ESG report multi-task classification model construction method and device, equipment and medium

Abstract

The application relates to an ESG report multi-task classification model construction method and device, equipment and a medium, wherein the method comprises the following steps: acquiring an initial multi-task classification model and sample ESG report texts, performing feature extraction, and obtaining a global context representation; generating first task features for an ESG theme classification task and second task features for a quality evaluation task; using a one-way gate mechanism, generating a one-way gate signal based on the first task features, modulating and updating a theme original prediction result by using a quality classification prediction result obtained by independent prediction of the second task features, dynamically weighting a prediction entropy of the quality classification prediction result according to a prediction confidence of the theme original prediction result, and constructing a dynamic consistency regularization loss; and jointly optimizing the classification labels corresponding to the sample ESG report texts, the dynamic consistency regularization loss and a multi-task classification loss to obtain a trained ESG report multi-task classification model.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to a method, apparatus, computer device, storage medium, and computer program product for constructing a multi-task classification model for ESG reports. Background Technology

[0002] With the global popularization of sustainable development concepts, Environmental, Social, and Governance (ESG) reports have become an important textual carrier reflecting a company's non-financial performance, long-term value, and risk management capabilities. Automated analysis of ESG reports typically involves two interrelated tasks: first, ESG topic classification, which identifies the specific environmental, social, or governance issues addressed in the sentences; and second, disclosure quality assessment, which determines the specificity, factuality, or sufficiency of the information contained in the sentences. Since the same ESG sentence often simultaneously reflects both issue attributes and disclosure quality attributes, how to achieve collaborative modeling of these two types of information within a unified model is a crucial issue in intelligent analysis of ESG reports.

[0003] Existing methods typically employ a multi-task learning framework based on a shared text encoder, achieving knowledge transfer between different tasks by sharing underlying semantic representations. However, such methods still have shortcomings in practical applications. First, topic classification tasks focus more on ESG issue boundaries, long-tail categories, and fine-grained semantic differences, while disclosure quality assessment tasks focus more on whether sentences contain factual, quantifiable, or verifiable information. The two types of tasks do not have entirely consistent focuses on feature representation. If modeling is based solely on a shared encoder or simple task branches, it can easily lead to insufficient representation of task-specific information.

[0004] Secondly, existing multi-task classification methods typically optimize topic classification results and quality assessment results as independent outputs, failing to fully utilize the potential semantic dependencies between the two types of labels. For example, disclosure sentences under different ESG topics may differ in factuality, quantification level, or expression style, and quality assessment results may also provide auxiliary discriminative information for topic classification. Without an effective cross-task information transfer mechanism, the model struggles to leverage such dependencies to improve classification performance.

[0005] Furthermore, some methods model inter-task relationships through feature concatenation, shared fusion layers, or additional interaction modules. However, these approaches often lack constraints on the flow of cross-task information and the intensity of feedback, easily introducing noise information unrelated to the target task and affecting the stable collaboration between the topic classification branch and the quality assessment branch. Especially in ESG sentence-level classification scenarios, where topic categories are highly granular and their distribution is uneven, a lack of selective control over cross-task interactions may further amplify inter-task interference.

[0006] In summary, traditional technical solutions struggle to accurately coordinate the two tasks of ESG topic classification and disclosure quality assessment. Therefore, a multi-task classification method is needed that can selectively transfer and dynamically constrain the correlation information between the ESG topic classification and disclosure quality assessment tasks while preserving task-specific representations, in order to improve the accuracy and training stability of ESG report text analysis. Summary of the Invention

[0007] Therefore, it is necessary to provide a method, apparatus, computer device, computer-readable storage medium, and computer program product for constructing an ESG report multi-task classification model that supports accurate ESG report multi-task processing, in response to the above-mentioned technical problems.

[0008] Firstly, this application provides a method for constructing a multi-task classification model for ESG reports. The method includes: The initial multi-task classification model and sample ESG report texts are obtained. The shared text encoder in the initial multi-task classification model is used to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The global context representation is input into the first task adapter and the second task adapter respectively to generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task. Using a one-way gating mechanism, a one-way gating signal is generated based on the first task feature. The quality classification prediction result obtained by independent prediction of the second task feature is modulated and updated using the one-way gating signal and the quality classification prediction result to obtain the original topic prediction result obtained by independent prediction of the first task feature, thus obtaining the final topic classification prediction result. The prediction entropy of the quality classification prediction results is dynamically weighted based on the prediction confidence of the original topic prediction results to construct a dynamic consistency regularization loss. Based on the classification labels corresponding to the sample ESG report text, and the dynamic consistency regularization loss and multi-task classification loss are jointly optimized, the parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0009] In one embodiment, the global context representation is input to a first task adapter and a second task adapter, respectively, to generate a first task feature for an ESG topic classification task and a second task feature for a quality assessment task, including: The global context representation is input into the first task adapter, and then subjected to layer normalization, dimensionality reduction, nonlinear activation, and random deactivation in sequence, followed by dimensionality increase to obtain the first task-specific offset. The first task-specific offset is then added element-wise to the global context representation to obtain the first task feature for the ESG topic classification task. The global context representation is input into the second task adapter, and then subjected to layer normalization, dimensionality reduction, nonlinear activation, and random deactivation in sequence, followed by dimensionality increase to obtain the second task-specific offset. The second task-specific offset is then added element-wise to the global context representation to obtain the second task features for the quality assessment task. The first task adapter and the second task adapter adopt the same bottleneck structure design with independent parameters; the weight matrix and bias vector of the dimension-up process are initialized to zero matrix and zero vector, so that the first task specialization offset and the second task specialization offset are both zero in the initial stage of training, and the first task feature and the second task feature are both equal to the global context representation in the initial stage of training; the dimension-down process reduces the dimension of the global context representation to the bottleneck dimension d.

[0010] In one embodiment, the initial multi-task classification model further includes a topic classification head and a quality classification head; Using a one-way gating mechanism, a one-way gating signal is generated based on the first task features. The quality classification prediction result obtained by independently predicting the second task features is then used to modulate and update the original topic prediction result obtained by independently predicting the first task features, along with the one-way gating signal and the quality classification prediction result. The final topic classification prediction result includes: The first task features are input into the topic classification head to generate the original topic prediction results, and a one-way gating signal is generated based on the prediction probability of the original topic prediction results. The second task features are input into the quality classification head to generate quality classification prediction results, and gradient truncation is performed on the quality classification prediction results to cut off the reverse gradient path of the quality classification prediction results. The first task features are concatenated with the quality classification prediction results after gradient truncation, and the concatenated features are used to generate a correction vector through a correction network. The one-way gating signal is multiplied by the correction vector, and the product is injected into the original topic prediction result by residual addition to obtain the final topic classification prediction result.

[0011] In one embodiment, the prediction entropy of the quality classification prediction result is dynamically weighted based on the prediction confidence of the original topic prediction result, and the dynamic consistency regularization loss is constructed by including: Calculate the prediction confidence of the original prediction results for the topic; Calculate the information entropy of the predicted distribution of the quality classification prediction results; The information entropy is weighted using the prediction confidence as a coefficient, and a weighted entropy loss is constructed as the dynamic consistency regularization loss.

[0012] In one embodiment, calculating the prediction confidence of the original prediction result for the topic includes: Obtain the predicted probability distribution of the original prediction results for the topic; The highest probability in the predicted probability distribution is determined as the prediction confidence level corresponding to the original prediction result of the topic.

[0013] In one embodiment, based on the classification labels corresponding to the sample ESG report texts, and through joint optimization of dynamic consistency regularization loss and multi-task classification loss, the parameters of the initial multi-task classification model are updated via end-to-end training, resulting in a trained ESG report multi-task classification model including: Obtain the topic classification labels and quality classification labels corresponding to the sample ESG report text; Calculate the topic classification loss based on the final topic classification prediction results and topic classification labels; Calculate the quality classification loss based on the quality classification prediction results and quality classification labels; We construct a joint objective function by weighted summation of topic classification loss, quality classification loss, and dynamic consistency regularization loss. Backpropagation is performed based on the joint objective function to update the parameters of the initial multi-task classification model until the model converges, resulting in the trained ESG report multi-task classification model.

[0014] In one embodiment, after jointly optimizing the classification labels corresponding to the sample ESG report texts, and the dynamic consistency regularization loss and multi-task classification loss, and updating the parameters of the initial multi-task classification model through end-to-end training to obtain the trained ESG report multi-task classification model, the method further includes: Obtain the text of the ESG report to be classified; The text of the ESG report to be classified is input into the trained ESG report multi-task classification model to obtain ESG topic classification results and disclosure quality assessment results.

[0015] Secondly, this application also provides an apparatus for constructing a multi-task classification model for ESG reports. The apparatus includes: The initialization module is used to obtain the initial multi-task classification model and sample ESG report texts. It uses the shared text encoder in the initial multi-task classification model to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The feature extraction module is used to input the global context representation into the first task adapter and the second task adapter respectively, and generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task. The classification prediction module is used to generate a one-way gating signal based on the first task feature using a one-way gating mechanism. It then modulates and updates the original topic prediction result obtained by independently predicting the first task feature using the one-way gating signal and the quality classification prediction result to obtain the final topic classification prediction result. The loss construction module is used to dynamically weight the prediction entropy of the quality classification prediction results based on the prediction confidence of the original topic prediction results, and construct a dynamic consistency regularization loss. The model training module is used to jointly optimize the classification labels corresponding to the sample ESG report text, as well as the dynamic consistency regularization loss and multi-task classification loss. The parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0016] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps: The initial multi-task classification model and sample ESG report texts are obtained. The shared text encoder in the initial multi-task classification model is used to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The global context representation is input into the first task adapter and the second task adapter respectively to generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task. Using a one-way gating mechanism, a one-way gating signal is generated based on the first task feature. The quality classification prediction result obtained by independent prediction of the second task feature is modulated and updated using the one-way gating signal and the quality classification prediction result to obtain the original topic prediction result obtained by independent prediction of the first task feature, thus obtaining the final topic classification prediction result. The prediction entropy of the quality classification prediction results is dynamically weighted based on the prediction confidence of the original topic prediction results to construct a dynamic consistency regularization loss. Based on the classification labels corresponding to the sample ESG report text, and the dynamic consistency regularization loss and multi-task classification loss are jointly optimized, the parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0017] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps: The initial multi-task classification model and sample ESG report texts are obtained. The shared text encoder in the initial multi-task classification model is used to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The global context representation is input into the first task adapter and the second task adapter respectively to generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task. Using a one-way gating mechanism, a one-way gating signal is generated based on the first task feature. The quality classification prediction result obtained by independent prediction of the second task feature is modulated and updated using the one-way gating signal and the quality classification prediction result to obtain the original topic prediction result obtained by independent prediction of the first task feature, thus obtaining the final topic classification prediction result. The prediction entropy of the quality classification prediction results is dynamically weighted based on the prediction confidence of the original topic prediction results to construct a dynamic consistency regularization loss. Based on the classification labels corresponding to the sample ESG report text, and the dynamic consistency regularization loss and multi-task classification loss are jointly optimized, the parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0018] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps: The initial multi-task classification model and sample ESG report texts are obtained. The shared text encoder in the initial multi-task classification model is used to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The global context representation is input into the first task adapter and the second task adapter respectively to generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task. Using a one-way gating mechanism, a one-way gating signal is generated based on the first task feature. The quality classification prediction result obtained by independent prediction of the second task feature is modulated and updated using the one-way gating signal and the quality classification prediction result to obtain the original topic prediction result obtained by independent prediction of the first task feature, thus obtaining the final topic classification prediction result. The prediction entropy of the quality classification prediction results is dynamically weighted based on the prediction confidence of the original topic prediction results to construct a dynamic consistency regularization loss. Based on the classification labels corresponding to the sample ESG report text, and the dynamic consistency regularization loss and multi-task classification loss are jointly optimized, the parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0019] The aforementioned ESG report multi-task classification model construction method, apparatus, computer equipment, storage medium, and computer program product, through the collaborative work of a shared text encoder and parallel first and second task adapters, effectively alleviates the representation pollution problem between heterogeneous tasks while preserving common semantic knowledge between tasks. This allows the topic classification task and the quality assessment task to each obtain more discriminative task-specific features. Utilizing a one-way gating mechanism, a one-way gating signal is generated based on the first task features, and the quality classification prediction result obtained by independently predicting the second task features is used as auxiliary information for the first task... The original topic predictions obtained through independent prediction of task features are modulated and updated to explicitly model the asymmetric semantic dependency between the topic classification task and the quality assessment task. This overcomes the shortcomings of existing multi-task learning methods that neglect the correlation of task outputs, significantly improving the accuracy of topic classification. Furthermore, a dynamic consistency regularization loss is constructed by dynamically weighting the prediction entropy of the quality classification prediction result with the prediction confidence of the original topic prediction result. This strengthens the logical dynamic consistency between the two task predictions at the loss function level, making the quality assessment result more certain when the topic prediction is clear, thereby improving the stability and rationality of joint prediction. Finally, end-to-end joint optimization training is performed based on classification labels, dynamic consistency regularization loss, and multi-task classification loss to obtain a multi-task classification model that can accurately perform ESG report topic classification and disclosure quality assessment. Attached Figure Description

[0020] Figure 1This is a diagram illustrating the application environment of an ESG report multi-task classification model construction method in one embodiment. Figure 2 This is a flowchart illustrating a method for constructing a multi-task classification model for ESG reports in one embodiment. Figure 3 This is a flowchart illustrating the method for constructing a multi-task classification model for ESG reports in another embodiment; Figure 4 This is a structural block diagram of an ESG report multi-task classification model construction device in one embodiment; Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] The ESG report multi-task classification model construction method provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. Terminal 102 sends a model building request to server 104. Server 104 obtains an initial multi-task classification model and sample ESG report text. It uses the shared text encoder in the initial multi-task classification model to extract features from the sample ESG report text to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The global context representation is input to the first and second task adapters respectively to generate first task features for the ESG topic classification task and second task features for the quality assessment task. Using a one-way gating mechanism, a one-way gating signal is generated based on the first task features, and the quality classification prediction result obtained by independently predicting the second task features is used. The system modulates and updates the original topic prediction results obtained by independently predicting the first task features using a one-way gating signal and the quality classification prediction results, to obtain the final topic classification prediction results. The prediction entropy of the quality classification prediction results is dynamically weighted according to the prediction confidence of the original topic prediction results to construct a dynamic consistency regularization loss. Based on the classification labels corresponding to the sample ESG report texts, and the dynamic consistency regularization loss and multi-task classification loss, joint optimization is performed. The parameters of the initial multi-task classification model are updated through end-to-end training to obtain a trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction results and the quality classification loss calculated based on the quality classification prediction results. Furthermore, the server 104 can also return the constructed trained ESG report multi-task classification model to the terminal 102. The terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers. Furthermore, this application can also be directly applied to terminals, and the specific processing procedure is similar to that described above, so it will not be repeated here.

[0023] In one embodiment, such as Figure 2 As shown, a method for constructing a multi-task classification model for ESG reports is provided, which can be applied to... Figure 1 Taking server 104 as an example, the following steps are included: S100: Obtain the initial multi-task classification model and sample ESG report texts. Use the shared text encoder in the initial multi-task classification model to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. Additionally, the entire initial multi-task classification model may also include a unidirectional gating unit and a dynamic consistency regularization unit.

[0024] A shared text encoder is a text feature extraction network capable of sharing underlying parameters across multiple tasks. Specifically, it encodes common semantic information required by different tasks into the same representation space, enabling the transfer and sharing of underlying knowledge and avoiding parameter redundancy and knowledge isolation when each task is modeled independently. The first task adapter is a network module used to extract topic classification task-specific information from the shared representation; the second task adapter is a network module used to extract quality assessment task-specific information from the shared representation. This parallel design allows both tasks to share underlying knowledge while retaining their ability to model task-specific features, effectively mitigating the representation pollution problem between heterogeneous tasks.

[0025] In practical applications, the shared text encoder can be implemented using a pre-trained language model. For example, bert-base-chinese can be used as the shared encoder, and its pooled_output can be used as the global context representation. This representation can be a 768-dimensional vector containing the aggregate semantic information of the input text. The first and second task adapters can be implemented using lightweight networks with bottleneck structures, generating their respective task-specific offsets with minimal parameter overhead through dimensionality reduction and dimensionality expansion.

[0026] S200: Input the global context representation into the first task adapter and the second task adapter respectively to generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task.

[0027] The first task feature refers to the feature representation with stronger discriminative power for ESG topic classification tasks after processing by the first task adapter; the second task feature refers to the feature representation with stronger discriminative power for disclosure quality assessment tasks after processing by the second task adapter. Here, two independent adapters differentiate the same global context representation, amplifying semantic signals relevant to their respective tasks and suppressing task-irrelevant noise information, thereby achieving the transformation from a general representation to a task-specific representation. In one embodiment, this feature generation process can achieve a smooth transition through residual connections. Specifically, the adapter output can be obtained by element-wise addition of the dimensionality-upgrading result and the original input. The weight matrix and bias vector of the dimensionality-upgrading process can adopt a near-zero initialization strategy, i.e., initialized as a zero matrix and a zero vector, so that the adapter behaves as an identity mapping in the initial training stage, and the first task feature and the second task feature are initially equal to the global context representation. As training progresses, the adapters (including the first task adapter and the second task adapter) gradually learn the specialized offsets of their respective tasks, thus achieving a smooth transition from general semantics to task-specific semantics.

[0028] S300: Using a one-way gating mechanism, a one-way gating signal is generated based on the first task feature. The quality classification prediction result obtained by independently predicting the second task feature is used to modulate and update the original topic prediction result obtained by independently predicting the first task feature using the one-way gating signal and the quality classification prediction result, so as to obtain the final topic classification prediction result.

[0029] A one-way gating signal refers to the dynamic weights generated from the prediction results of the first task features, used to control the intensity of correction. The original topic prediction result refers to the initial topic classification result obtained independently from the first task features, before being corrected by the gating mechanism. The final topic classification prediction result refers to the topic classification result after modulation and updating by the one-way gating signal. Here, modulation and updating can be understood as: using the quality classification prediction result as auxiliary information, and under the dynamic control of the gating signal, selectively correcting the original topic prediction result. Specifically, this explicitly models the one-way semantic dependency between the topic classification task and the quality assessment task. In the ESG report analysis scenario, there is an objective statistical correlation between the disclosure topic to which a sentence belongs and its disclosure quality—for example, topics related to environmental emissions are more likely to be accompanied by quantitative data disclosure—and known topics can provide prior evidence for quality judgment, while the reverse does not have this inference relationship. Through the one-way gating mechanism, topic prediction can benefit from the auxiliary information of the quality branch, while ensuring that the modulation process does not adversely affect the features used to generate the quality classification prediction result, maintaining the independent learning of the two tasks at the optimization level.

[0030] S400: Dynamically weight the prediction entropy of the quality classification prediction results based on the prediction confidence of the original topic prediction results to construct a dynamic consistency regularization loss.

[0031] Prediction confidence refers to the certainty measure of a model's prediction of the original topic, typically represented by the maximum probability value in the prediction probability distribution. Prediction entropy refers to the information entropy of the quality classification prediction distribution, used to measure the uncertainty of the model's prediction of the quality assessment result. Dynamic consistency regularization loss is a loss term constructed by weighting the prediction entropy with prediction confidence as the weight. Here, at the loss function level, logical dynamic consistency between the two task predictions is enforced. When the model is highly certain about the topic classification, it should be encouraged to make equally certain predictions about the quality assessment; when the topic classification is uncertain, the certainty constraint on the quality assessment should be relaxed accordingly. This dynamic weighting mechanism avoids the excessive constraints that may be introduced by traditional fixed-weight regularization, allowing the regularization strength to adaptively match the model's own prediction state.

[0032] S500: Based on the classification labels corresponding to the sample ESG report text, and the dynamic consistency regularization loss and multi-task classification loss are jointly optimized, the parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0033] Classification labels refer to the pre-labeled ESG topic category labels and disclosure quality category labels in the training samples. Topic classification loss is a measure of the difference between the final topic classification prediction and the topic classification label; quality classification loss is a measure of the difference between the quality classification prediction and the quality classification label. End-to-end training refers to using dynamic consistency regularization loss and multi-task classification loss as joint optimization objectives, updating all learnable parameters in the model all at once through backpropagation, rather than training in stages or independently for each task. Here, through joint optimization of multiple losses, the model pursues classification accuracy for each task while satisfying the logical dynamic consistency constraint between the predictions of the two tasks, ultimately obtaining a multi-task classification model that can accurately perform both topic classification and quality assessment simultaneously. Specifically, the objective function of joint optimization can be constructed as: Total Loss = Topic Classification Loss + Quality Classification Loss + α·Dynamic Consistency Regularization Loss, where α is an adjustable hyperparameter controlling the strength of regularization. The classification loss can be implemented using the cross-entropy loss function, and label smoothing techniques can be introduced during calculation to enhance the model's generalization performance. During training, a hierarchical learning rate scheduling strategy can be adopted, using a smaller learning rate for the shared text encoder to preserve pre-trained knowledge, and a larger learning rate for the adapter and classification head to accelerate task-specific learning.

[0034] The aforementioned method for constructing a multi-task classification model for ESG reports, through the collaborative work of a shared text encoder and parallel first and second task adapters, effectively alleviates the representation pollution problem between heterogeneous tasks while preserving common semantic knowledge between tasks. This allows the topic classification task and the quality assessment task to each acquire more discriminative task-specific features. Utilizing a one-way gating mechanism, a one-way gating signal is generated based on the first task features, and the quality classification prediction result obtained by independently predicting the second task features is used as auxiliary information to modulate and update the original topic prediction result obtained by independently predicting the first task features. This explicitly models the asymmetric semantic dependency between the topic classification task and the quality assessment task, making up for the shortcomings of existing multi-task learning methods that neglect task output correlation, and significantly improving the accuracy of topic classification. Furthermore, by dynamically weighting the prediction entropy of the quality classification prediction result with the prediction confidence of the original topic prediction result to construct a dynamic consistency regularized loss, the logical dynamic consistency of the two task predictions is strengthened at the loss function level, making the quality assessment result more certain when the topic prediction is clear, thereby improving the stability and rationality of joint prediction. Finally, end-to-end joint optimization training was performed based on classification labels, dynamic consistency regularization loss, and multi-task classification loss to obtain a multi-task classification model that can accurately perform ESG report topic classification and disclosure quality assessment.

[0035] In one embodiment, the global context representation is input to a first task adapter and a second task adapter, respectively, to generate a first task feature for an ESG topic classification task and a second task feature for a quality assessment task, including: Step 1: Input the global context representation into the first task adapter, and perform layer normalization, dimensionality reduction, nonlinear activation, and random deactivation in sequence, followed by dimensionality increase to obtain the first task-specific offset; add the first task-specific offset to the global context representation element by element to obtain the first task feature for the ESG topic classification task.

[0036] Step 2: Input the global context representation into the second task adapter, and perform layer normalization, dimensionality reduction, nonlinear activation, and random deactivation in sequence, followed by dimensionality increase to obtain the second task-specific offset; add the second task-specific offset to the global context representation element by element to obtain the second task feature for the quality assessment task.

[0037] The first task adapter and the second task adapter adopt the same bottleneck structure design with independent parameters; the weight matrix and bias vector of the dimension-up process are initialized to zero matrix and zero vector, so that the first task specialization offset and the second task specialization offset are both zero in the initial stage of training, and the first task feature and the second task feature are both equal to the global context representation in the initial stage of training; the dimension-down process reduces the dimension of the global context representation to the bottleneck dimension d.

[0038] Layer normalization refers to the process of normalizing input data to achieve a mean of zero and a variance of one. Its purpose is to stabilize data distribution and accelerate model convergence. Dimensionality reduction refers to the process of mapping input data from a high-dimensional space to a low-dimensional space, reducing the number of parameters and computational overhead by decreasing the feature dimension. Nonlinear activation refers to the process of using Gaussian error linear units as activation functions, introducing nonlinear transformation capabilities to enable the network to fit complex mapping relationships. Random deactivation refers to the process of randomly discarding some neurons with a certain probability during training, used to enhance the model's generalization performance and prevent overfitting. Dimensionality upscaling refers to the process of restoring the dimensionality-reduced low-dimensional representation to the original high-dimensional space to match the input dimensionality requirements of the downstream classifier. These processes are executed sequentially, constituting the complete forward computation process of the adapter.

[0039] The first task-specific offset refers to the topic classification task-specific incremental information learned by the first task adapter and superimposed on the global context representation. The second task-specific offset refers to the quality assessment task-specific incremental information learned by the second task adapter and superimposed on the global context representation. Adding the task-specific offsets element-wise to the global context representation signifies superimposing task-specific information on the general semantic representation using residual connections, forming a composite representation that integrates general semantics and task-specific discriminative features. This design allows the adapter to progressively learn the task-specific features required for each task while retaining the existing knowledge of the pre-trained model, rather than learning entirely new representations from scratch.

[0040] The bottleneck structure refers to the internal processing structure of the adapter that first reduces dimensionality and then increases it. By limiting the flow of information through a low-dimensional bottleneck layer, the adapter is forced to learn the most core task-specific information, while achieving multi-task adaptation with relatively small parameter overhead. The first and second task adapters adopt the same internal processing flow and hierarchical organization, and each has its own independent set of learnable parameters, which are not shared during training. This parameter independence design ensures that the specific feature extraction processes of the two tasks are isolated from each other, avoiding interference between training signals from different tasks, thereby effectively mitigating the problem of inter-task representation pollution in the shared representation space. In this embodiment, the weight matrix and bias vector of the dimensionality increase processing are initialized to zero matrices and zero vectors; this initialization strategy can be called near-zero initialization. The purpose of this setting is that, in the initial stage of training, the output of the dimensionality increase processing is zero, making the first task-specific offset and the second task-specific offset both zero, and the first task features and the second task features equal to the global context representation. This means that the adapter behaves as an identity mapping in the early stages of training, without changing the representation output of the pre-trained model, thus protecting the knowledge already learned by the pre-trained model from being destroyed by stochastic gradients in the early stages of training. As the training process progresses, the weight matrix and bias vector of the dimensionality-upgraded processing are gradually updated to non-zero values ​​through backpropagation. The adapter gradually learns the specialized offsets for each task, achieving a smooth transition from a general representation to a task-specific representation.

[0041] The bottleneck dimension *d* refers to the dimensionality of the feature vector after dimensionality reduction, and is a key control parameter for information flow within the adapter. This bottleneck dimension *d* can be configured according to the capacity-complexity matching principle, that is, proportionally set according to the number of task categories and the difficulty of discrimination. Specifically, when the adapter completes the generation of task-specific representations, its output features are the superposition of the original input and the adapter's learned bias. The parameters predicted by the network model are: shared pooling features... Dimensionality reduction matrix Upgraded Dimensional Matrix Residual connection parameters ; and non-linear activation functions. The specific functions corresponding to the adapter are as follows:

[0042] In one embodiment, the initial multi-task classification model further includes a topic classification head and a quality classification head; such as Figure 3 As shown, using a one-way gating mechanism, a one-way gating signal is generated based on the first task features. The quality classification prediction result obtained by independently predicting the second task features is then used to modulate and update the original topic prediction result obtained by independently predicting the first task features, using the one-way gating signal and the quality classification prediction result. The final topic classification prediction result includes: S320: Input the first task features into the topic classification head, generate the original topic prediction results, and generate a one-way gating signal based on the prediction probability of the original topic prediction results.

[0043] The topic classification head refers to the network module used to map the features of the first task to topic category prediction scores, and its input is connected to the output of the first task adapter. The one-way gating signal is a dynamic weight signal generated from the prediction probability of the original prediction result. Its magnitude reflects the model's certainty about the current topic prediction result and is used to control the activation intensity of subsequent correction operations. Here, firstly, the original topic prediction result is generated as the basis for subsequent correction operations; secondly, a one-way gating signal is generated based on the uncertainty of the original prediction, serving as a control switch for subsequent correction operations. The inherent logic of this design is: when the model is highly certain about the topic classification, the gating signal approaches zero, the subsequent correction mechanism is almost closed, and the topic prediction remains unchanged; when the model is uncertain about the topic classification, the gating signal increases, the subsequent correction mechanism is activated, and auxiliary information from the quality branch is introduced for adjustment.

[0044] Specifically, when the topic task completes its initial prediction, its confidence and uncertainty are calculated from the prediction probability distribution. The parameters predicted by the network model are the output of the topic task adapter. Theme category header Gating threshold parameters Temperature coefficient parameter And the gated activation function Sigmoid.

[0045] ; .

[0046] This design ensures that the adjudication module is activated only for samples with uncertain predictions, while for samples with highly certain topic predictions, the gating weights approach zero, the adjudication module skips directly, and the topic prediction remains unchanged.

[0047] S340: Input the second task features into the quality classification head, generate the quality classification prediction result, and perform gradient truncation processing on the quality classification prediction result to cut off the reverse gradient path of the quality classification prediction result.

[0048] The quality classification head is a network module used to map the features of the second task to predicted quality class scores. Gradient truncation refers to the operation of cutting off the back propagation path of a specified tensor in the computation graph. After this processing, the optimization gradients generated in subsequent calculations will not be propagated back to their upstream network modules through this tensor. Specifically, when the quality task completes independent forward computation, its specialized representation and probability distribution are generated by the adapter and the classification head. The parameters predicted by the network model are shared pooled features. Quality Task Adapter Quality task classification head And the probability distribution function Softmax. Its corresponding functional expression is as follows: .

[0049] After obtaining the probability distribution of the quality prediction, a gradient truncation operation is immediately performed to cut off the backward gradient path of the tensor. This design ensures that the optimization gradient of the subsequent adjudication module does not backpropagate to the quality classification branch, strictly guaranteeing the independent learning of the two tasks at the optimization level.

[0050] S360: The first task features are concatenated with the quality classification prediction results after gradient truncation, and the concatenated features are used to generate a correction vector through a correction network.

[0051] Feature concatenation refers to the operation of connecting two or more feature vectors along a specified dimension. The concatenated joint feature simultaneously carries feature information from the topic task and prediction information from the quality task. The correction network is a network module used to map the concatenated joint feature into a correction vector. Here, the auxiliary information contained in the quality classification prediction result is fused with the topic task features, and an incremental vector for correcting topic prediction is generated based on the fused comprehensive information. Through feature concatenation, the correction network can simultaneously perceive two aspects of information: first, the feature representation of the topic task, reflecting the semantic topic tendency of the sentence; and second, the prediction result of the quality classification, reflecting the judgment tendency of the sentence regarding disclosure adequacy. The correction network utilizes the interaction of these two aspects of information to generate more targeted correction quantities. For example, in the ESG report analysis scenario, if the quality classification prediction result indicates that the sentence tends towards quantitative disclosure, while the topic features indicate that the sentence involves environmental emissions issues, the correction network can use this comprehensive information to generate a corresponding correction vector, specifically adjusting the category probability distribution of the original topic prediction, thereby improving the accuracy of topic classification.

[0052] S380: Multiply the one-way gating signal with the correction vector, and inject the product into the original topic prediction result by residual addition to obtain the final topic classification prediction result.

[0053] Residual additive injection refers to the operation of adding the correction amount to the original prediction result in an additive manner. Its calculation form can be expressed as: Final prediction result = Original prediction result + Gating signal × Correction vector. Here, the correction of the original prediction result of the topic is completed in a controllable and smooth manner. The one-way gating signal plays a key role in dynamically adjusting the correction intensity: when the gating signal approaches zero, the product of the correction vector and the gating signal approaches zero, and the final classification prediction result of the topic is approximately equal to the original prediction result of the topic, and the model maintains its original judgment unchanged; when the gating signal is large, the correction vector is amplified accordingly, and the original prediction result of the topic is significantly corrected. This gated weighted residual injection design achieves "selective correction" of topic prediction—the model only introduces auxiliary information from the quality branch for appropriate adjustment when it is uncertain about its own judgment of the topic; when the model has sufficient confidence, the correction mechanism automatically degenerates into an identity mapping, without interfering with the original prediction.

[0054] Specifically, when the topic task completes confidence-aware prediction correction, its correction vector is generated by fusing features and a lightweight network. The parameters predicted by the network model are topic task-specific representations. Quality probability distribution after gradient truncation The features include the Concat feature concatenation operation, the lightweight correction network DeltaNet, and the gating weights.

[0055] .

[0056] The output layer of this DeltaNet is initialized to a zero matrix to ensure that the adjudication module does not intervene at all during the initial training phase, and the output correction is zero. The final correction is achieved through residual addition.

[0057] In one embodiment, the prediction entropy of the quality classification prediction result is dynamically weighted based on the prediction confidence of the original topic prediction result, and the dynamic consistency regularization loss is constructed by including: Step 1: Calculate the prediction confidence of the original prediction results for the topic.

[0058] Prediction confidence is a quantitative indicator of the certainty with which a model predicts the original topic, reflecting the model's confidence in the topic category to which the current sample belongs. Higher prediction confidence indicates greater certainty in the model's topic classification of the current sample; lower prediction confidence indicates greater uncertainty in the model's assessment of the current sample's topic affiliation.

[0059] In one embodiment, the prediction probability distribution of the original topic prediction results after Softmax normalization can be obtained first, and the maximum probability value in this distribution can be determined as the prediction confidence level ct. Taking a 36-category topic classification task as an example, the original topic prediction results, after Softmax, yield a 36-dimensional probability vector, and the maximum value of this vector is taken as the current prediction confidence level. In another embodiment, the prediction confidence level can also adopt other uncertainty metrics, such as the information entropy of the prediction probability distribution, the variance between each probability value, the margin between the highest probability and the second highest probability, etc., or a weighted combination of multiple metrics can be used as a comprehensive confidence level measure, as long as it can reflect the degree of certainty of the model regarding the prediction results.

[0060] Step 2: Calculate the information entropy of the predicted distribution of the quality classification prediction results.

[0061] The quality classification prediction result refers to the prediction result output by the quality classification head based on the second task features. Information entropy is an indicator used in information theory to measure the uncertainty of random variables. The larger the entropy value, the higher the uncertainty and the more uniform the distribution; the smaller the entropy value, the lower the uncertainty and the more concentrated the distribution. In quality assessment tasks, if the information entropy of the quality classification prediction result is low, it indicates that the model has a clear bias towards disclosing the quality category; if the information entropy is high, it indicates that the model is hesitant in making qualitative or quantitative judgments, and the probabilities of the two classes are close to a uniform distribution.

[0062] Here, we can assume that the probability distribution of the quality classification prediction results is obtained after Softmax normalization. The information entropy calculation function The specific calculations are as follows: ; Step 3: Weight the information entropy with the prediction confidence as the coefficient, and construct the weighted entropy loss as the dynamic consistency regularization loss.

[0063] Dynamic consistency regularization loss refers to a regularization loss term that uses prediction confidence as a dynamic weight and the quality prediction information entropy as the weighted object. Specifically, dynamic consistency regularization loss... The calculation function is as follows: .

[0064] In one embodiment, based on the classification labels corresponding to the sample ESG report texts, and through joint optimization of dynamic consistency regularization loss and multi-task classification loss, the parameters of the initial multi-task classification model are updated via end-to-end training, resulting in a trained ESG report multi-task classification model including: Step 1: Obtain the topic classification label and quality classification label corresponding to the sample ESG report text.

[0065] Topic classification labels refer to the annotation results of the ESG issue domain to which the ESG report sentence belongs, indicating the specific ESG topic category involved in the sentence. Quality classification labels refer to the annotation results of the adequacy of the information disclosure in the sentence, such as indicating whether the sentence belongs to quantitative disclosure or qualitative description. These classification labels are part of the training sample set and are pre-determined by manual or automatic annotation methods before model training. In practice, an ESG report sentence-level annotation dataset can be pre-built, with each sentence equipped with corresponding topic classification labels and quality classification labels.

[0066] Step 2: Calculate the topic classification loss based on the final topic classification prediction results and topic classification labels.

[0067] Topic classification loss is a measure of the difference between the final topic classification prediction and the topic classification label, used to measure the model's prediction bias for ESG topic classification tasks. The final topic classification prediction refers to the topic classification output modulated and updated by a one-way gating mechanism, not the original topic prediction before correction. Specifically, the topic classification loss can be calculated using the cross-entropy loss function.

[0068] Step 3: Calculate the quality classification loss based on the quality classification prediction results and quality classification labels.

[0069] Quality classification loss is a measure of the difference between the quality classification prediction result and the quality classification label, used to measure the model's prediction bias for the disclosure quality assessment task. The quality classification prediction result refers to the prediction result output by the quality classification head from the features of the second task. Step 3 quantifies the prediction error of the quality assessment task into an optimizable scalar value, which, together with the topic classification loss in step 2, constitutes the main part of the multi-task classification loss, driving the model to pursue classification accuracy in both tasks simultaneously. In practice, the quality classification loss can also be calculated using the cross-entropy loss function.

[0070] Step 4: Weight the topic classification loss, quality classification loss and dynamic consistency regularization loss to construct a joint objective function.

[0071] The joint objective function refers to the overall optimization objective formed by combining multiple loss terms in a weighted manner. It serves as the basis for backpropagation and parameter updates during end-to-end training. Here, multiple optimization objectives are unified into a single scalar function, ensuring that the model simultaneously considers topic classification accuracy, quality classification accuracy, and the logical dynamic consistency between the predictions of the two tasks in a single training iteration. The contribution ratio of each loss term can be adjusted using weight coefficients through weighted summation to adapt to different application scenarios and their varying emphases on various metrics. In practice, the joint objective function can be constructed as: Total Loss = Topic Classification Loss + Quality Classification Loss + × Dynamic consistency regularization loss, where An adjustable hyperparameter for controlling the strength of dynamic consistency regularization. When When the setting is larger, the model will place greater emphasis on the dynamic consistency of the deterministic predictions between the two tasks during training; when When the settings are smaller, the model focuses more on the classification accuracy of each task.

[0072] Step 5: Perform backpropagation based on the joint objective function to update the parameters of the initial multi-task classification model until the model converges, thus obtaining the trained ESG report multi-task classification model.

[0073] Backpropagation refers to the process of calculating parameter updates layer by layer from the output layer to the input layer based on the gradient of the joint objective function with respect to each network parameter. Model convergence refers to the state where the value of the joint objective function no longer decreases significantly on the validation set, and the model performance tends to stabilize. Here, during model training, through iterative optimization, all learnable parameters in the initial multi-task classification model are gradually adjusted, enabling the model to accurately perform ESG topic classification and disclosure quality assessment. The significance of end-to-end training is that the parameter updates of each module within the model—including the shared text encoder, the first task adapter, the second task adapter, the topic classification head, the quality classification head, and the one-way gating adjudication module—are all carried out collaboratively under the guidance of a unified optimization objective, rather than being trained independently in stages or by modules, thereby ensuring that the cooperation between modules reaches global optimum.

[0074] Specifically, the overall loss is minimized when the model converges to its optimal state. The parameters predicted by the network model are the topic classification loss parameters. Dynamic consistency regularization loss The model's overall loss E is as follows, along with the regularization loss weight parameter α: .

[0075] In practical applications, to verify the effectiveness of the multi-task framework consisting of a shared encoder and a task-specific adapter, experiments were conducted on a Chinese ESG report sentence-level annotation dataset. Experimental results show that the Macro-F1 score of the single-task topic classification model is 82.34%, and that of the single-task quality classification model is 92.50%. After introducing task-specific adapters for ESG topic classification and quality assessment tasks on top of the shared text encoder, the multi-task joint model achieves a topic classification Macro-F1 score of 84.03%, an improvement of 1.69 percentage points compared to the single-task topic classification baseline; and a quality classification Macro-F1 score of 93.02%, an improvement of 0.52 percentage points compared to the single-task quality classification baseline. These results demonstrate that the shared encoder and task-specific adapter can provide an effective multi-task feature foundation for ESG topic identification and disclosure quality assessment, and indicate the existence of usable collaborative information between topic classification and quality assessment tasks, providing an experimental basis for further introducing a one-way gating modulation mechanism and dynamic consistency regularization loss.

[0076] In one embodiment, after jointly optimizing the classification labels corresponding to the ESG report text, and the dynamic consistency regularization loss and multi-task classification loss, and updating the parameters of the initial multi-task classification model through end-to-end training to obtain the trained ESG report multi-task classification model, the method further includes: Step 1: Obtain the ESG report text to be classified.

[0077] Step 2: Input the ESG report text to be classified into the trained ESG report multi-task classification model to obtain ESG topic classification results and disclosure quality assessment results.

[0078] The ESG report text to be classified refers to a brand-new ESG report sentence that needs to be automatically analyzed by a trained model. It does not have topic classification labels or quality classification labels before being input into the model. The ESG topic classification result refers to the predicted output of the trained model for the ESG issue domain to which the input sentence belongs, and the disclosure quality assessment result refers to the predicted output of the trained model for the sufficiency of information disclosure in the input sentence. After the text to be classified is input into the trained model, the model automatically performs a forward propagation process, sequentially passing through feature extraction by a shared text encoder, task-specific representation generation by the first and second task adapters, and modulation and updating by a one-way gating mechanism. Finally, the topic classification head and quality classification head output the classification predictions for the two tasks, respectively.

[0079] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0080] Based on the same inventive concept, this application also provides an ESG reporting multi-task classification model construction apparatus for implementing the aforementioned ESG reporting multi-task classification model construction method. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more ESG reporting multi-task classification model construction apparatus embodiments provided below can be found in the limitations of the ESG reporting multi-task classification model construction method described above, and will not be repeated here.

[0081] In one embodiment, such as Figure 4 As shown, an apparatus for constructing a multi-task classification model for ESG reports is provided, comprising: The initialization module 100 is used to obtain the initial multi-task classification model and sample ESG report text, and to extract features from the sample ESG report text using the shared text encoder in the initial multi-task classification model to obtain a global context representation; the initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter; The feature extraction module 200 is used to input the global context representation into the first task adapter and the second task adapter respectively, and generate the first task feature for the ESG topic classification task and the second task feature for the quality assessment task. The classification prediction module 300 is used to generate a one-way gating signal based on the first task feature using a one-way gating mechanism, and modulate and update the original topic prediction result obtained by the independent prediction of the second task feature using the one-way gating signal and the quality classification prediction result to obtain the final topic classification prediction result. The loss construction module 400 is used to dynamically weight the prediction entropy of the quality classification prediction results based on the prediction confidence of the original topic prediction results, and construct a dynamic consistency regularization loss. The model training module 500 is used to jointly optimize the classification labels corresponding to the sample ESG report text, as well as the dynamic consistency regularization loss and multi-task classification loss. It updates the parameters of the initial multi-task classification model through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

[0082] In one embodiment, the feature extraction module 200 is further configured to input the global context representation into a first task adapter, perform layer normalization, dimensionality reduction, nonlinear activation, and random deactivation processing sequentially, and then perform dimensionality increase processing to obtain a first task-specific offset; add the first task-specific offset to the global context representation element-wise to obtain a first task feature for the ESG topic classification task; input the global context representation into a second task adapter, perform layer normalization, dimensionality reduction, nonlinear activation, and random deactivation processing sequentially, and then perform dimensionality increase processing to obtain a second task-specific offset; add the second task-specific offset to the global context representation element-wise to obtain a second task feature for the quality assessment task; wherein the first task adapter and the second task adapter adopt a bottleneck structure design with the same structure and independent parameters; the weight matrix and bias vector of the dimensionality increase processing are initialized to zero matrix and zero vector, so that the first task-specific offset and the second task-specific offset are both zero in the initial training stage, and the first task feature and the second task feature are both equal to the global context representation in the initial training stage; the dimensionality reduction processing reduces the dimension of the global context representation to the bottleneck dimension d.

[0083] In one embodiment, the initial multi-task classification model further includes a topic classification head and a quality classification head; the classification prediction module 300 is further configured to input the first task features into the topic classification head to generate the original topic prediction result, and generate a one-way gating signal based on the prediction probability of the original topic prediction result; input the second task features into the quality classification head to generate the quality classification prediction result, and perform gradient truncation processing on the quality classification prediction result to cut off the reverse gradient path of the quality classification prediction result; concatenate the first task features with the gradient-truncation-processed quality classification prediction result, and generate a correction vector for the concatenated features through a correction network; multiply the one-way gating signal with the correction vector, and inject the product into the original topic prediction result using residual addition to obtain the final topic classification prediction result.

[0084] In one embodiment, the loss construction module 400 is further used to calculate the prediction confidence of the original prediction result of the topic; calculate the information entropy of the prediction distribution of the quality classification prediction result; and weight the information entropy with the prediction confidence as a coefficient to construct a weighted entropy loss as a dynamic consistency regularization loss.

[0085] In one embodiment, the loss construction module 400 is further configured to obtain the prediction probability distribution of the original prediction result of the topic; and determine the maximum probability in the prediction probability distribution as the prediction confidence corresponding to the original prediction result of the topic.

[0086] In one embodiment, the model training module 500 is further configured to obtain the topic classification label and quality classification label corresponding to the sample ESG report text; calculate the topic classification loss based on the final topic classification prediction result and the topic classification label; calculate the quality classification loss based on the quality classification prediction result and the quality classification label; weight and sum the topic classification loss, quality classification loss and dynamic consistency regularization loss to construct a joint objective function; perform backpropagation based on the joint objective function to update the parameters of the initial multi-task classification model until the model converges, thus obtaining the trained ESG report multi-task classification model.

[0087] In one embodiment, the above-mentioned ESG report multi-task classification model construction apparatus further includes: The classification module is used to obtain the ESG report text to be classified; the ESG report text to be classified is input into the trained ESG report multi-task classification model to obtain the ESG topic classification results and disclosure quality assessment results.

[0088] The modules in the aforementioned ESG reporting multi-task classification model construction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0089] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores preset data. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer program implements a method for constructing a multi-task classification model for ESG reporting.

[0090] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0091] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method for constructing a multi-task classification model for ESG reports.

[0092] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the above-described method for constructing a multi-task classification model for ESG reports.

[0093] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described method for constructing a multi-task classification model for ESG reports.

[0094] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for constructing a multi-task classification model for ESG reports, characterized in that, The method includes: An initial multi-task classification model and sample ESG report texts are obtained. The shared text encoder in the initial multi-task classification model is used to extract features from the sample ESG report texts to obtain a global context representation. The initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter. The global context representation is input into the first task adapter and the second task adapter respectively to generate a first task feature for the ESG topic classification task and a second task feature for the quality assessment task. Using a one-way gating mechanism, a one-way gating signal is generated based on the first task feature. The quality classification prediction result obtained by independent prediction of the second task feature is modulated and updated using the one-way gating signal and the quality classification prediction result to obtain the original topic prediction result obtained by independent prediction of the first task feature, so as to obtain the final topic classification prediction result. The prediction entropy of the quality classification prediction result is dynamically weighted based on the prediction confidence of the original prediction result of the topic, and a dynamic consistency regularization loss is constructed. Based on the classification labels corresponding to the sample ESG report text, and the joint optimization of the dynamic consistency regularization loss and multi-task classification loss, the parameters of the initial multi-task classification model are updated through end-to-end training to obtain the trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

2. The method according to claim 1, characterized in that, The step of inputting the global context representation into the first task adapter and the second task adapter respectively to generate a first task feature for the ESG topic classification task and a second task feature for the quality assessment task includes: The global context representation is input into the first task adapter, and after performing layer normalization, dimensionality reduction, nonlinear activation, and random deactivation processing in sequence, it is then subjected to dimensionality increase processing to obtain the first task-specific offset. The first task-specific offset is added element-wise to the global context representation to obtain the first task feature for the ESG topic classification task. The global context representation is input into the second task adapter, and after performing layer normalization, dimensionality reduction, nonlinear activation, and random deactivation processing in sequence, it is then subjected to dimensionality-up mapping processing to obtain the second task-specific offset. The second task-specific offset is then added element-wise to the global context representation to obtain the second task feature for the quality assessment task. The first task adapter and the second task adapter adopt the same bottleneck structure design with independent parameters; the weight matrix and bias vector of the dimensionality increase process are initialized to zero matrix and zero vector, so that the first task specialization offset and the second task specialization offset are both zero in the initial training stage, and the first task feature and the second task feature are both equal to the global context representation in the initial training stage; the dimensionality reduction process reduces the dimension of the global context representation to the bottleneck dimension d.

3. The method according to claim 1, characterized in that, The initial multi-task classification model also includes a topic classification head and a quality classification head; The method of utilizing a one-way gating mechanism to generate a one-way gating signal based on the first task feature, and using the quality classification prediction result obtained by independently predicting the second task feature to modulate and update the original topic prediction result obtained by independently predicting the first task feature using the one-way gating signal and the quality classification prediction result, to obtain the final topic classification prediction result includes: The first task feature is input into the topic classification head to generate the original topic prediction result, and the one-way gating signal is generated based on the prediction probability of the original topic prediction result. The second task feature is input into the quality classification head to generate the quality classification prediction result, and gradient truncation processing is performed on the quality classification prediction result to cut off the reverse gradient path of the quality classification prediction result. The first task features are concatenated with the quality classification prediction results after gradient truncation, and the concatenated features are used to generate a correction vector through a correction network. The one-way gating signal is multiplied by the correction vector, and the product is injected into the original topic prediction result using residual addition to obtain the final topic classification prediction result.

4. The method according to claim 1, characterized in that, The prediction entropy of the quality classification prediction result is dynamically weighted based on the prediction confidence of the original prediction result of the topic, and a dynamic consistency regularization loss is constructed, including: Calculate the prediction confidence of the original prediction results for the topic; Calculate the information entropy of the predicted distribution of the quality classification prediction results; The information entropy is weighted using the prediction confidence as a coefficient to construct a weighted entropy loss as a dynamic consistency regularization loss.

5. The method according to claim 4, characterized in that, The calculation of the prediction confidence of the original prediction result of the topic includes: Obtain the predicted probability distribution of the original prediction results for the topic; The maximum probability in the predicted probability distribution is determined as the prediction confidence level corresponding to the original prediction result of the topic.

6. The method according to claim 1, characterized in that, The process of jointly optimizing the classification labels corresponding to the sample ESG report texts, and the dynamic consistency regularization loss and multi-task classification loss, and updating the parameters of the initial multi-task classification model through end-to-end training to obtain the trained ESG report multi-task classification model includes: Obtain the topic classification label and quality classification label corresponding to the sample ESG report text; Based on the final topic classification prediction result and the topic classification label, the topic classification loss is calculated; Based on the quality classification prediction results and the quality classification labels, calculate the quality classification loss; The topic classification loss, the quality classification loss, and the dynamic consistency regularization loss are weighted and summed to construct a joint objective function; Backpropagation is performed based on the joint objective function to update the parameters of the initial multi-task classification model until the model converges, thus obtaining the trained ESG report multi-task classification model.

7. The method according to claim 1, characterized in that, After jointly optimizing the classification labels corresponding to the sample ESG report texts, and the dynamic consistency regularization loss and multi-task classification loss, and updating the parameters of the initial multi-task classification model through end-to-end training to obtain the trained ESG report multi-task classification model, the process further includes: Obtain the text of the ESG report to be classified; The ESG report text to be classified is input into the trained ESG report multi-task classification model to obtain ESG topic classification results and disclosure quality assessment results.

8. A device for constructing a multi-task classification model for ESG reports, characterized in that, The device includes: An initialization module is used to obtain an initial multi-task classification model and sample ESG report texts, and to extract features from the sample ESG report texts using a shared text encoder in the initial multi-task classification model to obtain a global context representation; the initial multi-task classification model includes a shared text encoder, a first task adapter, and a second task adapter; The feature extraction module is used to input the global context representation into the first task adapter and the second task adapter respectively, and generate a first task feature for the ESG topic classification task and a second task feature for the quality assessment task. The classification prediction module is used to generate a one-way gating signal based on the first task feature using a one-way gating mechanism, and modulate and update the original topic prediction result obtained by the independent prediction of the second task feature using the one-way gating signal and the quality classification prediction result to obtain the final topic classification prediction result. The loss construction module is used to dynamically weight the prediction entropy of the quality classification prediction result based on the prediction confidence of the original prediction result of the topic, and construct a dynamic consistency regularization loss. The model training module is used to jointly optimize the classification labels corresponding to the sample ESG report text, as well as the dynamic consistency regularization loss and multi-task classification loss. It updates the parameters of the initial multi-task classification model through end-to-end training to obtain a trained ESG report multi-task classification model. The multi-task classification loss includes the topic classification loss calculated based on the final topic classification prediction result and the quality classification loss calculated based on the quality classification prediction result.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images