Large-model-based continuous fusion method and device for spatio-temporal data, and medium

Abstract

The present application relates to the technical field of artificial intelligence. Disclosed are a large-model-based continuous fusion method and device for spatio-temporal data, and a medium. The method comprises: on the basis of a preset data fusion requirement, determining parameters to be fine-tuned that are in a large model to be applied, and fine tuning said parameters by means of a large-model fine tuning algorithm, so as to obtain a large model to be trained; acquiring a training corpus, and on the basis of the training corpus, training the large model to be trained, so as to obtain an applied large model; when an actual question needs to be processed, performing prompt design on the basis of question information input by a user; and on the basis of a design result corresponding to the prompt design, performing question answering by means of the applied large model, so as to determine a corresponding actual answer to the question. By means of the method, the present application realizes efficient and accurate data processing and fusion when faced with massive amounts of spatio-temporal data.

Description

A large model-based cross-temporal and spatial data continuous fusion method, device and medium

[0001] The present application claims priority to the Chinese patent application No. 202410627222.X, filed on May 21, 2024, and entitled "A large model-based cross-temporal and spatial data continuous fusion method, device and medium", the whole content of which is incorporated herein by reference. TECHNICAL FIELD

[0002] The present application relates to the field of artificial intelligence, in particular to a large model-based cross-temporal and spatial data continuous fusion method, device and medium. BACKGROUND

[0003] With the advent of the big data era, the generation and accumulation of data have shown explosive growth. Cross-temporal and spatial data are of great concern because they can reflect the state changes of things at different times and spaces. Such data have wide application value in weather forecasting, traffic management, financial prediction and other fields. However, in the face of massive cross-temporal and spatial data, how to efficiently and accurately process and fuse these data to extract valuable information and provide more accurate and valuable decision support for related fields has become a major problem faced by current technology.

[0004] Firstly, the quality and integrity of data have important influence on the processing results. Due to the influence of data acquisition equipment, transmission network and other factors, there are often missing, abnormal and other problems in cross-temporal and spatial data, which need to be effectively cleaned and preprocessed. Secondly, the fusion of cross-temporal and spatial data needs to consider the correlation in time and space dimensions. How to effectively fuse data of different sources, different times and spatial resolutions to extract valuable information is a challenging problem. In addition, with the expansion of data scale and the increase of complexity, traditional data processing methods have been difficult to meet the requirements of real-time and accuracy. Therefore, in the face of massive cross-temporal and spatial data, how to efficiently and accurately process and fuse has become a technical problem to be solved.

[0005] SUMMARY

[0006] The embodiments of the present application provide a large model-based cross-temporal and spatial data continuous fusion method, device and medium, to solve the technical problem of how to efficiently and accurately process and fuse in the face of massive cross-temporal and spatial data.

[0007] In a first aspect, the embodiments of the present application provide a cross-temporal and spatial data continuous fusion method based on a large model, characterized in that the method comprises: determining to-be-tuned parameters in a to-be-applied large model based on a preset data fusion requirement, and tuning the to-be-tuned parameters through a large model tuning algorithm to obtain a to-be-trained large model; obtaining training corpus, and training the to-be-trained large model based on the training corpus to obtain an applied large model; in the case of needing to process an actual problem, designing a prompt based on problem information input by a user; and determining a reply to the actual problem by performing question and answer through the applied large model based on a design result corresponding to the prompt design.

[0008] In an implementation manner of the present application, the to-be-tuned parameters in the to-be-applied large model are determined based on the preset data fusion requirement, specifically comprising: analyzing to-be-required data features according to the data fusion requirement; matching corresponding to-be-tuned parameters in the to-be-applied large model based on the to-be-required data features, and determining tuning targets of the to-be-tuned parameters.

[0009] In an implementation manner of the present application, the to-be-tuned parameters are tuned through the large model tuning algorithm to obtain the to-be-trained large model, specifically comprising: determining a tuning target matrix corresponding to each to-be-tuned parameter based on the tuning target of each to-be-tuned parameter; decomposing the tuning target matrix to obtain an initial dimension reduction matrix and an initial dimension increase matrix; initializing the initial dimension increase matrix through a Gaussian distribution to obtain an applied dimension increase matrix, and initializing the initial dimension reduction matrix through a 0 matrix to obtain an applied dimension reduction matrix; constructing a tuning application matrix based on the applied dimension increase matrix and the applied dimension reduction matrix, and adding a corresponding to-be-applied branch to the to-be-applied large model based on the tuning application matrix to obtain the to-be-trained large model.

[0010] In an implementation manner of the present application, the to-be-trained large model is trained based on the training corpus to obtain the applied large model, specifically comprising: training the to-be-applied large model in the to-be-trained large model based on the training corpus to obtain initial applied parameters converging; training the to-be-applied branch in the to-be-trained large model based on the training corpus to obtain branch applied parameters converging; superimposing the initial applied parameters and the branch applied parameters to obtain applied parameters; and updating the to-be-applied large model based on the applied parameters to obtain the applied large model.

[0011] In an implementation manner of the present application, the prompt is designed based on the problem information input by the user, specifically comprising: analyzing the problem information to decompose the problem information into a plurality of sub-questions; determining a plurality of sub-replies by performing question and answer through the applied large model based on the plurality of sub-questions; and integrating the plurality of sub-questions and the plurality of sub-replies.

[0012] In an implementation form of the present application, after the question and answer are performed by the application large model to determine the corresponding actual problem reply, the method further comprises: based on the preset Bradley-Terry preference model and the reward function, optimizing the application large model, specifically comprising: calculating the preference score of the reply according to the quality of the actual problem reply through the Bradley-Terry preference model; calculating the reward value through the reward function based on the preference score; using the policy gradient method to update and optimize the parameters of the application large model according to the reward value.

[0013] In an implementation form of the present application, after the training corpus is obtained, the method further comprises: cleaning the training corpus to remove irrelevant information, duplicate data and outliers; labeling the cleaned training corpus; wherein the labeling comprises semantic role labeling and named entity recognition; performing word segmentation processing on the labeled training corpus to realize word segmentation, stem extraction and stop word removal.

[0014] In an implementation form of the present application, the method further comprises: after the user inputs the question information, performing security check and filtering on the question information; and after determining the corresponding actual problem reply, monitoring and auditing the actual problem reply; when an abnormal situation is found, processing based on the preset abnormal handling logic.

[0015] In a second aspect, the embodiments of the present application also provide a cross-time and space data continuous fusion device based on a large model, characterized in that the device comprises: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: based on a preset data fusion requirement, determine to-be-tuned parameters in the application large model, and tune the to-be-tuned parameters through a large model tuning algorithm to obtain a to-be-trained large model; obtain training corpus, and train the to-be-trained large model based on the training corpus to obtain the application large model; in the case of needing to process an actual problem, perform prompt design based on user input question information; through the application large model, perform question and answer based on the prompt design corresponding design result to determine the corresponding actual problem reply.

[0016] In a third aspect, the embodiments of the present application also provide a non-volatile computer storage medium based on continuous fusion of cross-temporal and spatial data of a large model, which stores computer executable instructions, and the computer executable instructions are configured to: determine to-be-fine-tuned parameters in the large model based on a preset data fusion requirement, fine-tune the to-be-fine-tuned parameters by using a large model fine-tuning algorithm, and obtain a to-be-trained large model; obtain training corpus, and train the to-be-trained large model based on the training corpus to obtain an application large model; in a case where an actual problem needs to be processed, perform prompt design based on problem information input by a user; and perform question and answer by using the application large model based on a design result corresponding to the prompt design, to determine a reply to the actual problem.

[0017] The method, device and medium for continuous fusion of cross-temporal and spatial data based on a large model provided by the embodiments of the present application have the following beneficial effects:

[0018] 1. The large model is fine-tuned based on the preset data fusion requirement, so that the method can flexibly adapt to different data fusion scenarios, thereby improving the flexibility of the model and enabling the model to better adapt to specific data characteristics, and thus improving the accuracy and efficiency of data processing.

[0019] 2. The to-be-trained large model after fine-tuning is trained to obtain a more accurate and efficient application large model, and this training method can fully utilize the learning ability of the large model to extract more useful information from the training corpus, thereby improving the performance of the model in actual application.

[0020] 3. The prompt design is performed based on the problem information input by the user, and the question and answer are performed by using the application large model, to provide the user with intuitive and easy-to-understand replies, thereby improving the user experience and reducing the difficulty of the user in obtaining information.

[0021] 4. The application large model is optimized by using the preset Bradley-Terry preference model and reward function, to continuously improve the reply quality and accuracy of the model, and this optimization strategy enables the model to continuously learn and progress in actual application, and better meets the user demand.

[0022] 5. After the problem information input by the user and the reply to the actual problem are determined, safety and reliability checks are performed, including safety checks and filtering on the problem information, monitoring and auditing on the reply, and timely processing when an abnormal situation is found, to ensure the safety and reliability of the model in actual application. BRIEF DESCRIPTION OF DRAWINGS

[0023] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0024] Fig. 1 is a flowchart of a large model-based cross-temporal and cross-spatial data continuous fusion method according to an embodiment of the application;

[0025] Fig. 2 is a schematic diagram of the internal structure of a large model-based cross-temporal and cross-spatial data continuous fusion device according to an embodiment of the application. DETAILED DESCRIPTION

[0026] In order to make the objectives, technical solutions, and advantages of the application clearer, the following will describe the technical solutions of the application in conjunction with specific embodiments of the application and corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the application.

[0027] The embodiments of the application provide a large model-based cross-temporal and cross-spatial data continuous fusion method, device, and medium, to solve the technical problem of how to efficiently and accurately process and fuse massive cross-temporal and cross-spatial data.

[0028] The technical solutions of the embodiments of the application will be described in detail below with reference to the drawings.

[0029] Fig. 1 is a flowchart of a large model-based cross-temporal and cross-spatial data continuous fusion method according to an embodiment of the application. As shown in Fig. 1, the large model-based cross-temporal and cross-spatial data continuous fusion method according to an embodiment of the application specifically includes the following steps:

[0030] Step 101, based on a preset data fusion requirement, determine the to-be-tuned parameters in the to-be-applied large model, and tune the to-be-tuned parameters by a large model tuning algorithm to obtain a to-be-trained large model.

[0031] In an embodiment of the application, to implement the large model-based cross-temporal and cross-spatial data continuous fusion, first, based on a preset data fusion requirement, the to-be-tuned parameters in the to-be-applied large model need to be determined.

[0032] Specifically, according to the data fusion requirement, analyze the to-be-required data features; based on the to-be-required data features, match the corresponding to-be-tuned parameters in the to-be-applied large model, and determine the tuning targets of the to-be-tuned parameters.

[0033] In one embodiment, first, a deep understanding and clear specific requirements of data fusion are needed. This includes but is not limited to determining which types of data need to be fused (such as text, image, audio, etc.), what the goal of data fusion is (such as classification, regression, clustering, etc.), and how the data after fusion will be applied (such as for decision support, predictive analysis, etc.). After the data fusion requirements are clear, the next step is to conduct a deep feature analysis of the data to be fused. This includes the type of data (such as structured data, unstructured data, etc.), the distribution of data (such as normal distribution, skew distribution, etc.), the statistical properties of data (such as mean, variance, correlation, etc.), and possible outliers, missing values, etc. In addition, the temporal and spatial characteristics of the data also need to be considered, which is particularly important for the fusion of cross-temporal and spatial data. Based on the deep understanding of the characteristics of the data to be fused, the next step is to find the parameters corresponding to these data characteristics in the large model to be applied. These parameters may include the weights, biases, convolution kernels, etc. of the model, which directly affect the way the model processes data and the effect. By matching data characteristics and model parameters, the model can better adapt to and process the data to be fused. Finally, a clear fine-tuning target needs to be set for each fine-tuning parameter. These targets should be consistent with the ultimate goal of data fusion, such as improving classification accuracy, reducing regression error, etc. By setting clear fine-tuning targets, the subsequent model fine-tuning process can be more targeted and efficient. At the same time, it is also helpful to evaluate whether the performance of the fine-tuned model meets the expected requirements.

[0034] In one embodiment of the present application, after determining the fine-tuning parameters in the large model to be applied, the fine-tuning parameters are fine-tuned by a large model fine-tuning algorithm to obtain a large model to be trained.

[0035] Specifically, based on the fine-tuning target of each fine-tuning parameter, a corresponding fine-tuning target matrix is determined; the fine-tuning target matrix is decomposed to obtain an initial dimension reduction matrix and an initial dimension increase matrix; the initial dimension increase matrix is initialized by a Gaussian distribution to obtain an application dimension increase matrix, and the initial dimension reduction matrix is initialized by a 0 matrix to obtain an application dimension reduction matrix; based on the application dimension increase matrix and the application dimension reduction matrix, a fine-tuning application matrix is constructed, and based on the fine-tuning application matrix, a corresponding application branch of the large model to be applied is added to obtain a large model to be trained.

[0036] In one embodiment, the determined fine-tuning target matrix can be represented as AW based on the fine-tuning target of each parameter to be fine-tuned. It can be understood that other parameters in the fine-tuning target matrix AW that do not need to be fine-tuned are set to 0. The matrix decomposition technique is used to decompose the fine-tuning target matrix AW into two one-dimensional matrices, an initial dimension reduction matrix B0 and an initial dimension increase matrix A0. Then, the initial dimension increase matrix is initialized by a Gaussian distribution to obtain an application dimension increase matrix A, and the initial dimension reduction matrix is initialized by a 0 matrix to obtain an application dimension reduction matrix B, so as to ensure that the fine-tuning application matrix BA is a 0 matrix and does not produce noise. Finally, according to the fine-tuning application matrix, a corresponding to-be-applied branch is added to the to-be-applied large model. This branch will be responsible for adjusting the parameters of the model according to the guidance of the fine-tuning application matrix. By integrating this new branch into the large model, a to-be-trained, more flexible and adjustable large model is obtained.

[0037] Step 102, obtaining training corpus, and training the to-be-trained large model based on the training corpus to obtain an application large model.

[0038] In one embodiment of the present application, after obtaining the training corpus, the method further comprises: cleaning the training corpus to remove irrelevant information, duplicate data and outliers; annotating the cleaned training corpus; wherein the annotation includes semantic role annotation and named entity recognition; performing word segmentation processing on the annotated training corpus to realize word segmentation, stem extraction and stop word removal.

[0039] Further, after obtaining the training corpus, the to-be-trained large model is trained based on the training corpus to obtain an application large model.

[0040] Specifically, the to-be-applied large model in the to-be-trained large model is trained based on the training corpus to obtain converged initial application parameters; and the to-be-applied branch in the to-be-trained large model is trained based on the training corpus to obtain converged branch application parameters; the initial application parameters and the branch application parameters are superimposed to obtain application parameters; the to-be-applied large model is updated based on the application parameters to obtain an application large model.

[0041] In one embodiment, first, the training corpus is input into the to-be-applied large model. Through the back propagation algorithm and gradient descent optimization technique, the parameters of the to-be-applied large model are constantly adjusted to minimize the gap between the predicted value and the true value. During the training process, the performance of the model is continuously monitored, and the model parameters after each iteration are recorded. When the performance of the model on the validation set reaches the preset standard or the number of training rounds reaches the preset upper limit, the training is stopped, and the model parameters at this time are the converged initial application parameters. Training the to-be-trained branch in the to-be-trained large model: similarly, the training corpus is input into the to-be-applied branch. The to-be-applied branch is trained using similar optimization techniques to obtain optimal branch parameters. The performance of the branch is monitored, and the training is stopped when the performance reaches a stable state or the number of training rounds reaches an upper limit, and the branch parameters at this time are the converged branch application parameters. Superimpose the initial application parameters and the branch application parameters: superimpose or fuse the converged initial application parameters and the branch application parameters. This can be achieved through weighted averaging, direct addition, or other suitable parameter fusion methods. The fused parameters are the final application parameters, which will be used to update the to-be-applied large model. Update the to-be-applied large model based on the application parameters: use the fused application parameters to update the weights and biases of the to-be-applied large model. The updated to-be-applied large model is the final application large model, which has been optimized through the training corpus and has better prediction and generalization capabilities.

[0042] Step 103, in the case of needing to handle actual problems, prompt design is performed based on the problem information input by the user.

[0043] In one embodiment of the present application, after training the to-be-trained large model based on the training corpus to obtain the application large model, if it is necessary to handle actual problems, prompt design is performed based on the problem information input by the user.

[0044] Specifically, the problem information is parsed to decompose the problem information into a plurality of sub-questions; based on the plurality of sub-questions, the application large model is used to answer the questions to determine a plurality of corresponding sub-replies; and the plurality of sub-questions and the plurality of corresponding sub-replies are integrated.

[0045] In one embodiment, when we fully train the large model to be trained through the training corpus and obtain the application large model, we can use this well-trained large model to process actual problems. In order to more effectively obtain information from users and generate corresponding replies, we will design prompts based on the problem information input by users. After receiving the problem information input by users, we first need to deeply analyze it. Using natural language processing techniques such as word segmentation, part-of-speech tagging, and semantic role labeling, we can analyze the problem information in detail. Through analysis, we can decompose complex problem information into several sub-questions, each focusing on a specific point, which facilitates more accurate answers from the application large model. We input the decomposed sub-questions one by one into the already trained application large model. The large model will generate corresponding sub-replies based on the specific content of each sub-question using its learned knowledge and reasoning ability. After obtaining all sub-replies, we need to effectively integrate them. According to the logical order and importance of sub-questions, we can reasonably sort and organize the sub-replies. If necessary, we can further process and edit the sub-replies to ensure the coherence and readability of the final reply.

[0046] Step 104, based on the corresponding design results of the prompt design, the application large model is used to answer the question to determine the corresponding actual problem reply.

[0047] In one embodiment of the present application, after the prompt design based on the problem information input by the user, the application large model is used to answer the question based on the corresponding design results of the prompt design to determine the corresponding actual problem reply.

[0048] In one embodiment of the present application, after the application large model is used to answer the question to determine the corresponding actual problem reply, the application large model is optimized based on the preset Bradley-Terry preference model and reward function.

[0049] Specifically, according to the quality of the actual problem reply, the preference score of the reply is calculated through the Bradley-Terry preference model; based on the preference score, the reward value is calculated through the reward function; using the policy gradient method, the parameters of the application large model are updated and optimized according to the reward value.

[0050] In one embodiment, the preference score of the reply is calculated by the Bradley-Terry preference model according to the quality of the reply. The Bradley-Terry model is a statistical model widely used in comparative evaluation and ranking, which can effectively measure the relative advantages and disadvantages between different replies. In this step, the reply generated by the large model is compared with the standard answer or the reply of other models, so as to obtain a quantitative preference score. With this preference score, the reward value can be calculated by a pre-set reward function. This reward function is designed to reflect the quality and accuracy of the reply, so that the reply with a high preference score will obtain a higher reward value. Finally, the strategy gradient method is adopted to update and optimize the parameters of the large model according to the calculated reward value. The strategy gradient is a reinforcement learning algorithm that can maximize the expected reward value by adjusting the parameters of the model. In this way, our model can gradually learn how to generate replies that meet the user's preferences and actual needs.

[0051] In one embodiment of the present application, after the user inputs the question information, the question information is also subjected to security check and filtering; and after determining the actual question reply corresponding thereto, the actual question reply is monitored and audited; when an abnormal situation is found, the preset abnormal handling logic is handled based on the same.

[0052] In one embodiment, after the user inputs the question information, the system first performs security check and filtering on the question information. This process is mainly to prevent the input of malicious information, sensitive information or inappropriate remarks. The present application uses natural language processing and regular expression technology to scan and analyze the input text to identify and filter out malicious code, sensitive words or inappropriate content. Secondly, after determining the actual question reply corresponding thereto, the reply content is strictly monitored and audited. This link aims to ensure the accuracy and compliance of the reply, and prevent misleading users or spreading false information. During the monitoring and auditing process, once any abnormal situation is found, such as the reply content not matching the question, containing inappropriate remarks or illegal information, etc., the system will immediately trigger the preset abnormal handling logic. This handling logic may include but is not limited to: shielding or deleting inappropriate content, sending warning information to the user, limiting the user's rights, recording and reporting abnormal situations, etc. Through these security measures, the user can be effectively protected from harmful information, while ensuring the compliance and accuracy of the platform content. In addition, these measures can also help us to discover and handle potential security risks in a timely manner, and provide a safe and reliable Q&A environment for users.

[0053] The above is the method embodiment of the present application. Based on the same inventive concept, the embodiment of the present application also provides a cross-time-space data continuous fusion device based on a large model, the structure of which is shown in FIG. 2.

[0054] Figure 2 is a schematic diagram of the internal structure of a cross-temporal and spatial data continuous fusion device based on a large model according to an embodiment of the present application. As shown in Figure 2, the device comprises:

[0055] at least one processor 201;

[0056] and a memory 202 in communication connection with the at least one processor;

[0057] The memory 202 stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor 201 to enable the at least one processor 201 to:

[0058] determine the to-be-tuned parameters in the to-be-applied large model based on the preset data fusion requirements, and tune the to-be-tuned parameters through a large model tuning algorithm to obtain a to-be-trained large model;

[0059] obtain training corpus and train the to-be-trained large model based on the training corpus to obtain an applied large model;

[0060] in the case of needing to process actual problems, prompt design is performed based on the problem information input by the user;

[0061] based on the design result corresponding to the prompt design, question and answer are performed through the applied large model to determine the reply to the corresponding actual problem.

[0062] Some embodiments of the present application provide a non-volatile computer storage medium for cross-temporal and spatial data continuous fusion based on a large model corresponding to Figure 1, which stores computer executable instructions, and the computer executable instructions are configured to:

[0063] determine the to-be-tuned parameters in the to-be-applied large model based on the preset data fusion requirements, and tune the to-be-tuned parameters through a large model tuning algorithm to obtain a to-be-trained large model;

[0064] obtain training corpus and train the to-be-trained large model based on the training corpus to obtain an applied large model;

[0065] in the case of needing to process actual problems, prompt design is performed based on the problem information input by the user;

[0066] based on the design result corresponding to the prompt design, question and answer are performed through the applied large model to determine the reply to the corresponding actual problem.

[0067] The various embodiments in the present application are described in a progressive manner, and the same or similar parts among the various embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, the IoT device and medium embodiments are described simply because they are basically similar to the method embodiments.

[0068] The system and medium provided by the embodiments of the present application are one-to-one corresponding to the method, and therefore, the system and medium also have similar beneficial technical effects to the method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the system and medium will not be described here.

[0069] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. In addition, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage, etc.) containing computer-usable program code.

[0070] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions specified in one or more flows of the flowcharts and / or one or more blocks of the block diagrams.

[0071] These computer program instructions can also be stored in a computer-readable memory that can cause the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including instruction devices that implement the functions specified in one or more flows of the flowcharts and / or one or more blocks of the block diagrams.

[0072] These computer program instructions can also be loaded into a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks.

[0073] In one typical arrangement, the computing device includes one or more processors (GPUs), input / output interfaces, network interfaces, and memory.

[0074] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) and / or cache memory, non-volatile memory, such as read-only memory (ROM), EPROM, and / or flash memory, etc. The memory is an example of computer readable media.

[0075] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0076] It should also be noted that the terms "comprising", "including", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element defined by the statement "comprising a" does not exclude the existence of additional identical elements in the process, method, article, or apparatus including the stated element.

[0077] The above merely provides an example of the present application, but is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the scope of claims of the present application.

Claims

1. A large model-based cross-temporal and spatial data continuous fusion method, wherein, The method comprises: Based on the preset data fusion requirement, determine the to-be-fine-tuned parameters in the to-be-applied large model, and fine-tune the to-be-fine-tuned parameters through a large model fine-tuning algorithm to obtain a to-be-trained large model; Obtain training corpus, and train the to-be-trained large model based on the training corpus to obtain an applied large model; In the case of needing to process actual problems, prompt design is performed based on user input problem information; Based on the design result corresponding to the prompt design, the applied large model is used for question and answer to determine the actual problem reply corresponding thereto.

2. The large model-based cross-temporal and spatial data continuous fusion method according to claim 1, wherein, Based on the preset data fusion requirement, determine the to-be-fine-tuned parameters in the to-be-applied large model, specifically comprising: According to the data fusion requirement, analyze the to-be-required data characteristics; Based on the to-be-required data characteristics, match the corresponding to-be-fine-tuned parameters in the to-be-applied large model, and determine the fine-tuning target of each to-be-fine-tuned parameter.

3. The method of claim 2, wherein, Fine-tune the to-be-fine-tuned parameters through a large model fine-tuning algorithm to obtain a to-be-trained large model, specifically comprising: Based on the fine-tuning target of each to-be-fine-tuned parameter, determine the corresponding fine-tuning target matrix; Decompose the fine-tuning target matrix to obtain an initial dimension reduction matrix and an initial dimension increase matrix; Initialize the initial dimension increase matrix through Gaussian distribution to obtain an application dimension increase matrix, and initialize the initial dimension reduction matrix through a 0 matrix to obtain an application dimension reduction matrix; Based on the application dimension increase matrix and the application dimension reduction matrix, construct a fine-tuning application matrix, and based on the fine-tuning application matrix, add a corresponding to-be-applied branch to the to-be-applied large model to obtain a to-be-trained large model.

4. The large model-based cross-temporal and spatial data continuous fusion method according to claim 3, wherein, Based on the training corpus, train the to-be-trained large model to obtain an applied large model, specifically comprising: Based on the training corpus, train the to-be-applied large model in the to-be-trained large model to obtain a converged initial application parameter; and Based on the training corpus, train the to-be-applied branch in the to-be-trained large model to obtain a converged branch application parameter; Superimpose the initial application parameter and the branch application parameter to obtain an application parameter; Update the to-be-applied large model based on the application parameter to obtain an applied large model.

5. The large model-based cross-temporal and spatial data continuous fusion method according to claim 1, wherein, Based on user input problem information, prompt design is performed, specifically comprising: Parse the problem information to decompose the problem information into a plurality of sub-questions; Based on the plurality of sub-questions, the applied large model is used for question and answer to determine a plurality of sub-replies corresponding thereto; Integrate the plurality of sub-questions and the plurality of sub-replies corresponding thereto.

6. The large model-based cross-temporal and spatial data continuous fusion method according to claim 1, wherein, After question and answer through the applied large model to determine the actual problem reply corresponding thereto, the method further comprises: Based on the preset Bradley-Terry preference model and reward function, the applied large model is optimized, specifically comprising: According to the quality of the actual problem reply, the preference score of the reply is calculated through the Bradley-Terry preference model; Based on the preference score, the reward value is calculated through the reward function; Using the policy gradient method, the parameters of the applied large model are updated and optimized according to the reward value.

7. The large model-based cross-temporal and spatial data continuous fusion method according to claim 1, wherein, After obtaining the training corpus, the method further comprises: cleaning the training corpus to remove irrelevant information, duplicate data and outliers; annotating the cleaned training corpus; wherein the annotation includes semantic role labeling and named entity recognition; performing word segmentation processing on the annotated training corpus to realize word segmentation, stem extraction and stop word removal.

8. The large model-based cross-temporal and spatial data continuous fusion method according to claim 1, wherein, The method further comprises: after the user inputs the question information, performing security check and filtering on the question information; and, after determining the actual problem reply corresponding thereto, monitoring and auditing the actual problem reply; when an abnormal situation is found, processing based on the preset abnormal handling logic.

9. A large model-based cross-spatial and temporal data continuous fusion device, wherein, The device comprises: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to: based on the preset data fusion requirement, determine the to-be-fine-tuned parameters in the to-be-applied large model, and fine-tune the to-be-fine-tuned parameters through a large model fine-tuning algorithm to obtain a to-be-trained large model; obtain a training corpus, and train the to-be-trained large model based on the training corpus to obtain an application large model; in the case of needing to handle an actual problem, based on the question information input by the user, prompt design is performed; based on the design result corresponding to the prompt design, question and answer are performed through the application large model to determine the actual problem reply corresponding thereto. 10.A non-transitory computer storage medium storing computer-executable instructions for persistent fusion of cross-temporal and cross-spatial data based on large models, wherein, The computer executable instructions are configured to: based on the preset data fusion requirement, determine the to-be-fine-tuned parameters in the to-be-applied large model, and fine-tune the to-be-fine-tuned parameters through a large model fine-tuning algorithm to obtain a to-be-trained large model; obtain a training corpus, and train the to-be-trained large model based on the training corpus to obtain an application large model; in the case of needing to handle an actual problem, based on the question information input by the user, prompt design is performed; based on the design result corresponding to the prompt design, question and answer are performed through the application large model to determine the actual problem reply corresponding thereto.

Images