansion words with the seed words to form an ESG-related word dictionary.

[0032] III. Quantitative assessment of institutional investors’ ESG attention Based on the ESG-related word dictionary obtained by expanding the above-mentioned word2vec model, the frequency of ESG-related words in the question and answer text is counted using Python's word frequency statistics program for the company's individual annual institutional investor field survey records. The sum of the statistical ESG-related word frequencies is added by 1 and then the natural logarithm is taken to obtain the institutional investors' ESG attention.

[0033] Through the above method, the present invention can achieve quantitative evaluation of institutional investors' ESG attention, which has the following advantages: High accuracy: By training the word2vec model and building an ESG-related word dictionary, we can accurately extract ESG-related keywords, thereby accurately assessing the ESG attention of institutional investors.

[0034] Strong objectivity: This method is based on machine learning and text analysis technology, which avoids subjective interference from human factors and makes the evaluation results more objective and reliable.

[0035] Strong practicability: The data required for this method is easy to obtain and process, and the calculation process is simple and efficient, which is suitable for the analysis and evaluation of large-scale data.

[0036] Combination Figure 5 Based on the same inventive concept, this embodiment also proposes a quantitative evaluation system for investors' ESG attention based on machine learning, which is applied to execute the above evaluation method. The system includes: A data collection module is used to collect ESG reports and parse their text content to obtain text data of the ESG reports; A preprocessing module is used to preprocess text data to obtain a corpus to be processed; The word embedding module is used to embed words in the corpus to be processed based on the word2vec model to obtain vector representations of all words in the corpus to be processed and form a vector data set; the word2vec model is trained using the skip-gram model based on negative sampling; The dictionary building module is used to define seed words representing the corresponding dimensions in the ESG report, and use the pre-trained word2vec model to identify words similar to the seed words to expand the ESG-related words and obtain the ESG-related word dictionary; The calculation module is used to determine investors' attention to ESG based on the frequency of ESG-related words appearing in investors' set question-and-answer scenarios.

[0037] It should be noted here that each module in the above quantitative evaluation system corresponds to steps S1 to S5 in implementing the above quantitative evaluation method, and the examples and application scenarios implemented by multiple modules and corresponding steps are the same, but are not limited to the contents disclosed in the above embodiment 1.

[0038] Further preferably, the preprocessing module also includes a word segmentation unit and a stop word deletion unit, the word segmentation unit is used to perform word segmentation using the Jieba word segmentation tool, and the stop word deletion unit is used to delete special characters, numbers and punctuation marks to form a corpus to be processed.

[0039] Further preferably, the word embedding module includes: Initialization unit, used to call the Python library to confirm input and output, to initialize the word2vec model parameters, and to set the context window size, word vector dimension, and the number of iterations of the corpus, and to select the skip-gram model; A model building unit is used to build a skip-gram model structure, including an input layer, a hidden layer, and an output layer, wherein the input layer represents each word in the corpus to be processed as a one-hot row vector with a dimension equal to the vocabulary size, the hidden layer projects the one-hot vector of the input layer into a space vector through a parameter matrix, and the output layer receives the output of the hidden layer as input through another parameter matrix, predicts the probability of each word in the corpus to be processed appearing in the context of a given central word, and obtains the true probability distribution through a softmax activation function; The model training unit uses the gensim library to iterate the corpus to be processed according to the parameters set by the initialization unit and the optimization objective function defined by the negative sampling optimization unit until the set number of iterations is reached; The word vector acquisition unit is used to convert each word in the corpus to be processed into a vector representation of a set dimension after training is completed, and form a vector data set.

[0040] Further preferably, the dictionary building module includes: Seed word definition unit, used to define seed words representing the environmental, social, and governance dimensions of ESG; The word vector generation unit is used to generate a word vector for any seed word in each ESG dimension by combining the pre-trained word2vec model in the word embedding module; A similar word selection unit is used to calculate the average vector of each ESG dimension, and calculate the cosine similarity with each word vector in the vector data set, and select the most similar set number of words for each dimension as ESG related words; The dictionary construction unit is used to add the selected ESG related words to the vector data set to obtain the ESG related word dictionary.

[0041] In another embodiment of the present invention, a computer-readable storage medium is provided, in which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above quantitative evaluation methods are implemented.

[0042] In another embodiment of the present invention, a computer program product including instructions is provided. When the computer program product is run on a computer, the computer executes any quantitative evaluation method in the above embodiments.

[0043] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated.

[0044] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated therein. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a magnetic tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0045] The above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features can be replaced by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A quantitative assessment method of investors’ ESG attention based on machine learning, characterized by: The method comprises: S1. Collect ESG reports and parse their text content to obtain text data of ESG reports; S2, preprocessing the text data to obtain a corpus to be processed; S3, embedding the corpus to be processed based on the word2vec model to obtain vector representations of all words in the corpus to be processed to form a vector data set; wherein the word2vec model is trained using a skip-gram model based on negative sampling; S4. define seed words representing corresponding dimensions in the ESG report, use a pre-trained word2vec model to identify words similar to the seed words to expand the ESG-related words, and obtain an ESG-related word dictionary; S5. Determine investors’ attention to ESG based on the frequency of ESG-related words appearing in investors’ set question-and-answer scenarios.

2. The method for quantitatively evaluating investors' attention to ESG based on machine learning according to claim 1, characterized in that: Step S2 includes: S2.1, performing data cleaning on the text data; S2.

2. Use Jieba word segmentation tool to segment the cleaned text data and delete special characters, numbers and punctuation marks to form a corpus to be processed.

3. The method for quantitatively evaluating investors' attention to ESG based on machine learning according to claim 1, characterized in that: Step S3 includes: S3.

1. Call the python library to confirm the input and output, initialize the word2vec model parameters, set the context window size, word vector dimension, and the number of iterations of the corpus, and select the skip-gram model. S3.2, define the skip-gram model structure as follows: (1) Input layer: Each word in the corpus to be processed is represented as a word with a dimension equal to the vocabulary size One-hot row vector of dimension ; (2) Hidden layer: through Dimensional parameter matrix , projecting the one-hot vector of the input layer into Spatial ,in Corresponding to W The corresponding row vector; (3) Output layer: Dimensional parameter matrix , receiving the output of the hidden layer as input, predicting each word in the corpus to be processed Appear in The original probability of the context is obtained through the softmax activation function to obtain the true probability distribution. Zhongyu The corresponding column vector is expressed as ; S3.3, training model: negative sampling technology is used to optimize the skip-gram model, and the optimization objective function is: ; ; in is the positive sample set, is the negative sample set, Neighboring words of the central word From the positive sample set , Neighboring words of the central word From the negative sample set ; Using the gensim library, according to the parameters set in step S3.1 and the optimization objective function, traverse the corpus to be processed for iterative training until the set number of iterations is reached; S3.

4. Obtain word vector representation and save the model: After training is completed, each word in the corpus to be processed is converted into a vector representation of a set dimension to form a vector data set, and the trained model is saved.

4. The method for quantitatively evaluating investors' attention to ESG based on machine learning according to claim 3 is characterized in that: Step S4 includes: S4.

1. Define seed words representing the environmental, social, and governance dimensions of ESG, and use the pre-trained word2vec model in step S3.4 to generate word vectors for any seed word in each ESG dimension; S4.

2. Calculate the average vector of each ESG dimension, calculate the cosine similarity between the vectors of each word in the vector data set, and select the most similar set number of words for each dimension as ESG-related words; S4.

3. Add the ESG-related words to the vector data set to obtain an ESG-related word dictionary.

5. The method for quantitatively evaluating investors' attention to ESG based on machine learning according to claim 1, characterized in that: Step S5 includes: Based on the ESG-related word dictionary, the investor's attention to ESG in a set question-and-answer scenario is calculated as follows: ； in Indicated in Year Company of investors pay attention to ESG. Indicates the company No. The frequency of ESG-related terms in ESG reports in 2018.

6. A quantitative evaluation system for investors' ESG attention based on machine learning, applied to the evaluation method according to any one of claims 1 to 5, characterized in that: The system comprises: A data collection module is used to collect ESG reports and parse their text content to obtain text data of the ESG reports; A preprocessing module, used for preprocessing the text data to obtain a corpus to be processed; A word embedding module, used for embedding the corpus to be processed based on a word2vec model to obtain vector representations of all words in the corpus to be processed and form a vector data set; wherein the word2vec model is trained using a skip-gram model based on negative sampling; A dictionary building module is used to define seed words representing corresponding dimensions in the ESG report, and use a pre-trained word2vec model to identify words similar to the seed words to expand the ESG-related words and obtain an ESG-related word dictionary; The calculation module is used to determine investors' attention to ESG based on the frequency of ESG-related words appearing in investors' set question-and-answer scenarios.

7. The quantitative evaluation system for investors' ESG attention based on machine learning according to claim 6 is characterized in that: The preprocessing module also includes a word segmentation unit and a stop word deletion unit. The word segmentation unit is used to perform word segmentation using the Jieba word segmentation tool, and the stop word deletion unit is used to delete special characters, numbers and punctuation marks to form a corpus to be processed.

8. The quantitative evaluation system for investors' ESG attention based on machine learning according to claim 6 is characterized in that: The word embedding module includes: Initialization unit, used to call the Python library to confirm input and output, to initialize the word2vec model parameters, and to set the context window size, word vector dimension, and the number of iterations of the corpus, and to select the skip-gram model; A model building unit is used to build a skip-gram model structure, including an input layer, a hidden layer, and an output layer, wherein the input layer represents each word in the corpus to be processed as a one-hot row vector with a dimension equal to the vocabulary size, the hidden layer projects the one-hot vector of the input layer into a space vector through a parameter matrix, and the output layer receives the output of the hidden layer as input through another parameter matrix, predicts the probability of each word in the corpus to be processed appearing in the context of a given central word, and obtains the true probability distribution through a softmax activation function; The model training unit uses the gensim library to iterate the corpus to be processed according to the parameters set by the initialization unit and the optimization objective function defined by the negative sampling optimization unit until the set number of iterations is reached; The word vector acquisition unit is used to convert each word in the corpus to be processed into a vector representation of a set dimension after training is completed, and form a vector data set.

9. The quantitative evaluation system for investors' ESG attention based on machine learning according to claim 8 is characterized in that: The dictionary building module includes: Seed word definition unit, used to define seed words representing the environmental, social, and governance dimensions of ESG; The word vector generation unit is used to generate a word vector for any seed word in each ESG dimension by combining the pre-trained word2vec model in the word embedding module; A similar word selection unit is used to calculate the average vector of each ESG dimension, and calculate the cosine similarity with each word vector in the vector data set, and select the most similar set number of words for each dimension as ESG related words; The dictionary construction unit is used to add the selected ESG related words to the vector data set to obtain the ESG related word dictionary.

Images