A method and device for enterprise risk identification combined with sentiment classification

Abstract

This invention relates to a method and apparatus for identifying corporate risks by combining sentiment classification. The method involves manually constructing and encoding discrete sentiment prompts to obtain continuous sentiment prompts; adding these continuous sentiment prompts before the embedded representation of a corporate news report, and then inputting this data into a first language model to obtain text features; performing sentiment classification based on these text features to obtain negative corporate news reports; adding a risk prompt matrix vector and an embedded representation of the marker [CLS] before the embedded representation of the negative corporate news report, and then inputting this data into a second language model to obtain sentence vectors corresponding to the negative corporate news reports; and finally matching corresponding risk tags to the negative corporate news reports based on these sentence vectors. This invention employs a two-stage approach: sentiment classification of corporate news reports and risk tag identification of negative news reports. This reduces interference from non-negative news reports and improves the accuracy of risk tag identification results.

Description

Technical Field

[0001] This invention relates to the field of data mining technology, specifically a method and apparatus for identifying enterprise risks by combining sentiment classification. Background Technology

[0002] If a company mishandles negative news, it can trigger a chain reaction, generate related public opinion, and severely damage its overall image. Therefore, companies need to handle crises promptly and correctly, minimizing the impact of negative events and creating a favorable public discourse environment. With the support of big data and artificial intelligence technologies, various financial risk control methods are developing rapidly. Among them, collecting corporate public opinion information from the internet to uncover potential risk events is a relatively effective approach. Identifying corporate risk events from internet news allows companies to grasp the risk situation more promptly and comprehensively, improving their risk identification and monitoring capabilities. Risk events exist in textual form, and corporate risk identification is essentially a form of text classification. The internet is filled with a wealth of new information, making the ability to accurately identify negative news and risk labels for companies an urgent need. Summary of the Invention

[0003] This invention addresses the technical problems existing in the prior art by providing a method and apparatus for identifying corporate risks that combines sentiment classification. By adding continuous sentiment prompts and deeply continuous risk prompts, it can more accurately identify the attitudes of internet users towards relevant companies and the risk tags corresponding to corporate news reports.

[0004] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0005] In a first aspect, the present invention provides a method for identifying enterprise risks by combining sentiment classification, comprising:

[0006] Obtain corporate news reports using web crawling technology;

[0007] For the aforementioned corporate news reports, discrete sentiment prompt statements are manually constructed, and these discrete sentiment prompt statements are encoded to obtain continuous sentiment prompt statements.

[0008] After adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text is input into a pre-trained first language model to obtain the text features corresponding to the corporate news report. Based on the text features, the sentiment of the corporate news report is classified to obtain negative corporate news reports.

[0009] After adding a pre-trained risk warning matrix vector and an embedding representation of the marker [CLS] before the embedded representation of the negative news report of the enterprise, the vector is input into a pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise.

[0010] The sentence vector is used to match the corresponding risk label for the negative news reports about the company.

[0011] Furthermore, for the aforementioned corporate news reports, discrete sentiment prompt statements are manually constructed, and these discrete sentiment prompt statements are encoded to obtain continuous sentiment prompt statements, including:

[0012] The discrete sentiment prompt statement consists of a guiding word generated for the target sentiment tag, a position slot [MASK] for generating the sentiment tag, and other words. The guiding word generated for the target sentiment tag is denoted as context word C, the position slot [MASK] for generating the sentiment tag is denoted as M, and the other words are denoted as pseudo-prompt words P.

[0013] The pseudo-prompt word P is encoded using an emotion prompting statement encoder to obtain a continuous vector representation of the pseudo-prompt word P;

[0014] In the embedding layer e of the first language model, the continuous vector representation of the pseudo-cue word P, the embedding representation e(C) of the context word C, and the embedding representation e(M) of the position slot [MASK] are combined to form the continuous sentiment cue corresponding to the discrete sentiment cue.

[0015] Furthermore, the emotion prompting statement encoder is implemented through a bidirectional long short-term memory network.

[0016] Furthermore, based on the aforementioned text features, sentiment classification of corporate news reports is performed to obtain negative corporate news reports, including:

[0017] Let V be the vocabulary of the first language model;

[0018] The text features are sequentially transformed through a linear layer, regularization, and another linear layer to obtain the probability of generating any word in the vocabulary V in the position slot [MASK].

[0019] The word with the highest probability is selected as the generated word for the position slot [MASK], and recorded as the sentiment tag of the corporate news report;

[0020] The emotional tags are classified according to the mapping relationship between the emotional tags and the emotional classification tags to obtain negative news reports about enterprises. The emotional classification tags include positive, negative and neutral.

[0021] Furthermore, the training method for the first language model includes:

[0022] Prepare training data and initialize the first language model. The training data consists of corporate news reports with known sentiment labels.

[0023] In the embedding layer of the first language model, after adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text features corresponding to the corporate news report are obtained, and the sentiment tags of the corporate news report are generated based on the text features.

[0024] Calculate the loss value, and terminate training when the loss value is minimized.

[0025] Furthermore, the training method for the risk warning matrix vector and the second language model includes:

[0026] Randomly initialize the risk warning matrix vector H and the marker "[CLS]";

[0027] In the embedding layer of the second language model, a randomly initialized risk warning matrix vector and the embedding representation of the marker "[CLS]" are added before the known corporate negative news report embedding representation e[negative news text], and then input into the first layer encoder of the second language model; the second language model includes multiple encoders;

[0028] The input to the next encoder layer is a randomly initialized risk warning matrix vector and the updated e[CLS] and e[negative news text] output by the previous encoder layer after cross-learning. The model parameters of each encoder layer are saved during training.

[0029] The risk warning matrix vector output by the last encoder layer of the second language model is the trained risk warning matrix vector. The parameters of the trained second language model are the model parameters of each encoder layer saved during the training process.

[0030] Furthermore, the process of matching corresponding risk tags for negative news reports about the company based on the sentence vector includes:

[0031] The feature vector corresponding to [CLS] is converted into a probability value for each label using softmax, and the risk label with the highest probability value is selected as the final risk label.

[0032] Secondly, the present invention provides a business risk identification device that combines emotion classification, comprising:

[0033] The data crawling module uses web crawling technology to obtain corporate news reports;

[0034] The encoding module manually constructs discrete sentiment prompt statements for the company's news reports and encodes these discrete sentiment prompt statements to obtain continuous sentiment prompt statements.

[0035] The sentiment classification module adds the continuous sentiment prompts before the embedded representation of the corporate news report, inputs them into a pre-trained first language model, obtains the text features corresponding to the corporate news report, and performs sentiment classification of the corporate news report based on the text features to obtain negative corporate news reports.

[0036] The sentence vector generation module adds a pre-trained risk warning matrix vector and the embedding representation of the marker [CLS] before the embedding representation of the negative news report of the enterprise. Then, it inputs the vector into the pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise.

[0037] The risk tag matching module matches the corresponding risk tags for the negative news reports about the company based on the sentence vector.

[0038] Thirdly, the present invention provides an electronic device, comprising:

[0039] Memory, used to store computer software programs;

[0040] A processor is used to read and execute the computer software program, thereby implementing the enterprise risk identification method combining sentiment classification as described in the first aspect of the present invention.

[0041] Fourthly, the present invention provides a non-transitory computer-readable storage medium storing a computer software program for implementing the enterprise risk identification method combining sentiment classification as described in the first aspect of the present invention.

[0042] The beneficial effects of this invention are as follows: Firstly, sentiment classification of crawled corporate news reports mitigates the interference of non-negative news on subsequent risk label identification to some extent. Furthermore, for sentiment classification of corporate news reports and risk label identification of negative news reports, adding continuous sentiment prompts and deeply continuous risk prompts enables the automatic construction of these prompts, resulting in more accurate risk label identification results. The purpose of using continuous sentiment prompts and deeply continuous risk prompts is to find a method that enables language models to better perform the task of risk label identification. Therefore, it is not necessary to consider whether the continuous sentiment prompts and deeply continuous risk prompts are composed of natural language or whether humans can understand them. In addition, the advantage of using continuous sentiment prompts and deeply continuous risk prompts is that suitable sentiment prompts and risk prompts can be automatically sought without heavily relying on prior human knowledge.

[0043] This invention combines continuous sentiment prompts and deeply continuous risk prompts to classify corporate news reports by sentiment and identify risk tags for negative news reports. This results in more accurate risk tag identification and provides accurate data for enterprises to conduct risk event mining and risk monitoring. Attached Figure Description

[0044] Figure 1 A schematic diagram of a corporate risk identification method combining sentiment classification is provided in an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of a corporate risk identification device combining emotion classification is provided in an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of an embodiment of the electronic device provided in this invention;

[0047] Figure 4 This is a schematic diagram of an embodiment of a computer-readable storage medium provided in this invention. Detailed Implementation

[0048] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0049] like Figure 1 As shown, this embodiment of the invention provides a method for identifying enterprise risks by combining sentiment classification, including the following steps:

[0050] S1 uses web crawling technology to collect news data from companies on the Internet.

[0051] S2 involves manually constructing discrete sentiment prompt statements for the aforementioned corporate news reports, and encoding these discrete sentiment prompt statements to obtain continuous sentiment prompt statements.

[0052] The objective is to classify corporate news reports by sentiment, and discrete sentiment prompts are designed manually. There are two ways to design discrete sentiment prompts: one is to design them as complete sentences, such as "What is the sentiment of this news article?" in a QA task; the second is to design them as sentences with empty slots, such as "The sentiment of this news article is [MASK]." in Prompt Learning, where "[MASK]" represents an empty slot where a sentiment label needs to be generated. This embodiment uses the second approach to design sentiment prompts.

[0053] Discrete sentiment prompts are unstable, dependent on the experience of the prompt designer, and because language models are inherently continuous, using discrete sentiment prompts cannot achieve global optimum. Finding a continuous sentiment prompt is crucial for sentiment classification in corporate news reports. Taking the prompt with empty slots, "The sentiment tendency of this news is [MASK]," as an example, the guiding word "sentiment tendency" generated for the target sentiment label is the context word C, [MASK] is denoted as M, and the other words "The...of this news is..." are called pseudo-prompt words P. Psuppository words P are selected from the discrete sentiment prompts, and a sentiment prompt encoder is used to obtain a continuous vector representation of P. The sentiment prompt encoder consists of a bidirectional Long Short-Term Memory (LSTM) network, using a two-layer Multilayer Perceptron (MLP) activated by ReLU to excite the discreteness. The resulting h sequence serves as the corresponding vector feature, which is the continuous vector prompt for the pseudo-prompt word P.

[0054] Then, in the embedding layer e of the first language model, the continuous vector representation of the pseudo-cue word P, the embedding representation e(C) of the context word C, and the embedding representation e(M) of the position slot [MASK] are combined to form the continuous sentiment cue corresponding to the discrete sentiment cue.

[0055] S3, after adding the continuous sentiment prompts before the embedded representation of the corporate news report, input it into the pre-trained first language model to obtain the text features corresponding to the corporate news report, and perform sentiment classification of the corporate news report based on the text features to obtain negative corporate news reports.

[0056] For a first language model LM, its embedding layer e∈LM, word vectors are used to represent the input of the first language model in the embedding layer e, and the vocabulary of the language model LM is denoted as V.

[0057] In the embedding layer e, the continuous sentiment cues h, the embedding representation "e(C)" of the context word C, and the embedding representation "e(M)" of M are combined to form a continuous sentiment cues statement. This continuous sentiment cues statement serves as a prefix to the news text embedding representation "e(news text)" in the embedding layer, and is encoded by a multi-layer sentiment encoder to obtain text features. The obtained text features are then transformed through a linear layer, regularization, and another linear layer to obtain the probability of generating any word in the vocabulary V at the [MASK] position (i.e., position M). The word with the highest probability is selected as the generated word for the empty slot [MASK] and recorded as the sentiment label of the corporate news report. These sentiment labels include: fines or investigations by regulatory agencies, changes in actual controllers, bankruptcy reorganization, overdue debts, regulatory measures taken, main board, growth enterprise board, SME board, bond delisting, excessively high stock pledge ratio, environmental pollution, information disclosure violations and safety accidents, penalties imposed by government departments, etc.

[0058] Since the sentiment tags generated by the first language model are not limited to positive, negative, and neutral, it is necessary to perform a tag mapping between the generated sentiment tags and the sentiment tag set S (S = {positive, negative, neutral}), that is, to classify the sentiment tags to obtain negative news reports about enterprises.

[0059] The training process of the first language model is as follows:

[0060] Prepare training data and initialize the first language model. The training data consists of corporate news reports with known sentiment labels.

[0061] In the embedding layer of the first language model, after adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text features corresponding to the corporate news report are obtained, and the sentiment tags of the corporate news report are generated based on the text features.

[0062] Calculate the loss value and terminate training when the loss value is minimized. Generally, the loss value gradually decreases first, reaches a minimum, and then increases. The model with the minimum loss value is considered the best-performing model.

[0063] S4. After adding the pre-trained risk warning matrix vector and the embedding representation of the marker [CLS] before the embedding representation of the negative news report of the enterprise, it is input into the pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], that is, the sentence vector corresponding to the negative news report of the enterprise.

[0064] The risk warning matrix vector and the training method of the second language model include:

[0065] Randomly initialize the risk warning matrix vector H and the marker "[CLS]";

[0066] In the embedding layer of the second language model, a randomly initialized risk warning matrix vector and the embedding representation of the marker "[CLS]" are added before the known corporate negative news report embedding representation e[negative news text], and then input into the first layer encoder of the second language model; the second language model includes multiple encoders;

[0067] The input to the next encoder layer is a randomly initialized risk warning matrix vector and the updated e[CLS] and e[negative news text] output by the previous encoder layer after cross-learning. The model parameters of each encoder layer are saved during training.

[0068] The risk warning matrix vector output by the last encoder layer of the second language model is the trained risk warning matrix vector. The parameters of the trained second language model are the model parameters of each encoder layer saved during the training process.

[0069] S5. Match the corresponding risk label to the negative news report about the company based on the sentence vector. Convert the feature vector corresponding to [CLS] into a probability value for each label using softmax, and select the risk label with the highest probability value as the final risk label.

[0070] The first and second language models can be of the same type or different types. In this embodiment of the invention, both the first and second language models are based on a 12-layer BERT, where LM is composed of 12 layers of Transformer Encoders stacked together, and the embedding layer e∈LM.

[0071] By adding continuous sentiment prompts and deep continuous risk prompts before the embedded representation of news text, vectors specific to sentiment classification and risk label recognition tasks are optimized only, thus preserving different prefixes for different tasks and reducing the cost of fine-tuning. Secondly, using continuous sentiment prompts and deep continuous risk prompts allows for the search for suitable and optimal prompts during the prompt optimization process, making the sentiment classification and risk label recognition results more accurate. Finally, adopting a two-stage approach—sentiment classification of corporate news reports and risk label recognition of negative news reports—reduces interference from non-negative news reports and improves the accuracy of risk label recognition results.

[0072] like Figure 2 As shown, this embodiment of the invention also provides a business risk identification device that combines emotion classification, including:

[0073] The data crawling module uses web crawling technology to obtain corporate news reports;

[0074] The encoding module manually constructs discrete sentiment prompt statements for the company's news reports and encodes these discrete sentiment prompt statements to obtain continuous sentiment prompt statements.

[0075] The sentiment classification module adds the continuous sentiment prompts before the embedded representation of the corporate news report, inputs them into a pre-trained first language model, obtains the text features corresponding to the corporate news report, and performs sentiment classification of the corporate news report based on the text features to obtain negative corporate news reports.

[0076] The sentence vector generation module adds a pre-trained risk warning matrix vector and the embedding representation of the marker [CLS] before the embedding representation of the negative news report of the enterprise. Then, it inputs the vector into the pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise.

[0077] The risk tag matching module matches the corresponding risk tags for the negative news reports about the company based on the sentence vector.

[0078] Please see Figure 3 , Figure 3 This is a schematic diagram illustrating an embodiment of the electronic device provided in this invention. For example... Figure 3 As shown, this embodiment of the invention provides an electronic device 500, including a memory 510, a processor 520, and a computer program 511 stored in the memory 520 and executable on the processor 520. When the processor 520 executes the computer program 511, it performs the following steps:

[0079] Obtain corporate news reports using web crawling technology;

[0080] For the aforementioned corporate news reports, discrete sentiment prompt statements are manually constructed, and these discrete sentiment prompt statements are encoded to obtain continuous sentiment prompt statements.

[0081] After adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text is input into a pre-trained first language model to obtain the text features corresponding to the corporate news report. Based on the text features, the sentiment of the corporate news report is classified to obtain negative corporate news reports.

[0082] After adding a pre-trained risk warning matrix vector and an embedding representation of the marker [CLS] before the embedded representation of the negative news report of the enterprise, the vector is input into a pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise.

[0083] The sentence vector is used to match the corresponding risk label for the negative news reports about the company.

[0084] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. For example... Figure 4 As shown, this embodiment provides a computer-readable storage medium 600, on which a computer program 611 is stored. When the computer program 611 is executed by a processor, it performs the following steps:

[0085] Obtain corporate news reports using web crawling technology;

[0086] For the aforementioned corporate news reports, discrete sentiment prompt statements are manually constructed, and these discrete sentiment prompt statements are encoded to obtain continuous sentiment prompt statements.

[0087] After adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text is input into a pre-trained first language model to obtain the text features corresponding to the corporate news report. Based on the text features, the sentiment of the corporate news report is classified to obtain negative corporate news reports.

[0088] After adding a pre-trained risk warning matrix vector and an embedding representation of the marker [CLS] before the embedded representation of the negative news report of the enterprise, the vector is input into a pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise.

[0089] The sentence vector is used to match the corresponding risk label for the negative news reports about the company.

[0090] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0091] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0092] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0093] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0094] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0095] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0096] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for identifying enterprise risks by combining sentiment classification, characterized in that, include: Obtain corporate news reports using web crawling technology; For the aforementioned corporate news reports, discrete sentiment prompt statements are manually constructed, and these discrete sentiment prompt statements are encoded to obtain continuous sentiment prompt statements. After adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text is input into a pre-trained first language model to obtain the text features corresponding to the corporate news report. Based on the text features, the sentiment of the corporate news report is classified to obtain negative corporate news reports. After adding a pre-trained risk warning matrix vector and an embedding representation of the marker [CLS] before the embedded representation of the negative news report of the enterprise, the vector is input into a pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise. Based on the sentence vector, match the corresponding risk label for the negative news report about the company; The encoding of discrete emotional prompts includes: The discrete sentiment prompt statement consists of a guiding word generated for the target sentiment tag, a position slot [MASK] for generating the sentiment tag, and other words. The guiding word generated for the target sentiment tag is denoted as context word C, the position slot [MASK] for generating the sentiment tag is denoted as M, and the other words are denoted as pseudo-prompt words P. The pseudo-prompt word P is encoded using an emotion prompting statement encoder to obtain a continuous vector representation of the pseudo-prompt word P.

2. The method according to claim 1, characterized in that, For the aforementioned corporate news reports, discrete sentiment prompts are manually constructed, and these discrete sentiment prompts are encoded to obtain continuous sentiment prompts, including: In the embedding layer e of the first language model, the continuous vector representation of the pseudo-cue word P, the embedding representation e(C) of the context word C, and the embedding representation e(M) of the position slot [MASK] are combined to form the continuous sentiment cue corresponding to the discrete sentiment cue.

3. The method according to claim 2, characterized in that, The emotion prompting encoder is implemented through a bidirectional long short-term memory network.

4. The method according to claim 2, characterized in that, Based on the aforementioned text features, sentiment classification of corporate news reports is performed to obtain negative corporate news reports, including: Let V be the vocabulary of the first language model; The text features are sequentially transformed through a linear layer, regularization, and another linear layer to obtain the probability of generating any word in the vocabulary V in the position slot [MASK]. The word with the highest probability is selected as the generated word for the position slot [MASK], and recorded as the sentiment tag of the corporate news report; The emotional tags are classified according to the mapping relationship between the emotional tags and the emotional classification tags to obtain negative news reports about enterprises. The emotional classification tags include positive, negative and neutral.

5. The method according to claim 1, characterized in that, The training methods for the first language model include: Prepare training data and initialize the first language model. The training data consists of corporate news reports with known sentiment labels. In the embedding layer of the first language model, after adding the continuous sentiment prompts before the embedded representation of the corporate news report, the text features corresponding to the corporate news report are obtained, and the sentiment tags of the corporate news report are generated based on the text features. Calculate the loss value, and terminate training when the loss value is minimized.

6. The method according to claim 1, characterized in that, The risk warning matrix vector and the training method of the second language model include: Randomly initialize the risk warning matrix vector H and the marker "[CLS]"; In the embedding layer of the second language model, a randomly initialized risk warning matrix vector and the embedding representation of the marker "[CLS]" are added before the known corporate negative news report embedding representation e[negative news text], and then input into the first layer encoder of the second language model; the second language model includes multiple encoders; The input to the next encoder layer is a randomly initialized risk warning matrix vector and the updated e[CLS] and e[negative news text] output by the previous encoder layer after cross-learning. The model parameters of each encoder layer are saved during training. The risk warning matrix vector output by the last encoder layer of the second language model is the trained risk warning matrix vector. The parameters of the trained second language model are the model parameters of each encoder layer saved during the training process.

7. The method according to claim 1, characterized in that, The method of matching corresponding risk tags for negative news reports about the company based on the sentence vector includes: The feature vector corresponding to [CLS] is converted into a probability value for each label using softmax, and the risk label with the highest probability value is selected as the final risk label.

8. A business risk identification device combining emotion classification, characterized in that, include: The data crawling module uses web crawling technology to obtain corporate news reports; The encoding module manually constructs discrete sentiment prompt statements for the company's news reports and encodes these discrete sentiment prompt statements to obtain continuous sentiment prompt statements. The sentiment classification module adds the continuous sentiment prompts before the embedded representation of the corporate news report, inputs them into a pre-trained first language model, obtains the text features corresponding to the corporate news report, and performs sentiment classification of the corporate news report based on the text features to obtain negative corporate news reports. The sentence vector generation module adds a pre-trained risk warning matrix vector and the embedding representation of the marker [CLS] before the embedding representation of the negative news report of the enterprise. Then, it inputs the vector into the pre-trained second language model to obtain the feature vector corresponding to the marker [CLS], which is the sentence vector corresponding to the negative news report of the enterprise. The risk label matching module matches the corresponding risk labels for the negative news reports about the company based on the sentence vector. The encoding of discrete emotional prompts includes: The discrete sentiment prompt statement consists of a guiding word generated for the target sentiment tag, a position slot [MASK] for generating the sentiment tag, and other words. The guiding word generated for the target sentiment tag is denoted as context word C, the position slot [MASK] for generating the sentiment tag is denoted as M, and the other words are denoted as pseudo-prompt words P. The pseudo-prompt word P is encoded using an emotion prompting statement encoder to obtain a continuous vector representation of the pseudo-prompt word P.

9. An electronic device, characterized in that, include: Memory, used to store computer software programs; A processor for reading and executing the computer software program, thereby implementing the enterprise risk identification method combining sentiment classification as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer software program for implementing the enterprise risk identification method combining emotion classification as described in any one of claims 1-7.

Images