Risk plan generation method and device, nonvolatile storage medium and electronic equipment

Abstract

The invention discloses a risk plan generation method and device, a nonvolatile storage medium and electronic equipment. The method comprises the following steps: inputting to-be-identified data into a problem identification model, and obtaining a problem identification result set output by the problem identification model; the problem recognition result is input into a risk recognition sub-model, a risk recognition result output by the risk recognition sub-model is obtained, the risk recognition result comprises a risk type corresponding to the problem recognition result and a risk value corresponding to the problem recognition result, and the risk value is used for indicating the severity degree of the risk; the risk identification result is input into a plan generation model, a risk plan output by the plan generation model is obtained, and the plan generation model is a large language model. The technical problems of high generation cost and low accuracy caused by the fact that an existing risk plan generation technology only supports single-mode data input and depends on a large model are solved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and more particularly, to a method, apparatus, non-volatile storage medium, and electronic device for generating risk contingency plans. Background Art

[0002] LLM (Large Language Models) has made significant breakthroughs in recent years and has impressed in many fields. Both the deployment and training of large models require costs. If large models are used regardless of the task size, the overhead will undoubtedly be huge and the cost performance will be extremely low.

[0003] In view of the above problems, no effective solution has been proposed yet. Summary of the Invention

[0004] Embodiments of this application provide a method, apparatus, non-volatile storage medium, and electronic device for generating risk contingency plans to at least solve the technical problems of high generation cost and low accuracy caused by the existing risk contingency plan generation technology that only supports single-modal data input and relies on large models.

[0005] According to one aspect of the embodiments of this application, a method for generating a risk contingency plan is provided, including: inputting data to be recognized into a problem recognition model to obtain a set of problem recognition results output by the problem recognition model, where the data type of the data to be recognized includes at least one of the following: text, audio, image, and video, the problem recognition model includes multiple problem recognition sub-models, the problem recognition sub-model is used to recognize the data to be recognized corresponding to the data type corresponding to the problem recognition sub-model, and the problem recognition result includes the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized; inputting the problem recognition result into a risk recognition sub-model to obtain a risk recognition result output by the risk recognition sub-model, where the risk recognition result includes the risk type corresponding to the problem recognition result, the risk value corresponding to the problem recognition result, and the risk value is used to indicate the severity of the risk; inputting the risk recognition result into a contingency plan generation model to obtain a risk contingency plan output by the contingency plan generation model, where the contingency plan generation model is a large language model.

[0006] Optionally, the method further includes: inputting the problem recognition result, the risk recognition result, and the risk contingency plan into an achievement evaluation model to obtain an evaluation result output by the achievement evaluation model, where the achievement evaluation model is a large language model and is used to evaluate the input problem recognition result, risk recognition result, and risk contingency plan according to preset evaluation conditions; obtaining an artificial feedback result, and adjusting the parameters of the problem recognition model, the risk recognition sub-model, and the contingency plan generation model according to the artificial feedback result and the evaluation result.

[0007] Optionally, input the data to be recognized into the problem recognition model, and obtain the problem recognition result set output by the problem recognition model, including: input each data to be recognized into the problem recognition sub-model corresponding to the data type according to the data type; obtain each problem recognition result output by each problem recognition sub-model; take the union of each problem recognition result as the problem recognition result set.

[0008] Optionally, when the data type of the data to be recognized is audio, the method further includes: converting the audio into text data; inputting the text data into the first problem recognition sub-model corresponding to the text data type; obtaining the first problem recognition result output by the first problem recognition sub-model; splicing the first problem recognition result with the text data as the problem recognition result output by the first problem recognition sub-model.

[0009] Optionally, when the data type of the data to be recognized is video, the method further includes: obtaining the image frames in the video; determining the similarity between each problem image in the database and the image frames according to the multi-modal pre-trained model, and sorting each problem image from large to small according to the similarity; obtaining the problems corresponding to the preset number of problem images ranked in the front as the similar problem set of the image frames; inputting the image frames into the second problem recognition sub-model corresponding to the picture data type; obtaining the second problem recognition result output by the second problem recognition sub-model, where the second problem recognition result includes the description information of the image frames; inputting the second problem recognition result into the large image recognition model to obtain the third problem recognition result output by the large image recognition model, where the third problem recognition result includes the semantic information of the image frames; adding the similar problem set and the third problem recognition result to the problem recognition result set.

[0010] Optionally, the risk recognition sub-model includes a classification sub-model and a regression sub-model; inputting the problem recognition result into the risk recognition sub-model, and obtaining the risk recognition result output by the risk recognition sub-model, including: inputting the problem recognition result into the classification sub-model to obtain the risk type corresponding to the problem recognition result output by the classification sub-model; inputting the problem recognition result into the regression sub-model to obtain the risk value corresponding to the problem recognition result output by the regression sub-model, where the risk value is used to indicate the severity of the risk.

[0011] Optionally, before inputting the risk recognition result into the pre-plan generation model, the method further includes: obtaining a preset prompt word corresponding to the data to be recognized, where the preset prompt word is used to indicate the generation requirements for the risk pre-plan; inputting the preset prompt word into the pre-plan generation model.

[0012] According to another aspect of the embodiments of the present application, a risk pre-plan generation device is further provided, including: a first recognition module, configured to input data to be recognized into a problem recognition model, and obtain a set of problem recognition results output by the problem recognition model, where the data types of the data to be recognized include at least one of the following: text, audio, image, and video, the problem recognition model includes multiple problem recognition sub-models, the problem recognition sub-model is configured to recognize the data to be recognized corresponding to the data type corresponding to the problem recognition sub-model, and the problem recognition result includes the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized; a second recognition module, configured to input the problem recognition result into a risk recognition sub-model, and obtain a risk recognition result output by the risk recognition sub-model, where the risk recognition result includes the risk type corresponding to the problem recognition result, the risk value corresponding to the problem recognition result, and the risk value is used to indicate the severity of the risk; a pre-plan generation module, configured to input the risk recognition result into a pre-plan generation model, and obtain a risk pre-plan output by the pre-plan generation model, where the pre-plan generation model is a large language model.

[0013] According to another aspect of the embodiments of the present application, a non-volatile storage medium is further provided. A program is stored in the non-volatile storage medium. When the program runs, it controls the device where the non-volatile storage medium is located to execute the risk pre-plan generation method.

[0014] According to another aspect of the embodiments of the present application, an electronic device is further provided, including: a memory and a processor, the processor is configured to run the program stored in the memory, where when the program runs, it executes the risk pre-plan generation method.

[0015] According to another aspect of the embodiments of the present application, a computer program product is further provided, including a computer program, and when the computer program is executed by a processor, it implements the risk pre-plan generation method.

[0016] In the embodiments of the present application, the data to be recognized is input into the problem recognition model to obtain the problem recognition result set output by the problem recognition model. Among them, the data types of the data to be recognized include at least one of the following: text, audio, image, and video. The problem recognition model includes multiple problem recognition sub-models, and the problem recognition sub-model is used to recognize the data to be recognized corresponding to the data type corresponding to the problem recognition sub-model. The problem recognition result includes the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized; the problem recognition result is input into the risk recognition sub-model to obtain the risk recognition result output by the risk recognition sub-model. Among them, the risk recognition result includes the risk type corresponding to the problem recognition result, the risk value corresponding to the problem recognition result, and the risk value is used to indicate the severity of the risk; the risk recognition result is input into the pre-plan generation model to obtain the risk pre-plan output by the pre-plan generation model. Among them, the pre-plan generation model is in the form of a large language model. By combining the large language model and the multi-modal small model for problem recognition, risk recognition and pre-plan generation, the purpose of efficiently and low-costly performing problem supervision and improvement in a multi-modal manner is achieved, thus realizing the technical effect of more accurate risk recognition and generating a pre-plan, and further solving the technical problems of high generation cost and low accuracy caused by the existing risk pre-plan generation technology only supporting single-modal data input and relying on large models. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0018] Figure 1 is a schematic structural diagram of a computer terminal provided according to an embodiment of the present application;

[0019] Figure 2 is a schematic flowchart of a risk pre-plan generation method provided according to an embodiment of the present application;

[0020] Figure 3 is a schematic diagram of the collaborative effect of a large model and a small model provided according to an embodiment of the present application;

[0021] Figure 4 is a schematic diagram of a model fine-tuning process provided according to an embodiment of the present application;

[0022] Figure 5 is a schematic structural diagram of a risk pre-plan generation device provided according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0023] To enable those skilled in the art to better understand the solution of this application, the following will clearly and completely describe the technical solution in the embodiments of this application with reference to the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of this application.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of this application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily need to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances, so that the embodiments of this application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily need to be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or are inherent to these processes, methods, products or devices.

[0025] To better understand the embodiments of this application, the following technical terms involved in the embodiments of this application are explained as follows:

[0026] Large Language Model (LLM): Refers to a complex algorithm model trained through deep learning technology that can understand and generate natural language text. It usually has a large number of parameters and can process and generate high-quality language data.

[0027] Small model: Refers to a machine learning model trained for a specific task or dataset. Its characteristics are relatively few parameters, low computational resource requirements, and can operate efficiently under limited hardware conditions. Small models usually perform well in specific fields or tasks, such as text classification, image recognition, etc. Due to their scale and complexity limitations, small models are more focused when processing tasks and can achieve high accuracy at a lower cost, especially suitable for resource-constrained environments or application scenarios that require quick responses. Compared with large language models, small models may be insufficient in terms of generality, but have obvious advantages in the efficiency and economy of specific tasks.

[0028] Multimodal data processing: Refers to methods and technologies for processing and analyzing data sources from multiple different types (such as text, images, videos, and audio).

[0029] Problem supervision and improvement model: Refers to an automated model used to discover problems, evaluate risks, and evaluate results, usually combining natural language processing and machine learning technologies.

[0030] Control Agent: An intelligent system based on large language models, responsible for controlling and coordinating the information flow and task execution among different modules in the system.

[0031] Data Flywheel: The data flywheel is a concept based on the continuous cycle and iteration of data. It is based on the following core idea: by continuously collecting data, training models, deploying models, collecting feedback and new data during model usage, and then using this new data to retrain the model to improve its performance, thus forming a closed-loop iterative process.

[0032] Text Clustering: It is an unsupervised learning method used to group text data according to content similarity to identify and generalize problem patterns.

[0033] Traditional problem supervision and improvement models and pre-plan generation models often only involve the processing of text information. However, as the saying goes, "A picture is worth a thousand words." There are many problems that cannot be represented by text, or require a large amount of text to represent, or are only reflected in images rather than text. Moreover, existing pre-plan generation models are mostly large language models, and both the deployment and training of large models require costs. If large models are used regardless of the task size, the cost will undoubtedly be huge and the cost performance will be extremely low. In specific vertical fields and for specific tasks, traditional small models can achieve comparable or even better performance at extremely low costs. However, the "high-order" capabilities such as planning and reasoning possessed by large models are indeed not available in small models.

[0034] To solve the above problems, relevant solutions are provided in the embodiments of this application, which are described in detail below.

[0035] According to the embodiments of this application, a method embodiment of a risk pre-plan generation method is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. And although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order than here.

[0036] The method embodiments provided by the embodiments of this application can be executed on mobile terminals, computer terminals, or similar computing devices. Figure 1 A hardware structure block diagram of a computer terminal for implementing the risk pre-plan generation method is shown. As Figure 1As shown, the computer terminal 10 may include one or more processors 102 (illustrated as 102a, 102b, ……, 102n in the figure) (the processor 102 may include, but is not limited to, processing devices such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may further include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply, and / or a camera. Those of ordinary skill in the art can understand that Figure 1 the structure shown is only schematic and does not limit the structure of the above-mentioned electronic device. For example, the computer terminal 10 may further include more or fewer components than Figure 1 shown in, or have a different configuration from Figure 1 that shown.

[0037] It should be noted that the above one or more processors 102 and / or other data processing circuits may generally be referred to as "data processing circuits" herein. The data processing circuit may be embodied in whole or in part as software, hardware, firmware, or any combination thereof. In addition, the data processing circuit may be a single independent processing module, or be incorporated in whole or in part into any one of the other elements in the computer terminal 10. As involved in the embodiments of the present application, the data processing circuit is a kind of processor control (such as the selection of a variable resistance terminal path connected to an interface).

[0038] The memory 104 may be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the risk pre-plan generation method in the embodiments of the present application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, implements the above-mentioned risk pre-plan generation method. The memory 104 may include a high-speed random access memory, and may further include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memories. In some instances, the memory 104 may further include a memory remotely set relative to the processor 102, and these remote memories may be connected to the computer terminal 10 through a network. Examples of the above network include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0039] The transmission device 106 is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a Radio Frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0040] The display can be, for example, a touch-screen liquid crystal display (LCD), which enables the user to interact with the user interface of the computer terminal 10.

[0041] Under the above operating environment, the embodiment of the present application provides a risk plan generation method, as Figure 2 shown, the method includes the following steps:

[0042] Step S202, input the data to be recognized into the problem recognition model, and obtain the problem recognition result set output by the problem recognition model. Among them, the data type of the data to be recognized includes at least one of the following: text, audio, image, and video. The problem recognition model includes multiple problem recognition sub-models, and the problem recognition sub-model is used to recognize the data to be recognized corresponding to the data type corresponding to the problem recognition sub-model. The problem recognition result includes the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized;

[0043] As an optional embodiment, the above-mentioned problem identification sub-models are all trained small models. By initializing a prompt (prompt word), a large language model is initialized as a management and control intelligent body to select and schedule the required multimodal small model. The large and small models are linked to solve the problem with the highest efficiency. By combining the advantages of the large language model (LLM) and the small model, and using multimodal data processing technology, comprehensive analysis and problem identification of various types of data are achieved. The efficiency of problem handling is improved through automated processes, manual intervention is reduced, and the high computing costs caused by relying on large models are reduced. In addition, by introducing multimodal data processing capabilities, the scope of problem identification is expanded, and potential problems can be found in various types of data such as text, images, videos and audio, thereby improving the accuracy and depth of identification. The problem discovery model (i.e., the problem identification model) is in the first ring of the embodiment, which is used to find problems (i.e., a set of problem identification results) from given information (i.e., data to be identified). The information found to have problems is further transmitted to the risk identification sub-model to determine the risks and levels of its existence, and output the corresponding risk management and control guidance plan (i.e., risk plan). Finally, the outcome evaluation model evaluates the judgment process and results of the above two models and gives feedback. The management and control agent is responsible for controlling the information interaction flow between modules at a higher dimension, intelligently selecting different models to handle different tasks and information, and regularly updating the entire system based on human feedback.

[0044] The problem discovery model performs relatively classic and easy understanding and classification tasks, so the use of small models (i.e., each problem identification sub-model) can meet the needs at a lower cost. In order to further expand the information scope of problem discovery, the small model is not limited to recognizing text information, but also supports recognizing multimodal information such as images, videos, and audio. The small model uses not only text models but also multimodal models, so it can extract problems from multiple modal information, expand the extraction scope, and improve recognition capabilities.

[0045] Optionally, before using the question recognition model, a multimodal question library is also constructed. Most of the traditional question libraries are composed of texts, that is, the question library contains many texts and questions corresponding to the texts. However, this single-modal question library is obviously limited in capacity, and the problems that can be represented are also limited. A picture is worth a thousand words. Often the amount of information contained in a picture requires a very lengthy text to describe, or it is difficult to describe using text. Therefore, if multimodal information such as images is not included in the question library, a large amount of information carriers containing problems will be lost, and effective extraction of many problems cannot be achieved. In this regard, the present embodiment constructs a multimodal database to expand the scope of problem recognition as much as possible and achieve more comprehensive and accurate problem extraction.

[0046] This question bank consists of four major components according to the modality: text, audio, images, and videos. The text modality is the most commonly used modality in the question bank. The text can be collected using traditional methods and the questions can be manually labeled. For the audio modality, there is no essential difference from the text as both are one-dimensional data. Audio can be converted into text using many open-source models with good performance for speech conversion. The data of the image modality is two-dimensional, quite different from text and audio, and contains more information. However, the annotation, training, and processing of images are often time-consuming and labor-intensive. For this, only a small number of images containing classic questions need to be manually sorted out and then annotated as the start of the data flywheel. Subsequently, new image data can be automatically annotated further through manual feedback, the understanding of multi-modal small models, and the reasoning of large language models to expand the image question bank. In addition, the image question bank also needs to save the CLIP (Contrastive Language-Image Pre-Training, a multi-modal pre-training model) feature representations corresponding to different images for subsequent retrieval by large models. Video is a high-dimensional extension of images in the time dimension and is three-dimensional data. However, the questions that can be extracted from videos often have little to do with time. Therefore, to simplify the processing flow of video data, it can be converted into a series of image frames for processing. After constructing the question bank, based on the established multi-modal question bank, intelligent identification and discovery of questions are carried out from the received multi-modal information.

[0047] In the technical solution provided in step S202, inputting the data to be recognized into the question recognition model and obtaining the set of question recognition results output by the question recognition model includes: inputting each data to be recognized into the question recognition sub-model corresponding to the data type according to the data type; obtaining each question recognition result output by each question recognition sub-model; and taking the union of each question recognition result as the set of question recognition results.

[0048] Optionally, since question recognition is a relatively basic classification task in the field of artificial intelligence and the performance of small models is already good enough, the question discovery model uses a series of multi-modal small models (each small model, that is, the question recognition sub-model processes data of one modality) and takes the union of the outputs of each small model as the set of question recognition results, achieving performance comparable to that of large models with quite small training, deployment, and inference costs.

[0049] Optionally, the process is as follows:

[0050] M = g(M modalities (x))

[0051] where M is the question recognition model, which combines multiple question recognition sub-models M modalitiesThe output, where g is the integration function of the model, i.e., taking the collection of each output, and x is the multi-modal input data.

[0052] As an optional embodiment, when the data type of the data to be recognized is audio, the method further includes: converting the audio into text data; inputting the text data into the first question recognition sub-model corresponding to the text data type; obtaining the first question recognition result output by the first question recognition sub-model; splicing the first question recognition result with the text data as the question recognition result output by the first question recognition sub-model.

[0053] Optionally, for text and audio (first converted to text data, and subsequent processing is the same as that of the text modality), the question recognition sub-model is directly trained according to the text question library. The text question library contains data pairs such as (text, question). Using supervised learning, the model takes the text as input and outputs question labels, and uses cross-entropy as the loss function. Optionally, it also supports fine-grained information extraction through some traditional natural language processing small models (such as named entity extraction, relationship extraction, etc.), and then splicing the extracted information with the original text to achieve data augmentation. Finally, question extraction is performed. This approach can significantly improve the accuracy of question recognition and enhance the model performance.

[0054] Optionally, for text and audio modalities, use supervised learning to train the question discovery sub-model M text :

[0055] Θ text = argmin Θ L CE (Θ, D text )

[0056] where Θ text represents the parameters of the question discovery sub-model, and these parameters need to be learned and optimized through training data. argmin Θ represents finding the parameter Θ that minimizes the subsequent loss function L CE . Here, Θ is a set of parameters, and L CE represents the cross-entropy loss function. L CE is the cross-entropy loss function, and D text is the text question library, which contains text data and its corresponding labels.

[0057] Optionally, use the trained first question discovery sub-model M text to extract fine-grained information:

[0058] z = M text (x text )

[0059] Among them, z is the first problem recognition result output by the first problem discovery sub-model.

[0060] Concatenate the extracted information (the first problem recognition result) with the original text x text to form enhanced data x t ′ ext , which is used as the problem recognition result:

[0061] x t ′ ext = [x text , z]

[0062] As an optional embodiment, when the data type of the data to be recognized is video, the method further includes: obtaining image frames in the video; determining the similarity between each problem image in the database and the image frames according to the multi-modal pre-trained model, and sorting each problem image from largest to smallest according to the similarity; obtaining the problems corresponding to the preset number of problem images ranked at the front as the similar problem set of the image frames; inputting the image frames into the second problem recognition sub-model corresponding to the picture data type; obtaining the second problem recognition result output by the second problem recognition sub-model, where the second problem recognition result includes the description information of the image frames; inputting the second problem recognition result into the image recognition large model to obtain the third problem recognition result output by the image recognition large model, where the third problem recognition result includes the semantic information of the image frames; adding the similar problem set and the third problem recognition result to the problem recognition result set.

[0063] Optionally, for modalities such as images and videos (first convert to image frames, and subsequent processing is the same as the text modality), the embodiments of the present application propose a multi-modal hierarchical problem recognition paradigm without training: first extract fine-grained information from the images. For example, use a variety of general multi-modal small models (i.e., the second problem recognition sub-model) (such as semantic segmentation small models for images, object detection small models for images, image annotation small models, OCR models, etc.) to extract fine-grained information. Then use a control agent as an intermediary to transfer the recognized fine-grained text information to the large language model. The large language model (i.e., the image recognition large model) first performs coarse-grained semantic recognition based on this information, and finally uses the reasoning ability of the large model and supplements it with relevant content retrieved from the CLIP model database to effectively extract fine-grained problems.

[0064] For the above retrieval, use the CLIP model to obtain the feature vector of the current image to be recognized (i.e., the image frame), and then calculate the similarity with the feature vectors corresponding to the images in the database to obtain the problems corresponding to the top k (i.e., the preset number) most similar images as the retrieval content (i.e., the similar problem set).

[0065] Optionally, for the image and video (converted to image frames) modalities, the process is as follows:

[0066] 1. Use the multi-modal small model M image to perform fine-grained information extraction:

[0067] z image = M image (x image )

[0068] 2. The control agent G passes the fine-grained information z image (the second problem recognition result) to the large image recognition model M large :

[0069] x context = G(z image )

[0070] where x context is the set of each second problem recognition result.

[0071] 3. The large image recognition model performs semantic recognition based on the passed information:

[0072]

[0073] where, is the third problem recognition result

[0074] 4. Use the CLIP model C to obtain the feature vector f of the image image , and calculate the similarity with the image features in the database:

[0075] S = C(f image ) = sim(f image , f database )

[0076] where f database is the feature vector of each image in the database.

[0077] 5. Obtain the top k most similar images' questions Q (i.e., the set of similar questions) according to the similarity S:

[0078] Q = retrieve(f database , S, k)

[0079] 6. Integrate the third problem recognition result and the set of similar questions to obtain the final image (video) modality problem recognition result and add it to the problem recognition result set:

[0080]

[0081] Over time and with the evolution of the information environment, new problems will constantly emerge. As an optional embodiment, to ensure the timeliness and relevance of the problem database, it is necessary to regularly apply text clustering algorithms to the collected problem texts. Text clustering is an unsupervised learning method that groups text data based on content similarity, which can help identify and generalize new problem patterns. The following is a detailed technical description of this process:

[0082] a Text preprocessing:

[0083] Before clustering, it is necessary to preprocess the problem texts, including removing stop words, punctuation marks, numbers, and performing stemming or lemmatization, etc.

[0084] b Feature extraction:

[0085] Use methods such as TF-IDF (Term Frequency-Inverse Document Frequency) to convert the text into a feature vector in numerical form for easy algorithm processing. The calculation of TF-IDF can be expressed as:

[0086]

[0087] where, t count is the number of occurrences of word t in document d, |D| is the total number of documents, and |{d ′ ∈D|t∈d ′}| is the number of documents containing word t.

[0088] c Clustering algorithm selection:

[0089] Select a suitable clustering algorithm, such as K-Means, hierarchical clustering, etc., according to the characteristics of the problems and the distribution of the data.

[0090] d Determine the number of clusters:

[0091] Determine the optimal number of clusters, that is, how many different problem categories there should be, through methods such as the elbow method and silhouette coefficient.

[0092] e Perform clustering:

[0093] Run the clustering algorithm to group the problem texts into different categories. Each category represents a group of similar problems. That is, assign the feature vector x i corresponding to the problem text to the nearest centroid c j :

[0094] c j =argmin c∈C ||x i -c||

[0095] where, C is the set of all centroids.

[0096] f Result analysis:

[0097] Analyze the clustering results to identify newly emerging problem categories and those that become irrelevant over time.

[0098] g Problem library update:

[0099] Update the problem library according to the clustering results, add new problem categories, delete obsolete problems, and possibly reclassify existing problems.

[0100] After updating the problem library, it also includes fine-tuning the text problem recognition small model based on the updated problem library.

[0101] In step S204, input the problem recognition result into the risk recognition sub-model to obtain the risk recognition result output by the risk recognition sub-model. Among them, the risk recognition result includes the risk type corresponding to the problem recognition result, the risk value corresponding to the problem recognition result, and the risk value is used to indicate the severity of the risk.

[0102] In the technical solution provided in step S204, the risk recognition sub-model includes a classification sub-model and a regression sub-model; inputting the problem recognition result into the risk recognition sub-model to obtain the risk recognition result output by the risk recognition sub-model includes: inputting the problem recognition result into the classification sub-model to obtain the risk type corresponding to the problem recognition result output by the classification sub-model; inputting the problem recognition result into the regression sub-model to obtain the risk value corresponding to the problem recognition result output by the regression sub-model.

[0103] Optionally, training a model based on the problem library for risk recognition is a classic classification or regression task, so a small model can be used here. Risk recognition needs to be based on the above problem discovery model, and there is an embedding relationship between the two, because where there is a risk, there is a problem, and the more serious the problem, the higher the risk. The risk value is a fine-grained numerical representation of the problem. Based on the existing problem database, manually classify the risk categories and calibrate the risk values for different problems, and then a classification small model (i.e., the classification sub-model) and a regression small model (i.e., the regression sub-model) can be trained using all the (problem, risk category), (problem, risk value) pairs. Furthermore, a complete pipeline from problem recognition to risk recognition and risk value prediction can be achieved. Since the problem recognition model has already recognized the multi-modal data to be recognized as single-modal (text) modal data, the risk recognition sub-model can use a single-modal small model, which reduces the computational pressure.

[0104] In step S206, input the risk recognition result into the pre-plan generation model to obtain the risk pre-plan output by the pre-plan generation model, where the pre-plan generation model is a large language model.

[0105] Optionally, generating a risk control guidance plan (i.e., risk plan) is relatively open and a generative task. Therefore, using a large language model (i.e., plan generation model) can better complete this task.

[0106] In the technical solution provided in step S206, before inputting the risk identification result into the plan generation model, the method further includes: obtaining a preset prompt word corresponding to the data to be identified, where the preset prompt word is used to indicate the generation requirements for the risk plan; inputting the preset prompt word into the plan generation model.

[0107] Optionally, the plan generation model supports formulating personalized solutions according to different core businesses and combining the general knowledge of the large model itself. This is also the advantage of the large model: First, different personalized Prompts are initialized according to different core businesses, so that the plan generation model can judge based on different criteria and generate risk plans. Second, the plan generation model uses the text or multimodal information with risks as the Context (context), and then gives different risk control guidance plans based on different core businesses.

[0108] Optionally, the method further includes: inputting the problem identification result, risk identification result, and risk plan into the result evaluation model, and obtaining the evaluation result output by the result evaluation model, where the result evaluation model is a large language model for evaluating the input problem identification result, risk identification result, and risk plan according to preset evaluation conditions; obtaining the manual feedback result, and adjusting the parameters of the problem identification model, risk identification sub-model, and plan generation model according to the manual feedback result and the evaluation result.

[0109] Optionally, it is necessary to conduct manual evaluation and feedback on the plan generated by the above result evaluation model. In addition, historical risk events and corresponding plans, combined with corresponding manual feedback, can be stored in the database. Thereafter, when the large model encounters similar problems, it can retrieve and recall for RAG (retrieval-augmented generation) to achieve more accurate and faster plan generation.

[0110] Optionally, since the result evaluation task is a generative task with a large degree of subjectivity, and it needs to adapt to different core construction goals involving various different consideration factors and combine manual preferences, the requirements are relatively complex and changeable, and need to be updated in a timely manner. Therefore, it is more suitable to use a large language model (i.e., plan generation model) here. For the continuously adjusted core construction goals, the large language model (control agent) adjusts the personalized setting Prompt and feedback knowledge base based on its own capabilities and manual feedback to continuously adapt, so as to conduct a more objective, fair, and personalized evaluation. Specifically as follows:

[0111] Update(Prompt, Knowledge Base) ← Feedback(Core Construction Goals, Human Preferences)

[0112] That is, the achievement evaluation model conducts evaluation and feedback based on preset evaluation conditions (core construction goals and human preferences), outputs the evaluation results, and the control agent updates the database and prompt words according to the evaluation results and human feedback.

[0113] Optionally, even though the cooperation between large and small models has high operating efficiency and low cost, it is obviously far from enough to inject human preferences into the entire system only by manually annotating data. Therefore, it is necessary to regularly provide manual feedback to all models, that is, it is necessary for humans to verify and provide feedback on the evaluation given by the output risk control guidance plan and the achievement evaluation model, and output the feedback results to the control agent. The control agent supports optimizing its own scheduling decisions based on these feedbacks, and fine-tuning, replacing, etc. specific small models at appropriate times. In addition, it supports humans to moderately modify the above results and then store them in the database for subsequent fine-tuning of small models or for the retrieval of large models.

[0114] Optionally, the information flow and interaction between all the large language models and multi-modal small models involved above do not require manual participation. A large language model is introduced as the control agent to achieve reasonable distribution of information, so as to achieve the functions of full automation and reasonable routing. And the control agent can update the database according to human feedback, thereby improving the accuracy of the problem discovery model and the risk identification model.

[0115] Optionally, the following functions of the control agent are realized by setting Prompt:

[0116] 1. Responsible for controlling the flow of information among the large models, text small models, and multi-modal small models involved in the system. Among them, the cooperation between large models and small models is as Figure 3 shown: The large model is responsible for routing and scheduling the information flow (control agent) and performing reasoning and generation, and the small model is responsible for simple tasks such as understanding, identifying, and classifying single-modal data.

[0117] The agent can choose to send multi-modal information to a specified model, and input the output of the model into a specified model or directly as the final output.

[0118] 2. Evaluate the results of recognition, understanding, generation, etc. of large and small models, and give appropriate feedback for subsequent optimization.

[0119] 3. Give fine-tuning instructions to achieve full-automatic fine-tuning and update of small models.

[0120] Specifically, the information flow process in the entire system is as follows. Let the large model be Mbig , the text small model is M text , the multi-modal small model is M multi , the input information is I, and the output information is O. The set Prompt of the control agent is P, and the decision function for information flow is:

[0121]

[0122] Assume that each model has a processing function, such as f big , f text , f multi , corresponding to the processing capabilities of the large model, text small model, and multi-modal small model respectively, then:

[0123] f model (I) = O

[0124] According to the result of the decision function , the information I is sent to the corresponding model for processing, and then the result O is output:

[0125]

[0126] If the information needs to be processed by multiple models, it can be expressed as a sequence of function applications:

[0127]

[0128]

[0129] Among them, M 1 , M 2 , …, M n is the sequence of models selected according to the decision function .

[0130] Optionally, since the question bank, core business, and core construction goals are constantly changing and dynamically updated, if the initially trained text question recognition small model remains unchanged, it is obvious that it cannot complete those OOD (Out of Distribution) tasks. Therefore, regular fine-tuning is required. However, if all parameters are fine-tuned, first, it consumes a lot of resources and is not cost-effective. Second, full-parameter fine-tuning will also bring problems such as knowledge forgetting. For this, the generally considered measures are Continue Pretrain or Lora fine-tuning. This embodiment adopts Lora fine-tuning. Lora (Low-Rank Adaptation) fine-tuning is a fine-tuning technology for large pre-trained models. It adjusts the weights of the model by introducing a low-rank structure instead of directly modifying the parameters of the entire pre-trained model. This method aims to reduce the consumption of computing resources and memory occupancy while maintaining or improving the performance of the model.

[0131] Figure 4 shows a fine-tuning process, as Figure 4 shown. The large language model automatically fine-tunes the problem discovery model and the risk identification (sub) model, and gives feedback on the identification results and risk contingency plans output by the model. At the same time, the text database is updated regularly. After a certain period of time or after receiving a fine-tuning instruction from the control agent or manually, the small model is fine-tuned using the updated new data, and then merged into the original model. However, due to the limited number of Lora parameters, the stored knowledge content is also limited. Therefore, it is necessary to set a fine-tuning times threshold T. Whenever the number of times of fine-tuning exceeds T, the original basic model is retrained on all the latest data, so as to better absorb and integrate all the knowledge.

[0132] This patent also utilizes the ability of the control agent to write code to automate the fine-tuning process. After the control agent is connected to the network and the configuration environment is initialized, the control agent for the subsequent updated function package can download it according to the error message, then update and start the training script, and automatically run the evaluation code to intelligently terminate the training process. Finally, the control agent also supports self-deployment to replace the original small model.

[0133] Through the above steps, a method including a multi-modal problem library can be realized. The model of the present application can be deployed in an enterprise environment, including a problem discovery model, a risk identification model, a result evaluation model, and a control agent, enhancing the adaptability and flexibility to a complex data environment. The control agent in the method embodiment can be updated and optimized regularly to ensure that the contingency plan generation model can adapt to new data patterns and business requirements. Through this method, risks can be evaluated and results can be evaluated more effectively, providing more objective and reliable decision-making support for users. Finally, through continuous learning and optimization, the satisfaction of users with the output risk contingency plan can be improved, ensuring the efficiency and effectiveness of the problem supervision and rectification process, realizing the automatic generation of risk contingency plans. By reasonably combining the large language model and the small model, the dependence on large models is reduced, and the computing cost is lowered. Introducing multi-modal data processing enhances the processing ability of various information types such as text, images, videos, and audio, expanding the scope of problem identification. An all-automated process for problem discovery, risk assessment, and result evaluation is realized, reducing manual intervention and improving efficiency. Using text clustering and multi-modal information extraction improves the accuracy of problem identification. A regular fine-tuning and automatic update mechanism is designed to ensure the continuous learning and optimization of the model. Specifically, the method embodiment of the present application has the following advantages:

[0134] 1. Construction of a multi-modal problem library: Collect and annotate data including text, audio, images, and videos to construct a multi-modal problem library, enhancing the processing ability of various information types such as text, images, videos, and audio, and expanding the scope of problem identification.

[0135] 2. Model Training and Fine-tuning: Use the data in the question library to train the question discovery model, train the risk identification model according to the risk assessment requirements, and use human feedback to train the achievement evaluation model, ensuring the continuous learning and optimization of the model.

[0136] 3. Risk Pre-plan Generation and Evaluation, Achievement Evaluation and Feedback: Use the large language model to generate a risk control guidance pre-plan, and optimize it through manual evaluation and feedback. Apply the achievement evaluation model to evaluate the project achievements, and collect manual feedback for model optimization to ensure the real-time effectiveness of the risk pre-plan output by the model.

[0137] 4. Automated Fine-tuning Process: The control agent automatically performs Lora fine-tuning or full-parameter fine-tuning of the small model according to preset conditions or manual instructions.

[0138] 5. System Update and Maintenance: The control agent is responsible for the daily update and maintenance of the system, including function package download, training script startup, evaluation code operation, and model deployment, reducing manual intervention and improving efficiency.

[0139] The embodiment of the present application provides a risk pre-plan generation device. Figure 5 It is a schematic structural diagram of the device, as Figure 5 shown. The device includes: a first identification module 50, configured to input the data to be identified into the question identification model, and obtain a set of question identification results output by the question identification model. Among them, the data type of the data to be identified includes at least one of the following: text, audio, image, and video. The question identification model includes multiple question identification sub-models, and the question identification sub-model is used to identify the data to be identified of the data type corresponding to the question identification sub-model. The question identification result includes the question type corresponding to the data to be identified and the question description corresponding to the data to be identified; a second identification module 52, configured to input the question identification result into the risk identification sub-model, and obtain a risk identification result output by the risk identification sub-model. Among them, the risk identification result includes the risk type corresponding to the question identification result, the risk value corresponding to the question identification result, and the risk value is used to indicate the severity of the risk; a pre-plan generation module 54, configured to input the risk identification result into the pre-plan generation model, and obtain a risk pre-plan output by the pre-plan generation model. Among them, the pre-plan generation model is a large language model.

[0140] In some embodiments of the present application, it further includes: inputting the problem recognition result, risk recognition result, and risk plan into the achievement evaluation model to obtain the evaluation result output by the achievement evaluation model, where the achievement evaluation model is a large language model for evaluating the input problem recognition result, risk recognition result, and risk plan according to preset evaluation conditions; obtaining the manual feedback result, and adjusting the parameters of the problem recognition model, risk recognition sub-model, and plan generation model according to the manual feedback result and the evaluation result.

[0141] In some embodiments of the present application, the first recognition module 50 inputs the data to be recognized into the problem recognition model, and the problem recognition result set output by the problem recognition model includes: inputting each data to be recognized into the problem recognition sub-model corresponding to the data type according to the data type; obtaining each problem recognition result output by each problem recognition sub-model; taking the union of each problem recognition result as the problem recognition result set.

[0142] In some embodiments of the present application, when the data type of the data to be recognized is audio, it further includes: converting the audio into text data; inputting the text data into the first problem recognition sub-model corresponding to the text data type; obtaining the first problem recognition result output by the first problem recognition sub-model; splicing the first problem recognition result with the text data as the problem recognition result output by the first problem recognition sub-model.

[0143] In some embodiments of the present application, when the data type of the data to be recognized is video, it further includes: obtaining the image frames in the video; determining the similarity between each problem image in the database and the image frames according to the multi-modal pre-trained model, and sorting each problem image from large to small according to the similarity; obtaining the problems corresponding to the preset number of problem images ranked at the top as the similar problem set of the image frames; inputting the image frames into the second problem recognition sub-model corresponding to the picture data type; obtaining the second problem recognition result output by the second problem recognition sub-model, where the second problem recognition result includes the description information of the image frames; inputting the second problem recognition result into the large image recognition model to obtain the third problem recognition result output by the large image recognition model, where the third problem recognition result includes the semantic information of the image frames; adding the similar problem set and the third problem recognition result to the problem recognition result set.

[0144] In some embodiments of the present application, the risk identification sub-model includes a classification sub-model and a regression sub-model; inputting the problem identification result into the risk identification sub-model to obtain the risk identification result output by the risk identification sub-model includes: inputting the problem identification result into the classification sub-model to obtain the risk type corresponding to the problem identification result output by the classification sub-model; inputting the problem identification result into the regression sub-model to obtain the risk value corresponding to the problem identification result output by the regression sub-model, where the risk value is used to indicate the severity of the risk.

[0145] In some embodiments of the present application, before inputting the risk identification result into the pre-plan generation model, it further includes: obtaining a preset prompt word corresponding to the data to be identified, where the preset prompt word is used to indicate the generation requirements for the risk pre-plan; inputting the preset prompt word into the pre-plan generation model.

[0146] It should be noted that each module in the above risk pre-plan generation device can be a program module (for example, a set of program instructions that implement a specific function), or a hardware module. For the latter, it can be presented in the following forms, but not limited to: the manifestation form of each of the above modules is a processor, or the functions of each of the above modules are implemented by a processor.

[0147] The embodiments of the present application provide a non-volatile storage medium, in which a program is stored. When the program runs, it controls the device where the non-volatile storage medium is located to execute the following risk pre-plan generation method: inputting the data to be identified into the problem identification model to obtain a set of problem identification results output by the problem identification model, where the data type of the data to be identified includes at least one of the following: text, audio, image, and video. The problem identification model includes multiple problem identification sub-models, and the problem identification sub-model is used to identify the data to be identified of the data type corresponding to the problem identification sub-model. The problem identification result includes the problem type corresponding to the data to be identified and the problem description corresponding to the data to be identified; inputting the problem identification result into the risk identification sub-model to obtain the risk identification result output by the risk identification sub-model, where the risk identification result includes the risk type corresponding to the problem identification result, the risk value corresponding to the problem identification result, and the risk value is used to indicate the severity of the risk; inputting the risk identification result into the pre-plan generation model to obtain the risk pre-plan output by the pre-plan generation model, where the pre-plan generation model is a large language model.

[0148] An embodiment of the present application provides an electronic device, including: a memory and a processor, where the processor is configured to run a program stored in the memory. When the program runs, it executes the following risk plan generation method: input the data to be recognized into a problem recognition model, and obtain a set of problem recognition results output by the problem recognition model. The data type of the data to be recognized includes at least one of the following: text, audio, image, and video. The problem recognition model includes multiple problem recognition sub-models, and the problem recognition sub-model is used to recognize the data to be recognized corresponding to the data type corresponding to the problem recognition sub-model. The problem recognition result includes the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized; input the problem recognition result into a risk recognition sub-model, and obtain a risk recognition result output by the risk recognition sub-model. The risk recognition result includes the risk type corresponding to the problem recognition result, the risk value corresponding to the problem recognition result, and the risk value is used to indicate the severity of the risk; input the risk recognition result into a plan generation model, and obtain a risk plan output by the plan generation model, where the plan generation model is a large language model.

[0149] An embodiment of the present application provides a computer program product, including a computer program, which when executed by a processor, implements the following risk plan generation method: input the data to be recognized into a problem recognition model, and obtain a set of problem recognition results output by the problem recognition model. The data type of the data to be recognized includes at least one of the following: text, audio, image, and video. The problem recognition model includes multiple problem recognition sub-models, and the problem recognition sub-model is used to recognize the data to be recognized corresponding to the data type corresponding to the problem recognition sub-model. The problem recognition result includes the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized; input the problem recognition result into a risk recognition sub-model, and obtain a risk recognition result output by the risk recognition sub-model. The risk recognition result includes the risk type corresponding to the problem recognition result, the risk value corresponding to the problem recognition result, and the risk value is used to indicate the severity of the risk; input the risk recognition result into a plan generation model, and obtain a risk plan output by the plan generation model, where the plan generation model is a large language model.

[0150] In the above embodiments of the present application, the descriptions of the respective embodiments have their own emphases. For the parts not detailed in a certain embodiment, reference may be made to the relevant descriptions of other embodiments.

[0151] In several embodiments provided in the present application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are merely illustrative. For example, the division of the units can be a logical function division. In actual implementation, there can be other division methods. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the couplings or direct couplings or communication connections shown or discussed with each other can be through some interfaces. The indirect couplings or communication connections of units or modules can be in electrical or other forms.

[0152] The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they can be located in one place or distributed to multiple units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0153] In addition, in each embodiment of the present application, the functional units can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0154] If the above-mentioned integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present application, in essence, or the part that contributes to the related technology, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present application. The foregoing storage medium includes: various media such as USB flash drives, read-only memories (ROMs), random access memories (RAMs), mobile hard disks, magnetic disks, or optical discs that can store program codes.

[0155] The above are only the preferred embodiments of the present application. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present application, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present application.

Claims

1. A risk plan generation method, characterized in that: include: Input the data to be identified into the problem identification model, and obtain a set of problem identification results output by the problem identification model, wherein the data type of the data to be identified includes at least one of the following: text, audio, image, and video, the problem identification model includes a plurality of problem identification sub-models, and the problem identification sub-models are used to identify the data to be identified of the data type corresponding to the problem identification sub-models, and the problem identification results include the problem type corresponding to the data to be identified and the problem description corresponding to the data to be identified; Inputting the problem identification result into the risk identification sub-model to obtain the risk identification result output by the risk identification sub-model, wherein the risk identification result includes the risk type corresponding to the problem identification result and the risk value corresponding to the problem identification result, and the risk value is used to indicate the severity of the risk; The risk identification result is input into a plan generation model to obtain a risk plan output by the plan generation model, wherein the plan generation model is a large language model.

2. The risk plan generation method according to claim 1, characterized in that: The method further comprises: Inputting the problem identification result, the risk identification result, and the risk plan into an achievement evaluation model to obtain an evaluation result output by the achievement evaluation model, wherein the achievement evaluation model is a large language model, which is used to evaluate the input problem identification result, the risk identification result, and the risk plan according to preset evaluation conditions; Obtain manual feedback results, and adjust the parameters of the problem identification model, the risk identification sub-model, and the plan generation model based on the manual feedback results and the evaluation results.

3. The risk plan generation method according to claim 1, characterized in that: Inputting the to-be-recognized data into a problem recognition model, and obtaining a problem recognition result set output by the problem recognition model comprises: Input each of the data to be identified into a problem identification sub-model corresponding to the data type according to the data type; Obtaining each problem identification result output by each of the problem identification sub-models; A collection of the problem identification results is taken as the problem identification result set.

4. The risk plan generation method according to claim 3 is characterized in that: In the case where the data type of the to-be-identified data is audio, the method further includes: Convert audio to text data; Inputting the text data into a first question recognition sub-model corresponding to the text data type; Obtaining a first question recognition result output by the first question recognition sub-model; The first question recognition result is concatenated with the text data as the question recognition result output by the first question recognition sub-model.

5. The risk plan generation method according to claim 3 is characterized in that: In the case where the data type of the to-be-identified data is video, the method further includes: Acquire image frames in the video; Determine the similarity between each problem image in the database and the image frame according to the multimodal pre-training model, and sort the problem images from large to small according to the similarity; Obtaining questions corresponding to a preset number of the question images that are ranked first as a similar question set of the image frame; Inputting the image frame into a second question recognition sub-model corresponding to the image data type; Obtaining a second question recognition result output by the second question recognition sub-model, wherein the second question recognition result includes description information of the image frame; Inputting the second question recognition result into the image recognition large model, and obtaining a third question recognition result output by the image recognition large model, wherein the third question recognition result includes semantic information of the image frame; The similar question set and the third question recognition result are added to the question recognition result set.

6. The risk plan generation method according to claim 1, characterized in that: The risk identification sub-model includes a classification sub-model and a regression sub-model; Inputting the problem identification result into the risk identification sub-model, and obtaining the risk identification result output by the risk identification sub-model includes: Inputting the problem identification result into the classification sub-model, and obtaining the risk type corresponding to the problem identification result output by the classification sub-model; The problem identification result is input into the regression sub-model, and a risk value corresponding to the problem identification result output by the regression sub-model is obtained, wherein the risk value is used to indicate the severity of the risk.

7. The risk plan generation method according to claim 1, characterized in that: Before inputting the risk identification result into the emergency plan generation model, the method further includes: Acquire a preset prompt word corresponding to the data to be identified, wherein the preset prompt word is used to indicate a requirement for generating the risk plan; The preset prompt words are input into the plan generation model.

8. A risk plan generation device, characterized in that: include: A first recognition module is used to input the data to be recognized into the problem recognition model, and obtain a set of problem recognition results output by the problem recognition model, wherein the data type of the data to be recognized includes at least one of the following: text, audio, image and video, the problem recognition model includes a plurality of problem recognition sub-models, the problem recognition sub-models are used to recognize the data to be recognized of the data type corresponding to the problem recognition sub-models, and the problem recognition results include the problem type corresponding to the data to be recognized and the problem description corresponding to the data to be recognized; A second identification module is used to input the problem identification result into the risk identification sub-model to obtain the risk identification result output by the risk identification sub-model, wherein the risk identification result includes the risk type corresponding to the problem identification result and the risk value corresponding to the problem identification result, and the risk value is used to indicate the severity of the risk; The contingency plan generation module is used to input the risk identification result into the contingency plan generation model to obtain the risk contingency plan output by the contingency plan generation model, wherein the contingency plan generation model is a large language model.

9. A non-volatile storage medium, characterized in that: The non-volatile storage medium stores a program, wherein when the program is executed, the device where the non-volatile storage medium is located is controlled to execute the risk plan generation method described in any one of claims 1 to 7.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program, when running, executes the risk plan generation method described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the risk plan generation method according to any one of claims 1 to 7 is implemented.

Images