Enterprise ESG index determination method and system based on artificial intelligence

Abstract

The invention provides an enterprise ESG (Enterprise Service Gateway) index determination method and system based on artificial intelligence, and aims at automatically and efficiently determining an ESG (Enterprise Service Gateway) index of an enterprise. The method comprises the following steps: data collection and processing: collecting and preprocessing enterprise announcements, news reports and social media information by utilizing a web crawler and an API (Application Program Interface); feature extraction: converting the preprocessed text data into feature vectors by adopting a BERT model; constructing a knowledge graph, and performing analysis and prediction based on the extracted features; model training: training the knowledge graph to form an ESG scoring model; and finally, index calculation and calibration are carried out, the trained model is applied to new data, the ESG index of the enterprise is calculated, and the model is adjusted and optimized by using actual tracking performance.

Description

Technical Field

[0001] The present invention belongs to the field of enterprise management, and more specifically to an enterprise ESG index determination method and system based on artificial intelligence. Background Art

[0002] The ESG (Environment, Social, Governance) index is a tool for measuring a company's environmental, social and corporate governance performance, and is widely used in socially responsible investment, portfolio management and corporate social responsibility assessment. However, traditional ESG assessments usually rely on subjective judgment and manual operations by experts, which is time-consuming and difficult to obtain large-scale data, and the efficiency and accuracy of the assessment are limited.

[0003] At the same time, public information sources such as corporate announcements, news reports and social media contain a large amount of ESG-related information. This information cannot be effectively extracted and utilized in massive text data, which greatly limits the effectiveness of ESG assessment.

[0004] To this end, with the help of artificial intelligence technology, especially natural language processing (NLP), deep learning and reinforcement learning technology, the above problems can be effectively solved and the efficiency and accuracy of ESG assessment can be improved.

[0005] Natural language processing technology can convert unstructured text information into structured data, deep learning technology can self-learn and optimize through large-scale data, and reinforcement learning technology allows the model to find the optimal strategy through continuous trial and error to improve the model's adaptability and predictive ability.

[0006] However, how to effectively integrate these technologies and develop a model and method suitable for corporate ESG assessment is still an important scientific problem that has not been solved. The present invention attempts to solve this problem and provides a method for determining corporate ESG index based on artificial intelligence. Summary of the invention

[0007] The technical problem to be solved by the present invention is how to effectively utilize artificial intelligence technologies, including natural language processing, deep learning, and reinforcement learning, to automatically extract and analyze ESG-related feature information from a large amount of unstructured public information such as corporate announcements, news reports, and social media, construct an enterprise ESG knowledge graph, and train a model based on this to automatically, quickly, and accurately determine the enterprise ESG index.

[0008] In order to achieve the above object, the present invention is implemented by adopting the following technical scheme: the method comprises:

[0009] Data collection and processing: Collect corporate announcements, news reports, and social media information, and use natural language processing technology for pre-processing, including entity recognition and keyword extraction;

[0010] Feature extraction: preprocessed text data is converted into feature vectors for quantitative analysis;

[0011] Knowledge graph construction: Analyze and predict the extracted features to construct the enterprise’s ESG knowledge graph;

[0012] Model training: Train the knowledge graph to form an ESG scoring model;

[0013] Index calculation and calibration: Apply the trained model to new data, calculate the company's ESG index, and use actual tracking performance to adjust and optimize the model.

[0014] In one embodiment, the data collection and processing includes:

[0015] Use web crawlers to collect public information from announcements and news reports of relevant companies on the Internet; use developer APIs of social media to obtain data;

[0016] The data preprocessing stage includes text cleaning and standardization steps, which include removing garbled characters and non-text content, punctuation, auxiliary words and other stop words, HTML tags, URLs, and identifying the language of the text.

[0017] In one solution, the feature extraction uses a BERT model, including:

[0018] First, the preprocessed text data is input into the BERT model; the bidirectional encoder is used to understand the context information to generate a more accurate text representation;

[0019] In the input stage, the text data is first segmented into word fragments, which are encoded by word embedding, position embedding, and segment embedding. Word embedding maps each word fragment to a high-dimensional vector space, position embedding provides position information for each word fragment, and segment embedding is used to distinguish different sentences or paragraphs.

[0020] Next, the input data processed by the embedding layer is processed by a multi-layer Transformer encoder; each layer of the encoder consists of a multi-head self-attention mechanism and a feedforward neural network; the core of the self-attention mechanism is to capture global context information by calculating the correlation between each word fragment and other word fragments in the input sequence; specifically, given the input matrix X, the self-attention calculation is expressed as:

[0021]

[0022] Among them, Q, K, and V are query, key, and value matrices, respectively, which are usually obtained by linearly transforming the input matrix X. k is the dimension of the key vector.

[0023] In one embodiment, the knowledge graph construction includes:

[0024] In the knowledge graph construction phase, the enterprise’s ESG knowledge graph will be constructed based on the feature vectors extracted in the previous steps;

[0025] First, define the structure of the knowledge graph. In the ESG scenario, the knowledge graph consists of nodes and edges. Nodes represent enterprises, events, policies, and indicator entities, while edges represent the relationships between these entities.

[0026] Next, these nodes and edges are constructed into a graph structure; given a graph G = (V, E), where V is the node set and E is the edge set; the node feature matrix Contains the feature vector of each node, where d is the dimension of the feature vector;

[0027] The graph neural network GNN is used to capture the relationship between graph structure information and node features by iteratively updating the representation of the node; the operation is expressed as:

[0028]

[0029] Among them, H (l) is the node representation of the lth layer, W (l) is the learnable weight matrix of this layer, σ is the activation function (ReLU), is the normalized adjacency matrix, usually defined as in I is the identity matrix, yes The degree matrix of .

[0031] In one embodiment, the model training includes:

[0032] In the model training phase, deep learning is used to train the constructed ESG knowledge graph to form a model that can provide ESG scoring;

[0033] First, a deep learning model is defined to process the feature representation extracted from the knowledge graph, and a graph neural network (GNN) is used to predict ESG scores using the feature and structure information of the nodes.

[0034] In order to further improve the adaptability of the model, reinforcement learning is introduced. In the reinforcement learning framework, the model is regarded as an intelligent agent whose task is to learn a strategy by interacting with the environment to maximize the cumulative reward; the intelligent agent takes actions by observing the state of the environment and obtains rewards based on the feedback from the environment; the strategy can be optimized by the policy gradient method, and its goal is to maximize the expected reward:

[0035]

[0036] Among them, τ represents the strategy π θ The generated trajectory, R(τ) is the cumulative reward of the trajectory. The policy gradient method optimizes J(θ) by adjusting the policy parameter θ, and its update rule is:

[0037] In the ESG scoring model, reinforcement learning is used to dynamically adjust model parameters to better adapt to different market conditions and corporate behaviors.

[0038] In one embodiment, the index calculation and calibration includes:

[0039] First, apply the trained model to a new data set. There is a new set of enterprise data X new , the data has been converted into feature representations suitable for model input through knowledge graph construction; use the trained model F to predict these data and obtain the ESG score of each company:

[0040] ESG score =F(X new )

[0041] These scores are normalized; the ESG index is calculated as:

[0042]

[0043] Next, the calculated ESG index is calibrated and feedback data is collected by actually tracking the ESG performance of the company. The feedback data is used to evaluate the prediction error of the model and guide the adjustment of the model;

[0044] Assume there is a set of actual observed ESG performance data Y actual , calculate the error between the model prediction and actual performance, using Mean Squared Error (MSE):

[0045]

[0046] By analyzing errors, systematic biases in model predictions can be identified and the model can be adjusted accordingly.

[0047] In one aspect, an enterprise ESG index determination system based on artificial intelligence, the system is applicable to the method described, and the system includes:

[0048] The data collection and processing module is used to collect announcements, news reports, public information, etc. of relevant enterprises and perform pre-processing operations;

[0049] Feature extraction module, used to convert preprocessed text data into feature vectors suitable for quantitative analysis by machine learning models;

[0050] The knowledge graph construction module is used to sort and predict based on the extracted features, and then build the knowledge graph of the enterprise ESG;

[0051] The model training module is used to use the constructed knowledge graph to train the machine learning model to form the corresponding ESG scoring model;

[0052] The index calculation and calibration module is used to calibrate the model parameters using supervised learning to make the model's prediction results more consistent with the actual ESG performance, and use the trained model to predict new data to calculate the company's ESG index;

[0053] The feedback module is used to find out the deficiencies of the model by comparing the model prediction results with the actual results, and to optimize and improve the model.

[0054] Beneficial effects of the present invention:

[0055] The AI-based method for determining corporate ESG index combines artificial intelligence technologies such as natural language processing, deep learning, and reinforcement learning. It can automatically extract and analyze ESG-related feature information from a large amount of unstructured information, and based on this, build an enterprise ESG knowledge graph, train models, and determine the enterprise ESG index.

[0056] This method can greatly improve the efficiency and accuracy of ESG assessment, freeing up expert resources from tedious data processing and focusing on tasks that require more human expertise. At the same time, due to its automated and intelligent features, this method can process large-scale data within an acceptable computational complexity, thereby better capturing and reflecting the ESG performance of enterprises.

[0057] In addition, this method also uses reinforcement learning technology, which enables the model to self-optimize through continuous trial and error, improves the adaptability and predictive ability of the model, and increases the objectivity and fairness of the evaluation.

[0058] Therefore, the method of the present invention has very important practical significance and far-reaching social impact for promoting the automation and intelligence of ESG evaluation, improving the ESG management level of enterprises, and promoting socially responsible investment and sustainable development. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of the method of the present invention;

[0060] Figure 2 This is a flow chart of the feature extraction stage of the present invention;

[0061] Figure 3 This is a flow chart of the knowledge graph construction phase of the present invention. DETAILED DESCRIPTION

[0062] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Typical embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0063] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used in the present invention in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Typical embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0064] Embodiment 1:

[0065] A method for determining an enterprise ESG index based on artificial intelligence, the implementation steps are as follows:

[0066] Environmental, Social and Governance (ESG) investment decision-making methods are gaining increasing attention.

[0067] S1. Data collection and processing: We collect public information including corporate announcements, news reports, social media, etc. through web crawler technology, and use natural language processing technology for pre-processing, including entity recognition, keyword extraction, etc.

[0068] The following is a detailed implementation of the data collection and processing phases:

[0069] Data collection, first of all, uses web crawlers to collect public information such as announcements, news reports, social media, etc. from the Internet. Depending on the data source, the implementation method of the crawler will also be different. For example, for some regular websites that often publish corporate announcements, we can periodically crawl newly released content; for social media, we need to use its developer API to obtain data.

[0070] The crawled data needs to be stored effectively, using a database or other data storage method. The raw data is stored in HTML, PDF, pictures, videos, etc., and needs to be cleaned and formatted initially, such as extracting text from HTML, OCR recognizing text from pictures, etc.

[0071] Preprocessing: The data preprocessing stage usually includes steps such as text cleaning and standardization. This includes (but is not limited to) removing garbled characters and non-text content, punctuation, auxiliary words and other stop words, HTML tags, URLs, and identifying the language of the text.

[0072] S2. Feature extraction: Introduce deep learning models, such as BERT, to convert preprocessed text data into feature vectors that can be used for quantitative analysis.

[0073] like Figure 2 As shown in Figure 1, in the feature extraction stage, the BERT model is used to convert the preprocessed text data into feature vectors. This process involves multiple steps and mathematical concepts to ensure that the text data can be represented in a form suitable for quantitative analysis.

[0074] S201. First, the preprocessed text data is input into the BERT model. BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained language model based on the Transformer architecture. It uses a bidirectional encoder to understand contextual information, thereby generating more accurate text representations.

[0075] S202, in the input stage, the text data is first segmented into word fragments (tokens). These word fragments are encoded by word embeddings, position embeddings, and segment embeddings. Word embeddings map each word fragment to a high-dimensional vector space, position embeddings provide position information for each word fragment, and segment embeddings are used to distinguish different sentences or paragraphs.

[0076] S203. Next, the input data processed by the embedding layer is processed by a multi-layer Transformer encoder. Each layer of the encoder consists of a multi-head self-attention mechanism and a feedforward neural network. The core of the self-attention mechanism is to capture global context information by calculating the correlation between each word fragment and other word fragments in the input sequence. Specifically, given the input matrix X, the self-attention calculation can be expressed as:

[0077]

[0078] Among them, Q, K, and V are query, key, and value matrices, respectively, which are usually obtained by linearly transforming the input matrix X. k is the dimension of the key vector.

[0079] After processing by multiple layers of encoders, BERT generates a contextual representation of each word fragment. In order to obtain the feature vector of the entire text, the output vector corresponding to the [CLS] tag is usually used, which is regarded as an aggregate representation of the entire input sequence.

[0080] Finally, the extracted feature vectors can be used for subsequent analysis and modeling steps. These vectors not only contain the semantic information of the text, but also capture rich contextual relationships through BERT's bidirectional encoder, providing a solid foundation for the calculation of the ESG index.

[0081] S3. Knowledge graph construction: Analyze and predict the extracted features to construct the enterprise’s ESG knowledge graph. This step can be done with the help of complex network analysis methods such as Graph Neural Networks.

[0082] like Figure 3 As shown in the figure, in the knowledge graph construction stage, based on the feature vectors extracted in the previous step, complex network analysis methods (such as graph neural network, GNN) are used to construct the enterprise's ESG knowledge graph. This process involves converting text features into graph structures and performing information propagation and node representation learning through graph neural networks.

[0083] S301. First, define the structure of the knowledge graph. In the ESG scenario, the knowledge graph is usually composed of nodes and edges. Nodes can represent entities such as companies, events, policies, indicators, etc., while edges represent the relationships between these entities, such as "participation", "influence" or "subordinate". The feature vectors extracted from S1 and S2 can help identify and define these nodes and their relationships.

[0084] S302. Next, we construct these nodes and edges into a graph structure. Given a graph G = (V, E), where V is the node set and E is the edge set. Node feature matrix Contains the feature vector of each node, where d is the dimension of the feature vector.

[0085] S303. In order to analyze and predict on the graph, a graph neural network (GNN) is used. GNN captures the relationship between graph structure information and node features by iteratively updating the representation of nodes. Graph convolutional network (GCN) is a common form of GNN, and its basic operation can be expressed as:

[0086]

[0087] Among them, H (l) is the node representation of the lth layer, W (l) is the learnable weight matrix of this layer, σ is the activation function (ReLU), is the normalized adjacency matrix, usually defined as in I is the identity matrix, yes The degree matrix of .

[0088] Through the propagation of multiple layers of GCN, the representation of a node not only contains its own features, but also incorporates the information of neighboring nodes. This information aggregation mechanism enables graph neural networks to effectively capture complex entity relationships and contextual information.

[0089] Finally, the node representations processed by GNN can be used for ESG-related analysis and prediction tasks. For example, we can predict the performance of a company on a specific ESG indicator, or identify key factors that have a significant impact on ESG scores. In this way, the company's ESG knowledge graph is constructed and improved, laying the foundation for subsequent model training and index calculation.

[0090] S4. Model training: Use deep learning models to train the knowledge graph to form an ESG scoring model. Consider introducing reinforcement learning algorithms to learn and optimize through continuous interaction with the environment.

[0091] In the model training phase, we will use deep learning and ensemble learning methods to train the constructed ESG knowledge graph to form a model that can provide ESG scores. In order to further improve the performance and adaptability of the model, reinforcement learning algorithms can be introduced to continuously optimize the model through interaction with the environment.

[0092] First, we need to define a deep learning model to process the feature representation extracted from the knowledge graph. A graph neural network (GNN) is used to predict ESG scores using the feature and structural information of the nodes.

[0093] In order to further improve the adaptability of the model, reinforcement learning (RL) can be introduced. In the reinforcement learning framework, the model is regarded as an agent whose task is to learn a policy by interacting with the environment to maximize the cumulative reward. The agent takes actions by observing the state of the environment and obtains rewards based on the feedback from the environment. The policy can be optimized by the policy gradient method, whose goal is to maximize the expected reward:

[0094]

[0095] Among them, τ represents the strategy π θ The generated trajectory, R(τ) is the cumulative reward of the trajectory. The policy gradient method optimizes J(θ) by adjusting the policy parameter θ, and its update rule is:

[0096] In ESG scoring models, reinforcement learning can be used to dynamically adjust model parameters to better adapt to different market conditions and corporate behaviors.

[0097] By combining deep learning and reinforcement learning, a powerful ESG scoring model can be built. This model can not only effectively use the information in the knowledge graph for scoring, but also continuously optimize and improve itself through interaction with the environment, thereby providing more accurate and reliable ESG scores.

[0098] S5. Index calculation and calibration: Apply the trained model to new data, calculate the company's ESG index, and use actual tracking performance to adjust and optimize the model.

[0099] In the index calculation and calibration phase, we will use the trained ESG scoring model to predict new data, calculate the company's ESG index, and adjust and optimize it according to the actual effect of the model. The following is the detailed implementation process of the index calculation and calibration phase:

[0100] Model application: First, we need to obtain new data, including the latest corporate announcements, news reports, social media and other public information. Since these data may contain content in different formats such as HTML, PDF, pictures, videos, etc., they need to be processed uniformly, such as extracting text from HTML, OCR recognizing text in pictures, etc., and extracting key features.

[0101] Then, we input the preprocessed data into the trained model to generate the ESG score. For example, if we are using a deep learning model, we convert the input data into a format that the model can process, and then get the score result through forward propagation; if reinforcement learning is used, we may need to select the best action based on the current environment state. This action can be a specific step in calculating the ESG score, such as selecting a model with a larger feature weight for calculation.

[0102] Index calculation: After the ESG score is calculated, it can be converted into an index form. The trained ESG score model is applied to new data to calculate the company's ESG index. We then adjust and optimize the model by actually tracking the company's performance to improve its accuracy and reliability.

[0103] First, we need to apply the trained model to a new dataset. Suppose we have a new set of enterprise data X new , these data have been converted into feature representations suitable for model input through the previous knowledge graph construction step. We use the trained model F to predict these data and obtain the ESG score of each company:

[0104] ESG score =F(X new )

[0105] These scores can be used directly as an ESG index for a company, or further normalized to facilitate comparison between different companies. The ESG index can be calculated as:

[0106]

[0107] Next, the calculated ESG index is calibrated. The purpose of calibration is to ensure the consistency between the model's predictions and actual corporate performance. To this end, feedback data can be collected by actually tracking the company's ESG performance. This feedback data can be used to evaluate the model's prediction error and guide the model's adjustment.

[0108] Assume there is a set of actual observed ESG performance data Y actual , we can calculate the error between the model prediction and the actual performance using the mean squared error (MSE):

[0109] By analyzing the errors, we can identify systematic biases in the model’s predictions and adjust the model accordingly. For example, we can adjust the model’s weight parameters or retrain the model to better fit the new data distribution.

[0110] Model adjustment and optimization: After obtaining the ESG index, we need to adjust and optimize it according to the actual performance of the model. First, we need to evaluate the model. We can choose MSE, MAE, R 2 Evaluation indicators such as , measure the performance of the model based on the difference between the predicted results and the actual values. If the model performs poorly, then the model parameters need to be optimized and adjusted, such as adjusting the loss function and optimizer of the deep learning model, or adjusting the weights of each base model in ensemble learning, or updating the policy parameters in reinforcement learning.

[0111] Embodiment 2

[0112] like Figure 2 As shown, based on the method of Embodiment 1, an enterprise ESG index determination system based on artificial intelligence is constructed. The data collection and processing module is used to collect announcements, news reports, public information, etc. of relevant enterprises and perform pre-processing operations;

[0113] Feature extraction module, used to convert preprocessed text data into feature vectors suitable for quantitative analysis by machine learning models;

[0114] The knowledge graph construction module is used to sort and predict based on the extracted features, and then build the knowledge graph of the enterprise ESG;

[0115] The model training module is used to use the constructed knowledge graph to train the machine learning model to form the corresponding ESG scoring model;

[0116] The index calculation and calibration module is used to calibrate the model parameters using supervised learning to make the model's prediction results more consistent with the actual ESG performance, and use the trained model to predict new data to calculate the company's ESG index;

[0117] The feedback module is used to find out the deficiencies of the model by comparing the model prediction results with the actual results, and to optimize and improve the model.

[0118] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0119] It should be understood that the detailed description of the technical solutions of the present invention by means of the preferred embodiments is illustrative rather than restrictive. A person skilled in the art may modify the technical solutions described in the embodiments, or replace some of the technical features by equivalents, based on reading the specification of the present invention; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The method for determining the enterprise ESG index based on artificial intelligence is characterized by: The method includes: Data collection and processing: Collect corporate announcements, news reports, and social media information, and use natural language processing technology for pre-processing, including entity recognition and keyword extraction; Feature extraction: preprocessed text data is converted into feature vectors for quantitative analysis; Knowledge graph construction: Analyze and predict the extracted features to construct the enterprise’s ESG knowledge graph; Model training: Train the knowledge graph to form an ESG scoring model; Index calculation and calibration: Apply the trained model to new data, calculate the company's ESG index, and use actual tracking performance to adjust and optimize the model.

2. The method for determining an enterprise ESG index based on artificial intelligence according to claim 1, characterized in that: The data collection and processing include: Use web crawlers to collect public information from announcements and news reports of relevant companies on the Internet; use developer APIs of social media to obtain data; The data preprocessing stage includes text cleaning and standardization steps, which include removing garbled characters and non-text content, punctuation, auxiliary words and other stop words, HTML tags, URLs, and identifying the language of the text.

3. The method for determining an enterprise ESG index based on artificial intelligence according to claim 1, characterized in that: The feature extraction adopts the BERT model, including: First, the preprocessed text data is input into the BERT model; the bidirectional encoder is used to understand the context information to generate a more accurate text representation; In the input stage, the text data is first segmented into word fragments, which are encoded by word embedding, position embedding, and segment embedding. Word embedding maps each word fragment to a high-dimensional vector space, position embedding provides position information for each word fragment, and segment embedding is used to distinguish different sentences or paragraphs. Next, the input data processed by the embedding layer is processed by a multi-layer Transformer encoder; each layer of the encoder consists of a multi-head self-attention mechanism and a feedforward neural network; the core of the self-attention mechanism is to capture global context information by calculating the correlation between each word fragment and other word fragments in the input sequence; specifically, given the input matrix X, the self-attention calculation is expressed as: Among them, Q, K, and V are query, key, and value matrices, respectively, which are usually obtained by linearly transforming the input matrix X. k is the dimension of the key vector.

4. The method for determining an enterprise ESG index based on artificial intelligence according to claim 1, characterized in that: The knowledge graph construction includes: In the knowledge graph construction phase, the enterprise’s ESG knowledge graph will be constructed based on the feature vectors extracted in the previous steps; First, define the structure of the knowledge graph. In the ESG scenario, the knowledge graph consists of nodes and edges. Nodes represent enterprises, events, policies, and indicator entities, while edges represent the relationships between these entities. Next, these nodes and edges are constructed into a graph structure; given a graph G = (V, E), where V is the node set and E is the edge set; the node feature matrix Contains the feature vector of each node, where d is the dimension of the feature vector; The graph neural network GNN is used to capture the relationship between graph structure information and node features by iteratively updating the representation of the node; the operation is expressed as: Among them, H (l) is the node representation of the lth layer, W (l) is the learnable weight matrix of this layer, σ is the activation function (ReLU), is the normalized adjacency matrix, usually defined as in I is the identity matrix, yes The degree matrix of .

5. The method for determining an enterprise ESG index based on artificial intelligence according to claim 1, characterized in that: The model training includes: In the model training phase, deep learning is used to train the constructed ESG knowledge graph to form a model that can provide ESG scoring; First, a deep learning model is defined to process the feature representation extracted from the knowledge graph, and a graph neural network (GNN) is used to predict ESG scores using the feature and structure information of the nodes. In order to further improve the adaptability of the model, reinforcement learning is introduced. In the reinforcement learning framework, the model is regarded as an intelligent agent whose task is to learn a strategy by interacting with the environment to maximize the cumulative reward; the intelligent agent takes actions by observing the state of the environment and obtains rewards based on the feedback from the environment; the strategy can be optimized by the policy gradient method, and its goal is to maximize the expected reward: Among them, τ represents the strategy π θ The generated trajectory, R(τ) is the cumulative reward of the trajectory; the policy gradient method optimizes J(θ) by adjusting the policy parameter θ, and its update rule is: In the ESG scoring model, reinforcement learning is used to dynamically adjust model parameters to better adapt to different market conditions and corporate behaviors.

6. The method for determining an enterprise ESG index based on artificial intelligence according to claim 1, characterized in that: The index calculation and calibration include: First, apply the trained model to a new data set. There is a new set of enterprise data X new , the data has been converted into feature representations suitable for model input through knowledge graph construction; use the trained model F to predict these data and obtain the ESG score of each company: ESG score =F(X new ) These scores are normalized; the ESG index is calculated as: Next, the calculated ESG index is calibrated; feedback data is collected by actually tracking the ESG performance of companies; feedback data is used to evaluate the model’s prediction error and guide the model’s adjustment; Assume there is a set of actual observed ESG performance data Y actual , calculate the error between the model prediction and actual performance, using Mean Squared Error (MSE): By analyzing errors, systematic biases in model predictions can be identified and the model can be adjusted accordingly.

7. An enterprise ESG index determination system based on artificial intelligence, wherein the system is applicable to the method according to any one of claims 1 to 6, characterized in that: The system comprises: The data collection and processing module is used to collect announcements, news reports, public information, etc. of relevant enterprises and perform pre-processing operations; Feature extraction module, used to convert preprocessed text data into feature vectors suitable for quantitative analysis by machine learning models; The knowledge graph construction module is used to sort and predict based on the extracted features, and then build the knowledge graph of the enterprise ESG; The model training module is used to use the constructed knowledge graph to train the machine learning model to form the corresponding ESG scoring model; The index calculation and calibration module is used to calibrate the model parameters using supervised learning to make the model's prediction results more consistent with the actual ESG performance, and use the trained model to predict new data to calculate the company's ESG index; The feedback module is used to find out the deficiencies of the model by comparing the model prediction results with the actual results, and to optimize and improve the model.

Images