Social world simulation method and system based on large model and electronic equipment

Abstract

The invention provides a social world simulation method and system based on a large model and electronic equipment, and the method comprises the steps: obtaining multi-modal social event information; inputting the social event information into an event cue word generator based on a deep learning model to obtain an event cue word; inputting the event prompt words into a pre-trained multi-modal large model to obtain social event multi-modal features; inputting the social spatio-temporal information into a spatio-temporal cue word generator based on a deep learning model to obtain spatio-temporal cue words; inputting the space-time cue words into a pre-trained large language model to obtain social event space-time features; carrying out feature fusion on the social event multi-modal features and the social event spatio-temporal features through a pre-trained multi-modal large model to obtain event embedding vectors; and inputting the reasoning prompt word corresponding to the event embedding vector into a pre-trained large language model to obtain a simulated social event set. According to the invention, the efficiency and accuracy of social event simulation can be improved.

Description

Technical Field

[0001] The present invention relates to the field of social computing technology, and in particular to a social world simulation method, system and electronic device based on a large model. Background Art

[0002] Social simulation technology focuses on simulating the potential impact of social behaviors, events, and policy fluctuations on social structure and individual behavior.

[0003] There are many social simulation systems, including those based on rules, social dynamics, and emotion diffusion models. However, these systems generally have the problem of single information modality, which may lead to the one-sidedness and limitations of social event simulation, that is, due to the use of only a single type of information, social event simulation may not accurately reflect the real state and behavior of the social system.

[0004] This shows that the social simulation method in the relevant technology has the technical problem of inaccurate simulation of social events due to the single information modality. Summary of the invention

[0005] The present invention provides a social world simulation method, system and electronic device based on a large model, so as to solve the defect of the social simulation method in the prior art that the social event simulation is inaccurate due to the single information mode, and to improve the efficiency and accuracy of the social event simulation.

[0006] The present invention provides a social world simulation method based on a large model, comprising the following steps.

[0007] Acquire multimodal social event information, wherein the multimodal social event information includes: social event information and social spatiotemporal information; input the social event information into an event prompt word generator based on a deep learning model to obtain event prompt words output by the event prompt word generator; input the event prompt words into a pre-trained multimodal large model to obtain multimodal features of social events output by the pre-trained multimodal large model; input the social spatiotemporal information into a spatiotemporal prompt word generator based on a deep learning model to obtain spatiotemporal prompt words output by the spatiotemporal prompt word generator; input the spatiotemporal prompt words into a pre-trained large language model to obtain spatiotemporal features of social events output by the pre-trained large language model; perform feature fusion on the multimodal features of social events and the spatiotemporal features of social events through the pre-trained multimodal large model to obtain event embedding vectors; input the inference prompt words corresponding to the event embedding vectors into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model.

[0008] According to a large model-based social world simulation method provided by the present invention, the inference prompt words corresponding to the event embedding vector are input into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, including: inputting the event embedding vector into an event reasoning model based on a long short-term memory mechanism to obtain the inference prompt words output by the event reasoning model; and performing reasoning based on the inference prompt words through the pre-trained large language model according to the reasoning-action reasoning mode to obtain a set of simulated social events.

[0009] According to a large-model-based social world simulation method provided by the present invention, after the pre-trained large language model is used to perform reasoning based on the reasoning prompt words according to the reasoning-action reasoning mode to obtain a set of simulated social events, the method further includes: determining the contextual reasoning process of each simulated social event in the set of simulated social events; based on a preset situational awareness algorithm, performing an importance evaluation based on the contextual reasoning process of each simulated social event to determine the importance score of each simulated social event.

[0010] According to a social world simulation method based on a large model provided by the present invention, before the event embedding vector is input into an event reasoning model based on a long short-term memory mechanism to obtain the reasoning prompt words output by the event reasoning model, the method also includes: obtaining training data with labels; iteratively performing the following steps until the similarity error reaches a preset value to obtain a trained event reasoning model: inputting the training data into the event reasoning model to obtain the reasoning prompt word samples output by the event reasoning model; determining the similarity error between the reasoning prompt word samples and the labels of the training data; based on the similarity error, determining the gradient of the model parameters of the event reasoning model; updating the model parameters by a preset gradient descent algorithm to obtain an updated event reasoning model.

[0011] According to a large-model-based social world simulation method provided by the present invention, after inputting the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, the method further includes: when there is a loss of a target simulated social event or a target simulated social event is lost in the set of simulated social events, determining the associated social events of the target simulated social event through a graph convolutional neural network; and restoring the target simulated social event based on the associated social events through a graph convolutional neural network.

[0012] According to a large-model-based social world simulation method provided by the present invention, after inputting the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, the method further includes: performing data conversion on the set of simulated social events to obtain an output result in a target format, wherein the target format at least includes: triples.

[0013] The present invention also provides a social world simulation system based on a large model, comprising the following modules: an event acquisition module, used to acquire multimodal social event information, wherein the multimodal social event information includes: social event information and social spatiotemporal information; an event prompt word module, used to input the social event information into an event prompt word generator based on a deep learning model, and obtain event prompt words output by the event prompt word generator; a multimodal large model module, used to input the event prompt words into a pre-trained multimodal large model, and obtain the multimodal features of social events output by the pre-trained multimodal large model; a spatiotemporal prompt word module, used to input the social spatiotemporal information into an event prompt word generator based on a deep learning model, and obtain the ...; a multimodal large model module, used to input the event prompt words into a pre-trained multimodal large model, and obtain the multimodal features of social events output by the pre-trained multimodal large model; a multimodal large model module, used to input the social spatiotemporal information into an event prompt word generator based on a deep learning model; a multimodal large model module, used to input the event prompt words into a pre-trained multimodal large model, and obtain the multimodal features of social events output by the pre-trained multimodal large model; a multi to a spatiotemporal prompt word generator based on a deep learning model, and obtain the spatiotemporal prompt words output by the spatiotemporal prompt word generator; a large language model module, used to input the spatiotemporal prompt words into a pre-trained large language model, and obtain the spatiotemporal features of social events output by the pre-trained large language model; a feature fusion module, used to perform feature fusion of the multimodal features of social events with the spatiotemporal features of social events through the pre-trained multimodal large model, and obtain an event embedding vector; a simulation output module, used to input the inference prompt words corresponding to the event embedding vector into the pre-trained large language model, and obtain a set of simulated social events output by the pre-trained large language model.

[0014] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the method for simulating the social world based on a large model as described above is implemented.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-described large-model-based social world simulation methods.

[0016] The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the above-mentioned large-model-based social world simulation methods.

[0017] The social world simulation method, system and electronic device based on a large model provided by the present invention can more comprehensively reflect various aspects of social events by acquiring multimodal social event information including social event information and social spatiotemporal information; through an event prompt word generator and a spatiotemporal prompt word generator based on a deep learning model, high-quality event prompt words and spatiotemporal prompt words can be automatically generated from original information, which not only helps the processing and analysis of subsequent models, but also improves the accuracy and efficiency of information extraction to a certain extent; through a pre-trained multimodal large model, event prompt words are converted into multimodal features of social events, which can more comprehensively reflect the actual situation of the event. situation; the spatiotemporal cue words are input into the pre-trained large language model to obtain the spatiotemporal characteristics of social events, which can reflect the distribution and changes of events in time and space; through the pre-trained multimodal large model, the multimodal characteristics of social events and the spatiotemporal characteristics of social events are fused to obtain event embedding vectors, which improves the richness and accuracy of feature expression; using the reasoning ability of the large language model, according to the reasoning cue words of the event embedding vector, a set of simulated social events is generated, realizing a comprehensive, accurate and intelligent simulation of social events; thereby solving the technical problem of inaccurate social event simulation due to the single information modality in the social simulation method in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced one by one below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0019] Figure 1 It is a flowchart of the social world simulation method based on a large model provided by the present invention.

[0020] Figure 2 It is a basic principle diagram of the social world simulation method based on a large model provided by the present invention.

[0021] Figure 3 It is a module flow chart of the social world simulation method based on a large model provided by the present invention.

[0022] Figure 4 It is the overall framework diagram of the social world simulation method based on the big model provided by the present invention.

[0023] Figure 5 It is a schematic diagram of the structure of the social world simulation system based on the big model provided by the present invention.

[0024] Figure 6It is a schematic diagram of the physical structure of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0026] Social simulation technology focuses on simulating the potential impact of social behavior, events, and policy fluctuations on social structure and individual behavior. There are many social simulation systems, including those based on rules, social dynamics, and emotional diffusion models. However, these systems generally have the following problems: single information modality and simple social interaction.

[0027] To solve the above problems, the present invention proposes a social world simulation method based on a large model, which can simulate the spatiotemporal elements, group cognitive psychological behaviors and the impact and propagation mechanism of major social events in the social environment, and can predict (simulate) the future development trend of input social events.

[0028] Optionally, the social world simulation method based on a big model in the embodiment of the present application can be executed by a server, or by a terminal device, or jointly by a server and a terminal device, taking the example of the social world simulation method based on a big model in this embodiment being executed by a server.

[0029] Figure 1 is a flow chart of the social world simulation method based on a large model provided by the present invention, such as Figure 1 As shown, the method includes the following: Step 101, obtaining multimodal social event information, wherein the multimodal social event information includes: social event information and social spatiotemporal information.

[0030] By collecting multimodal social event information including social event information and social spatiotemporal information, it is possible to describe and analyze social events more comprehensively, where multimodality includes: video, text, and audio.

[0031] Social event information provides the essential content of the event, such as event type, participants, location, etc.; while social spatiotemporal information provides the specific time and place of the event, as well as the relevant spatiotemporal background.

[0032] For example, social media platforms usually require users to authorize the acquisition of their geographic location information. By analyzing the geo-tagged information posted by users, we can understand the behavior and interests of users in different locations, and thus infer the social events that have occurred, as well as the geographic location and spatiotemporal distribution of social events. The data on social media platforms are usually arranged in chronological order. By analyzing time series data, we can understand the behavior and trends of users at different times, and then infer the temporal distribution and evolution of social events.

[0033] By using the embodiments of the present invention, multi-modal social event information data is obtained, and the evolution process of social events can be effectively analyzed.

[0034] Step 102: input the social event information into an event prompt word generator based on a deep learning model to obtain event prompt words output by the event prompt word generator.

[0035] Among them, the prompt generator can automatically generate prompt words suitable for specific models or tasks based on the information or requirements input by the user.

[0036] The preprocessed social event information is passed as input to an event prompt word generator based on a deep learning model. The event prompt word generator will generate event prompt words based on the input social event information. The event prompt words are used to summarize the core content of the event.

[0037] The preprocessing process includes cleaning the collected data, removing irrelevant information (such as advertisements and spam), and performing text preprocessing operations such as word segmentation, stop word removal, and stem extraction.

[0038] In an embodiment of the present invention, an event prompt word generator is used to simulate a single event containing complex modal information; the event prompt word generator generates prompt words based on the event information to facilitate the subsequent triggering of the multimodal related content generation capability of the multimodal large model to simulate social events with cross-modal characteristics.

[0039] The event prompt word generator is supervised by deep learning and combined with a large multimodal model to realize event simulation.

[0040] Through the embodiments of the present invention, a deep learning model is used to extract key information from social event information and generate event prompt words. These prompt words can summarize the core content of the event and provide targeted input for subsequent multimodal feature extraction.

[0041] Step 103, input the event prompt words into the pre-trained multimodal large model to obtain the multimodal features of the social event output by the pre-trained multimodal large model.

[0042] The event prompt words are passed as text input to the pre-trained multimodal large model. The pre-trained multimodal large model encodes and processes the received input to generate multimodal features of social events; multimodal features of social events include semantic information in text, visual features in images, acoustic features in audio, etc.

[0043] The pre-trained multimodal large model utilizes a cross-modal attention mechanism to fuse information from different modalities to generate a more comprehensive event representation.

[0044] Through the embodiment of the present invention, event prompt words are converted into multimodal features of social events through a pre-trained multimodal large model; the multimodal features of social events include information in multiple modalities such as images, texts, and audios, which can more comprehensively reflect the actual situation of the event.

[0045] Step 104: input the social spatiotemporal information into a spatiotemporal prompt word generator based on a deep learning model to obtain spatiotemporal prompt words output by the spatiotemporal prompt word generator.

[0046] Social spatiotemporal information includes timestamps and geographic locations (such as latitude and longitude coordinates, administrative divisions). The social spatiotemporal information is input into the trained spatiotemporal prompt word generator to obtain output spatiotemporal prompt words. The spatiotemporal prompt words are used to represent the temporal and spatial characteristics of the input information.

[0047] In the embodiment of the present invention, the spatiotemporal prompt word generator is used to describe the spatiotemporal attributes of an event. The spatiotemporal prompt word generator generates spatiotemporal prompt words based on the input social spatiotemporal information, and then inputs the spatiotemporal prompt words into the large language model to generate the time information and spatial information of the social event. The spatiotemporal prompt word generator is a deep learning model, which is combined with the large language model to perform supervised learning to achieve model training.

[0048] Step 105, input the spatiotemporal prompt words into the pre-trained large language model to obtain the spatiotemporal features of the social events output by the pre-trained large language model.

[0049] The spatiotemporal cue words are passed as text input to the pre-trained large language model. The pre-trained large language model encodes the received spatiotemporal cue words through the self-attention mechanism to obtain feature representation; and extracts the spatiotemporal features of social events related to the spatiotemporal nature of social events from the feature representation generated by the model.

[0050] Among them, the spatiotemporal characteristics of social events include: time-related characteristics (such as the time window of event occurrence, time trend, etc.), space-related characteristics (such as the geographical location of event occurrence, spatial distribution, etc.), and spatiotemporal interaction characteristics (such as the propagation mode and impact scope of events in different time and space).

[0051] Through the embodiment of the present invention, the spatiotemporal prompt words are input into a pre-trained large language model to obtain the spatiotemporal characteristics of social events, which can reflect the distribution and changes of events in time and space.

[0052] Step 106, through the pre-trained multimodal large model, the multimodal features of social events and the spatiotemporal features of social events are fused to obtain an event embedding vector.

[0053] In an embodiment of the present invention, based on a pre-trained multimodal large model, feature fusion is performed on multimodal social simulation events (i.e., multimodal features of social events) and spatiotemporal events in a social spatiotemporal environment (i.e., spatiotemporal features of social events) to achieve event alignment and generate event embedding vectors.

[0054] For a social event, the generation of features of different dimensions includes: the event prompt word generator drives the multimodal large model to generate the multimodal information of the above social event (i.e. the multimodal features of social events); the spatiotemporal prompt word generator drives the large language model to generate the spatiotemporal information of the above social event (i.e. the spatiotemporal features of social events); then the multimodal information of the above social event is integrated (fused) with the spatiotemporal information to obtain the event embedding vector.

[0055] In order to achieve event alignment, it is necessary to determine the correspondence between the multimodal features of social events and the spatiotemporal features of social events. This can be achieved by calculating the similarity between features, using matching algorithms, or building alignment models. The purpose of alignment is to ensure that events from different data sources can be correctly associated during the feature fusion process.

[0056] The aligned multimodal features and spatiotemporal features of social events are input into the pre-trained multimodal large model for further fusion and processing, and a fixed-dimensional event embedding vector is generated through the output layer of the pre-trained multimodal large model. The event embedding vector is used to represent the multimodal features and spatiotemporal features of social events while maintaining sufficient abstractness and generalization capabilities.

[0057] Step 107, input the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model.

[0058] In an embodiment of the present invention, an event embedding vector is input into an event reasoning model combined with a long short-term memory mechanism to obtain a reasoning prompt word; and the reasoning prompt word corresponding to the event embedding vector is input into a pre-trained large language model to trigger an event reasoning function driven by the large language model, thereby obtaining a set of simulated social events output by the pre-trained large language model.

[0059] Through the embodiments of the present invention, by inputting the reasoning prompt words corresponding to the event embedding vector into the pre-trained large language model, the model can be guided to perform more accurate reasoning and simulation, thereby further improving the intelligence and accuracy of the simulation results.

[0060] According to a social world simulation method based on a large model provided by the present invention, the reasoning prompt words corresponding to the event embedding vector are input into a pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, including: Input the event embedding vector into the event reasoning model based on the long short-term memory mechanism to obtain the reasoning prompt words output by the event reasoning model; Through the pre-trained large language model, according to the reasoning-action reasoning mode, reasoning is performed based on the reasoning prompt words to obtain a set of simulated social events.

[0061] In an embodiment of the present invention, an event embedding vector is input into an event reasoning model combined with a long short-term memory mechanism to generate reasoning prompts, and the event reasoning function driven by a pre-trained large language model is triggered by the reasoning prompts; the pre-trained large language model is based on the reasoning-action (ReAct, Reasoning-Acting) reasoning mode to obtain a set of simulated social events.

[0062] Here, the reasoning-action reasoning model is a framework that combines reasoning and action, and is used to optimize the model's response through logical reasoning and context analysis. That is, it emphasizes the introduction of reasoning mechanisms in the generation process, so that the model not only relies on existing knowledge, but can also perform logical reasoning and decision-making based on the characteristics of the problem.

[0063] The event embedding vector is used as the input of the event reasoning model based on the long short-term memory mechanism. The input event embedding vector is processed by the long short-term memory mechanism to extract the temporal relationship and contextual information between events, and output reasoning prompt words. The reasoning prompt words can be keywords, phrases or sentences to guide the subsequent large language model to perform reasoning.

[0064] The large language model performs reasoning and generation based on the input reasoning prompt words in combination with its internal knowledge base and language model; the large language model outputs a series of simulated social events (i.e., a set of simulated social events), which can be results or scenarios derived based on the reasoning prompt words.

[0065] Through the embodiments of the present invention, the long short-term memory (LSTM) mechanism can capture the long-term dependencies in the event sequence, so that the event reasoning model can output reasoning prompt words. The pre-trained large language model has powerful language understanding and generation capabilities, and can perform in-depth reasoning based on the reasoning prompt words to simulate a variety of possible social event sets.

[0066] According to a social world simulation method based on a large model provided by the present invention, after obtaining a set of simulated social events by reasoning based on reasoning prompt words according to the reasoning-action reasoning mode through a pre-trained large language model, the method further includes: determining a contextual reasoning process for each simulated social event in the set of simulated social events; Based on the preset situation-aware algorithm, an importance assessment is performed based on the contextual reasoning process of each simulated social event to determine the importance score of each simulated social event.

[0067] In the embodiment of the present invention, key social events are automatically identified and extracted from the simulation process. The importance of the event is evaluated by using the situational awareness algorithm and combined with the environmental context information, and the learning strategy is adjusted accordingly to achieve more effective model training; and the parameters are automatically adjusted according to the performance in the learning process to improve the learning efficiency and accuracy.

[0068] Context-aware algorithms can capture and understand the current context or contextual environment status. During model training, context-aware algorithms can identify the changing patterns of data in different contexts and provide a basis for evaluating the importance of events.

[0069] For example, by combining environmental context information and using situational awareness algorithms to evaluate the importance of extracted key social events, the evaluation indicators may include the scale of the event, the scope of impact, urgency, social attention, etc. By comprehensively analyzing these factors, an importance score or level can be assigned to each event.

[0070] According to the evaluation results of the importance of events, the learning strategy of the model is adjusted to give priority to those social events with higher importance, for example, increasing the number of simulations of important events, improving the data sampling rate, optimizing model parameters, etc. The model is trained using the adjusted learning strategy to improve the model's ability to identify and predict key social events.

[0071] Through the embodiments of the present invention, by combining simulation deduction, situational awareness algorithm and learning strategy adjustment, more effective model training can be achieved and the ability to identify and predict social events can be improved.

[0072] According to a social world simulation method based on a large model provided by the present invention, before inputting the event embedding vector into an event reasoning model based on a long short-term memory mechanism to obtain the reasoning prompt words output by the event reasoning model, the method further includes: Get labeled training data; Iterate the following steps until the similarity error reaches the preset value to obtain the trained event reasoning model: Input the training data into the event reasoning model to obtain the reasoning prompt word samples output by the event reasoning model; Determine the similarity error between the inference hint word sample and the label of the training data; Based on the similarity error, determining the gradient of the model parameters of the event reasoning model; The model parameters are updated through the preset gradient descent algorithm to obtain an updated event reasoning model.

[0073] In an embodiment of the present invention, the model training of the event reasoning model is based on a pre-trained large language model for supervised learning, and the specific steps include: Step 1: Input labeled training data; Prepare and input a set of training data with annotations (or labels), which are the true output or target value of the data, used to guide the model during the training process.

[0074] Step 2, calculate and output the prediction results; Use the current model parameters to predict the input data and output the prediction results.

[0075] Step 3, calculate the similarity between the result and the label; Evaluate the difference or error between the predicted result and the actual label. Commonly used loss functions include mean square error (MSE), cross entropy loss, etc. The calculated loss value represents the prediction error (i.e., similarity error) of the current state of the model.

[0076] Step 4, calculate the gradient based on the similarity; The loss function is used to derive the model parameters to obtain the gradient; the gradient represents the rate of change of the loss function relative to the model parameters and is used to guide the update direction of the model parameters; in the neural network, the gradient is calculated by the back propagation algorithm.

[0077] Step 5, optimization by gradient descent; Update the model parameters according to the gradient to reduce the prediction error; gradient descent is an optimization algorithm that minimizes the loss function by iteratively adjusting the model parameters.

[0078] Step 6: Return to step 1 and perform iterative optimization until the training error reaches the preset value, and then end the training.

[0079] Repeat steps 1 to 5 until the stopping condition is met; the stopping condition can be that the training error reaches a preset threshold, the number of iterations reaches an upper limit, the performance of the validation set no longer improves, etc. When the performance of the model on the training set is stable and no longer improves significantly, the model is considered to have converged.

[0080] Through iterative optimization of the above steps, the model can gradually learn the inherent rules of the training data and make accurate predictions on new unseen data.

[0081] According to a social world simulation method based on a large model provided by the present invention, after inputting the reasoning prompt words corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, the method further includes: When there is a target simulated social event loss or a target simulated social event is lost in the simulated social event set, the associated social events of the target simulated social event are determined through a graph convolutional neural network; Through graph convolutional neural networks, the target simulated social events are restored based on related social events.

[0082] In an embodiment of the present invention, data compression and encoding technology is used to optimize the use of storage space while ensuring rapid retrieval and updating of data; a self-healing mechanism is also integrated, which can mine the association information between memories through a graph convolutional neural network algorithm when data is damaged or lost, automatically restore key information, and ensure the integrity and reliability of memory data.

[0083] By compressing data, the storage space occupied can be effectively reduced and the storage cost can be lowered. The compression technology can be lossy compression (such as JPEG image compression) or lossless compression (such as ZIP file compression). The specific choice depends on the type and importance of the data. Coding technology can ensure that key information is not lost during the compression process, while improving the transmission efficiency and storage density of data. For example, Huffman coding and run-length coding are often used for lossless data compression.

[0084] Through the above compression and encoding technologies, the system can use storage space more efficiently and reduce the storage of redundant data. It can also adopt a hierarchical storage strategy to store data of different importance or access frequency on different levels of storage media (such as SSD, HDD) to balance performance and cost.

[0085] The self-healing mechanism is used to automatically detect, diagnose and repair system failures or data corruption. In the storage system, the self-healing mechanism can monitor the integrity of data in real time and immediately start the recovery process once data corruption or loss is detected.

[0086] Graph convolutional neural network is a neural network model specially designed for processing graph structured data; in storage systems, graph convolutional neural network can be used to mine the association information between memories, such as dependencies between data, access patterns, etc. When data is damaged or lost, graph convolutional neural network can use this association information to predict and restore key data.

[0087] Through the embodiments of the present invention, through the combination of the self-healing mechanism and the graph convolutional neural network, the system can automatically restore key information when the data is damaged or lost, ensuring the integrity and reliability of the data; in addition, redundant storage, data backup and other technologies can also be used to further improve the reliability of the data.

[0088] According to a social world simulation method based on a large model provided by the present invention, after inputting the reasoning prompt words corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, the method further includes: Data conversion is performed on the simulated social event set to obtain an output result in a target format, wherein the target format at least includes: a triple.

[0089] According to the characteristics of social events, define the structure of triples. For social events, triples can contain the following elements: Entity (Event): represents the social event itself. Relation (Relation): describes the association or attribute between entities. Attribute value (Value): provides a specific value or description for the relationship. For example, a possible triple structure is: <Event name, Relation type, Attribute value>.

[0090] The name or identifier of the social event is extracted from the simulated social event as the entity part of the triple; according to the defined triple structure, the relationship and attribute value related to the entity are extracted from the simulated social event; the extracted entity, relationship and attribute value are combined into a triple.

[0091] Output the constructed triples in the target format, which can be a text file, a database table, or any other format suitable for subsequent analysis and utilization.

[0092] Through the embodiments of the present invention, a set of simulated social events can be converted into a target format output result containing triples, providing support for subsequent analysis, prediction and decision-making.

[0093] The following describes an example of a practical application of the large-model-based social world simulation method provided by the present invention.

[0094] Social simulation technology focuses on simulating the potential impact of social behavior, events, and policy fluctuations on social structure and individual behavior. There are many social simulation systems, including those based on rules, social dynamics, and emotional diffusion models. However, these systems generally have the following problems: single information modality and simple social interaction.

[0095] To solve the above problems, the present invention proposes a method based on the joint drive of a multimodal large model and a large language model, and constructs a world simulator with the ability to understand multimodal information and model spatiotemporal information in the social field. The simulator can simulate the spatiotemporal elements, group cognitive psychological behavior, and the impact and propagation mechanism of major social events in the social environment, and can predict the future development trend of the input social events.

[0096] refer to Figure 2 , Figure 2 It is a basic principle diagram of the social world simulation method based on a large model provided by the present invention, which includes: inputting multimodal social event information into a social domain world simulator, and outputting predicted events.

[0097] refer to Figure 3 , Figure 3 It is a module flow chart of the social world simulation method based on a large model provided by the present invention, which includes: multimodal event simulation, social space-time simulation environment, dual-model driven event reasoning engine and prediction result output.

[0098] The present invention designs a world simulator for the humanities and social sciences. Its basic process is as follows: Figure 3 The simulator combines the intelligent agent technology driven by a multimodal large model and a large language model, and can simulate the evolution of multimodal social events in a dynamic social space-time environment. The present invention can obtain multimodal data related to social simulation from data sources such as the Internet space and field think tanks, covering social event data, network social data, and public opinion information. These data are used to build a social event simulator driven by a multimodal large model to simulate social scenarios from multiple dimensions.

[0099] In the simulation process, the present invention uses a large language model to construct a social space-time simulation environment, and expresses the space-time information of social events through language symbols. The event reasoning engine combines the multimodal large model with the social space-time simulation environment, assembles the multimodal event set with the simulation environment, and constructs a complete social field world simulation model. In addition, through the reasoning system driven by the large language model, social events are inferred and their future development trends are predicted.

[0100] refer to Figure 4 , Figure 4 It is the overall framework diagram of the social world simulation method based on the big model provided by the present invention.

[0101] The domain world simulator system based on large model agents consists of the following parts (see Figure 4 ). The dual-model driven event reasoning engine is mainly responsible for simulating events with multi-modal spatiotemporal information; the autonomous evolution unit is mainly used as a memory, recording important reasoning event information and can be updated autonomously; the output module will format and output the results of the dual-model driven event reasoning engine. The specific instructions are as follows: In an embodiment of the present invention, the dual-model driven event reasoning engine includes four main parts: a prompt generator for multimodal event simulation is used to generate prompt words. After the prompt words are input into the multimodal large model, the large model will generate multimodal features of the event; a prompt generator for social spatiotemporal environment simulation is also used to generate prompt words. After the prompt words are input into the large language model, the large language model will generate spatiotemporal features of the event; after the multimodal features and spatiotemporal features of the event are generated, the social domain world modeling fuses these two types of features to construct an event set with complete features; the large language model driven event reasoning model realizes event reasoning based on the event set and supported by the memory information in the word count evolution unit; driven by the multimodal and large language models, supported by the memory function of the autonomous evolution unit, the spatiotemporal reasoning of social events is realized.

[0102] The Prompt generator for multimodal event simulation (equivalent to the event prompt generator mentioned above) is used to simulate a single event containing complex modal information. The generator generates prompts based on event information, triggers the multimodal related content generation capability of the multimodal large model, and simulates social events with cross-modal characteristics. The generator performs supervised training through deep learning and realizes the simulation of domain events in combination with the multimodal large model.

[0103] The prompt generator for social spatiotemporal environment simulation (equivalent to the spatiotemporal prompt generator mentioned above) is used to describe the spatiotemporal attributes of events. The generator generates prompts based on the input social spatiotemporal information, and then inputs the prompts into the large language model to generate the time and space information of social events. The generator is a deep learning model that combines the large language model to conduct supervised learning and realize model training.

[0104] Social domain world modeling includes, based on the multimodal big model, feature fusion of multimodal social simulation events and spatiotemporal events in the social spatiotemporal environment, event alignment and generation of event embedding vectors. For a social event, the generation of features of different dimensions. The prompt generator for multimodal event simulation drives the multimodal big model to generate multimodal information of the event; while the prompt generator for social spatiotemporal environment simulation drives the big language model to generate spatiotemporal information of the event; then, the two types of information of the event are integrated. Through the prompt generation model (i.e., event reasoning model), prompts are generated according to the input event embedding vectors and input into the pre-trained big language model, thereby generating a social domain simulation model based on language symbols. Model training is supervised learning through labeled data sets and pre-trained big language models.

[0105] The event reasoning model driven by the large language model combines the long-term and short-term memory mechanism to generate prompts and trigger the event reasoning function driven by the large language model. Based on the ReAct (reasoning-acting) reasoning mode, the reasoning results generated at each step are fed back to the social domain world simulator, and the next step of reasoning is performed based on the feedback information. Model training is based on supervised learning based on the pre-trained large language model. The specific steps are as follows: Step 1: Input labeled training data; Step 2, calculate and output the prediction results; Step 3, calculate the similarity between the result and the label; Step 4, calculate the gradient based on the similarity; Step 5, optimization by gradient descent; Step 6: Return to step 1 and perform iterative optimization until the training error reaches the preset value, and then end the training.

[0106] In the embodiment of the present invention, the autonomous evolution unit includes: a dynamic adaptive memory and a context-aware autonomous learning model.

[0107] This paper proposes a new type of long-term and short-term memory, called dynamic adaptive memory, which can dynamically adjust its storage strategy according to the complexity and speed of change of the environment in which the agent is located. This memory uses advanced data compression and encoding technology to optimize the use of storage space while ensuring fast retrieval and update of data. The memory also integrates a self-healing mechanism, which can mine the association information between memories through the graph convolutional neural network algorithm when data is damaged or lost, automatically restore key information, and ensure the integrity and reliability of the agent's memory data.

[0108] The context-aware autonomous learning model is an advanced model based on deep learning that can automatically identify and extract key social events from the simulation process. This model uses context-aware algorithms and combines environmental context information to evaluate the importance of events and adjust the learning strategy accordingly to achieve more effective model training. The model also has self-optimization capabilities and can automatically adjust parameters based on performance during the learning process to improve learning efficiency and accuracy.

[0109] In the embodiment of the present invention, the result output module includes: a formatted output unit.

[0110] The deduction and prediction results are a set of social events generated by the social domain world simulator. The formatted output unit is used to convert the result data into a specified format, such as triples.

[0111] In an embodiment of the present invention, the pre-trained large model module includes: a pre-trained large language model and a pre-trained multimodal large model.

[0112] The social world simulation system based on a big model provided by the present invention is described below. The social world simulation system based on a big model described below and the social world simulation method based on a big model described above can refer to each other.

[0113] refer to Figure 5 , Figure 5 It is a schematic diagram of the structure of the social world simulation system based on the big model provided by the present invention.

[0114] An event acquisition module 501 is used to acquire multimodal social event information, wherein the multimodal social event information includes: social event information and social time-space information; An event prompt word module 502 is used to input social event information into an event prompt word generator based on a deep learning model to obtain event prompt words output by the event prompt word generator; The multimodal large model module 503 is used to input the event prompt words into the pre-trained multimodal large model to obtain the multimodal features of the social event output by the pre-trained multimodal large model; The spatiotemporal prompt word module 504 is used to input the social spatiotemporal information into the spatiotemporal prompt word generator based on the deep learning model to obtain the spatiotemporal prompt words output by the spatiotemporal prompt word generator; The large language model module 505 is used to input the spatiotemporal prompt words into the pre-trained large language model to obtain the spatiotemporal features of social events output by the pre-trained large language model; A feature fusion module 506 is used to perform feature fusion on multimodal features of social events and spatiotemporal features of social events through a pre-trained multimodal large model to obtain an event embedding vector; The simulation output module 507 is used to input the inference prompt words corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model.

[0115] Specifically, the above-mentioned social world simulation system based on big model provided by the present invention can implement all the method steps implemented by the above-mentioned social world simulation method embodiment based on big model, and can achieve the same technical effect. The parts and beneficial effects of this embodiment that are the same as those of the method embodiment will not be described in detail here.

[0116] Figure 6 is a schematic diagram of the physical structure of the electronic device provided by the present invention, such as Figure 6 As shown, the electronic device may include: a processor (processor) 610 , a communication interface (Communications Interface) 620 , a memory (memory) 630 and a communication bus 640 , wherein the processor 610 , the communication interface 620 , and the memory 630 communicate with each other through the communication bus 640 . The processor 610 can call the logic instructions in the memory 630 to execute the social world simulation method based on the big model, which includes: obtaining multimodal social event information, wherein the multimodal social event information includes: social event information and social spatiotemporal information; inputting the social event information into an event prompt word generator based on a deep learning model to obtain event prompt words output by the event prompt word generator; inputting the event prompt words into a pre-trained multimodal big model to obtain multimodal features of social events output by the pre-trained multimodal big model; inputting the social spatiotemporal information into a spatiotemporal prompt word generator based on a deep learning model to obtain spatiotemporal prompt words output by the spatiotemporal prompt word generator; inputting the spatiotemporal prompt words into a pre-trained big language model to obtain spatiotemporal features of social events output by the pre-trained big language model; performing feature fusion on the multimodal features of social events and the spatiotemporal features of social events through the pre-trained multimodal big model to obtain an event embedding vector; inputting the inference prompt words corresponding to the event embedding vector into the pre-trained big language model to obtain a set of simulated social events output by the pre-trained big language model.

[0117] In addition, the logic instructions in the above-mentioned memory 630 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0118] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the social world simulation method based on a large model provided by the above methods. The method includes: obtaining multimodal social event information, wherein the multimodal social event information includes: social event information and social spatiotemporal information; inputting the social event information into an event prompt word generator based on a deep learning model to obtain event prompt words output by the event prompt word generator; inputting the event prompt words into a pre-trained multimodal The pre-trained large model is used to obtain the multimodal features of social events output by the pre-trained multimodal large model; the social spatiotemporal information is input into the spatiotemporal prompt word generator based on the deep learning model to obtain the spatiotemporal prompt words output by the spatiotemporal prompt word generator; the spatiotemporal prompt words are input into the pre-trained large language model to obtain the spatiotemporal features of social events output by the pre-trained large language model; the multimodal features of social events and the spatiotemporal features of social events are fused through the pre-trained multimodal large model to obtain the event embedding vector; the inference prompt words corresponding to the event embedding vector are input into the pre-trained large language model to obtain the set of simulated social events output by the pre-trained large language model.

[0119] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the social world simulation method based on a large model provided by the above-mentioned methods, the method comprising: obtaining multimodal social event information, wherein the multimodal social event information comprises: social event information and social spatiotemporal information; inputting the social event information into an event prompt word generator based on a deep learning model to obtain event prompt words output by the event prompt word generator; inputting the event prompt words into a pre-trained multimodal large model to obtain a pre-trained multimodal large model. The model outputs the multimodal features of social events; the social spatiotemporal information is input into the spatiotemporal prompt word generator based on the deep learning model to obtain the spatiotemporal prompt words output by the spatiotemporal prompt word generator; the spatiotemporal prompt words are input into the pre-trained large language model to obtain the spatiotemporal features of social events output by the pre-trained large language model; the multimodal features of social events and the spatiotemporal features of social events are fused through the pre-trained multimodal large model to obtain the event embedding vector; the inference prompt words corresponding to the event embedding vector are input into the pre-trained large language model to obtain the set of simulated social events output by the pre-trained large language model.

[0120] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

[0121] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0122] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A social world simulation method based on a large model, characterized in that: include: Acquiring multimodal social event information, wherein the multimodal social event information includes: social event information and social spatiotemporal information; Inputting the social event information into an event prompt word generator based on a deep learning model to obtain an event prompt word output by the event prompt word generator; Inputting the event prompt words into a pre-trained multimodal large model to obtain multimodal features of social events output by the pre-trained multimodal large model; Inputting the social spatiotemporal information into a spatiotemporal prompt word generator based on a deep learning model to obtain a spatiotemporal prompt word output by the spatiotemporal prompt word generator; Inputting the spatiotemporal cue words into a pre-trained large language model to obtain spatiotemporal features of social events output by the pre-trained large language model; By using the pre-trained multimodal large model, the multimodal features of the social event and the spatiotemporal features of the social event are fused to obtain an event embedding vector; The inference prompt words corresponding to the event embedding vector are input into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model.

2. The social world simulation method based on a large model according to claim 1, characterized in that: The step of inputting the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model includes: Inputting the event embedding vector into an event reasoning model based on a long short-term memory mechanism to obtain a reasoning prompt word output by the event reasoning model; By using the pre-trained large language model, according to the reasoning-action reasoning mode, reasoning is performed based on the reasoning prompt words to obtain a set of simulated social events.

3. The social world simulation method based on a large model according to claim 2, characterized in that: After the pre-trained large language model is used to perform reasoning based on the reasoning prompt words according to the reasoning-action reasoning mode to obtain a set of simulated social events, the method further includes: Determining a contextual reasoning process for each simulated social event in the set of simulated social events; Based on a preset context-aware algorithm, an importance assessment is performed based on the contextual reasoning process of each simulated social event to determine the importance score of each simulated social event.

4. The social world simulation method based on a large model according to claim 2, characterized in that: Before inputting the event embedding vector into the event reasoning model based on the long short-term memory mechanism to obtain the reasoning prompt word output by the event reasoning model, the method further includes: Get labeled training data; Iterate the following steps until the similarity error reaches the preset value to obtain the trained event reasoning model: Inputting the training data into the event reasoning model to obtain a sample of reasoning prompt words output by the event reasoning model; Determining a similarity error between the inference hint word sample and the label of the training data; Based on the similarity error, determining the gradient of the model parameter of the event reasoning model; The model parameters are updated by a preset gradient descent algorithm to obtain an updated event reasoning model.

5. The social world simulation method based on a large model according to claim 1, characterized in that: After inputting the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, the method further includes: When there is a target simulated social event loss or a target simulated social event is lost in the simulated social event set, determining the associated social events of the target simulated social event through a graph convolutional neural network; The target simulated social event is restored based on the associated social event through a graph convolutional neural network.

6. The social world simulation method based on a large model according to claim 1, characterized in that: After inputting the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model, the method further includes: Data conversion is performed on the simulated social event set to obtain an output result in a target format, wherein the target format at least includes: a triple.

7. A social world simulation system based on a large model, characterized in that: include: An event acquisition module, used to acquire multimodal social event information, wherein the multimodal social event information includes: social event information and social time-space information; An event prompt word module, used for inputting the social event information into an event prompt word generator based on a deep learning model to obtain an event prompt word output by the event prompt word generator; A multimodal large model module, used for inputting the event prompt words into a pre-trained multimodal large model to obtain multimodal features of social events output by the pre-trained multimodal large model; A spatiotemporal prompt word module, used for inputting the social spatiotemporal information into a spatiotemporal prompt word generator based on a deep learning model to obtain the spatiotemporal prompt words output by the spatiotemporal prompt word generator; A large language model module, used for inputting the spatiotemporal prompt words into a pre-trained large language model to obtain the spatiotemporal features of social events output by the pre-trained large language model; A feature fusion module, used to perform feature fusion on the multimodal features of the social event and the spatiotemporal features of the social event through the pre-trained multimodal large model to obtain an event embedding vector; The simulation output module is used to input the inference prompt word corresponding to the event embedding vector into the pre-trained large language model to obtain a set of simulated social events output by the pre-trained large language model.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the social world simulation method based on a large model as claimed in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the social world simulation method based on a large model as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the social world simulation method based on a large model as claimed in any one of claims 1 to 6 is implemented.

Images