Emergency rescue assisting method based on large language model and computer equipment

Abstract

The invention provides an emergency rescue assisting method based on a large language model and computer equipment, and relates to the technical field of computers, and the method comprises the following steps: obtaining a current question text of a target user; performing intention recognition on the current question text to obtain a target intention type corresponding to the current question text and target intention recognition information under the target intention type; determining a target big language model matched with the target intention type from a plurality of pre-trained big language models according to the target intention type; the plurality of pre-trained large language models are pre-trained large language models of a plurality of different model frameworks in the field of emergency rescue; and by adopting the target large language model, according to the target intention identification information and a preset emergency knowledge base, generating target emergency rescue auxiliary information corresponding to the current problem text. The method does not need manual analysis, and improves the efficiency and accuracy of generating the target emergency rescue auxiliary information.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and in particular to an emergency rescue assistance method and computer equipment based on a large language model. Background Art

[0002] At present, it is crucial for grassroots staff to quickly find accurate information in daily office work and when responding to emergencies. However, the accumulation of a large number of documents often makes retrieval time-consuming and laborious, affecting the speed and effectiveness of emergency response. Grassroots staff need to deal with a large amount of daily paperwork and the document writing task is heavy, and they need to maintain professionalism and accuracy. Although existing writing assistance tools such as templates and automatic correction functions can provide help, they are often insufficient in the processing of specific official document formats and professional terms, especially the correctness check of quoted authoritative speeches, and staff need to invest a lot of time in content editing and proofreading.

[0003] Therefore, in the current emergency rescue field, when natural disasters, accidents and disasters occur, rescue teams usually rely on traditional plans and experience to develop emergency rescue plans. These plans are often prepared in advance by expert teams based on historical data and simulation situations, and have a certain degree of universality.

[0004] However, these plans may not be able to quickly adapt to various emergencies when faced with complex and ever-changing rescue scenes, which affects the rescue efficiency. In addition, due to the limited ability of the rescue team to process and analyze information in an emergency, it is difficult to generate a comprehensive and accurate emergency rescue plan in a short period of time. Summary of the invention

[0005] The purpose of this application is to address the deficiencies in the above-mentioned prior art and provide an emergency rescue assistance method and computer equipment based on a large language model, which does not require manual analysis and improves the efficiency and accuracy of generating target emergency rescue assistance information.

[0006] To achieve the above purpose, the technical solution adopted in the embodiment of the present application is as follows:

[0007] In a first aspect, an embodiment of the present application provides an emergency rescue assistance method based on a large language model, the method comprising:

[0008] Get the current question text of the target user;

[0009] Performing intent recognition on the current question text to obtain a target intent type corresponding to the current question text and target intent recognition information under the target intent type;

[0010] According to the target intent type, a target large language model matching the target intent type is determined from a plurality of pre-trained large language models; the plurality of pre-trained large language models are respectively: large language models of a plurality of different model frameworks pre-trained in the field of emergency rescue;

[0011] The target large language model is adopted to generate target emergency rescue auxiliary information corresponding to the current question text according to the target intention recognition information and a preset emergency knowledge base.

[0012] Optionally, performing intent recognition on the current question text to obtain a target intent type corresponding to the current question text and target intent recognition information under the target intent type includes:

[0013] Generate a first prompt word according to the current question text and the field branch list of the emergency rescue field; the field branch list includes: information of multiple field branches under the multiple emergency rescue fields;

[0014] Using a preset large language model, according to the first prompt word, the current question text is subjected to domain identification, so as to determine a target domain branch that matches the current question text from among the multiple domain branches;

[0015] Generate a second prompt word according to the intention type list under the target domain branch and the current question text; the intention type list includes: information of multiple preset intention types;

[0016] Using the preset large language model, according to the second prompt word, the intent recognition is performed on the current question text, so as to determine the target intent type matched by the current question text from the multiple preset intent types;

[0017] Generate a third prompt word according to the current question text and the word slot list corresponding to the target intention type; the word slot list includes: information of multiple preset word slots;

[0018] The preset large language model is used to perform semantic recognition on the current question text according to the third prompt word, and entity recognition data of the multiple preset word slots are obtained as the target intent recognition information.

[0019] Optionally, before determining, according to the target intent type, from a plurality of pre-trained large language models a target large language model that matches the target intent type, the method further comprises:

[0020] Extracting knowledge pairs from the preset emergency knowledge base to obtain a plurality of question-answer knowledge pairs in the field of emergency rescue;

[0021] According to the multiple question-answer knowledge pairs, a preset model fine-tuning mode is used to adjust the parameters of multiple basic large language models to generate model files of the multiple pre-trained large language models; the multiple basic large language models correspond to multiple different model frameworks respectively;

[0022] According to the model files of the multiple pre-trained large language models, the multiple pre-trained large language models are deployed respectively.

[0023] Optionally, the preset model fine-tuning mode is a local efficient fine-tuning mode, wherein the preset model fine-tuning mode is used to train multiple basic large language models according to the multiple question-answer knowledge pairs, so as to adjust the parameters of the multiple preset basic large language models and generate model files of the multiple pre-trained large language models, including:

[0024] According to the multiple question-answer knowledge pairs, the local efficient fine-tuning mode is used to perform initial parameter adjustment on each basic large language model to obtain an initial model file, and a preset model pruning algorithm is used to prune the initial model file;

[0025] The local efficient fine-tuning mode is adopted to fine-tune the model file after the pruning process to generate a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0026] Optionally, the preset model fine-tuning mode is a circular local aggregation mode, wherein the preset model fine-tuning mode is used to train multiple preset basic large language models according to the multiple question-answer knowledge pairs, so as to adjust the parameters of the multiple preset basic large language models and generate model files of the multiple pre-trained large language models, including:

[0027] According to the multiple question-answer knowledge pairs, the circular local aggregation mode is adopted to adjust the parameters of each preset basic large language model at multiple local nodes to obtain multiple partial model files, and the multiple partial model files are aggregated to obtain a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0028] Optionally, the method further comprises:

[0029] Evaluate each pre-trained large language model in the evaluation dimensions of multiple preset intent types to obtain evaluation parameter values ​​of each pre-trained large language model in the evaluation dimensions of the multiple preset intent types;

[0030] The step of determining, according to the target intent type, a target large language model matching the target intent type from a plurality of pre-trained large language models comprises:

[0031] According to the target intent type, a pretrained large language model with the highest evaluation parameter value under the evaluation dimension corresponding to the target intent is determined from the multiple pretrained large language models as the target large language model.

[0032] Optionally, the extracting knowledge pairs from the preset emergency knowledge base to obtain a plurality of question-answer knowledge pairs in the field of emergency rescue includes:

[0033] Generate a fourth prompt word according to the information in the emergency rescue field and the preset emergency knowledge base;

[0034] Using a preset large language model, knowledge is extracted from the preset emergency knowledge base according to the fourth prompt word to obtain the multiple question-answer knowledge pairs.

[0035] Optionally, before obtaining the current question text of the target user, the method further includes:

[0036] Building a user profile of the target user based on the historical interaction data of the target user;

[0037] According to the user portrait and the historical large language model corresponding to the historical natural language text, a preset knowledge graph embedding technology is used to map the structured semantic information of multiple preset emergency questions into a low-dimensional vector space to obtain target semantic vectors of the multiple preset emergency questions;

[0038] Determining at least one recommended emergency question for the target user according to the user portrait and the target semantic vectors of the plurality of preset emergency questions;

[0039] presenting the at least one recommended emergency question;

[0040] The step of obtaining the current question text of the target user includes:

[0041] A recommended emergency question selected by the target user from the at least one recommended emergency question is obtained as the current question text.

[0042] Optionally, the method further comprises:

[0043] Obtaining feedback information from the target user regarding the target emergency rescue plan;

[0044] According to the feedback information, parameters of the target large language model are updated.

[0045] In a second aspect, another embodiment of the present application provides an emergency rescue auxiliary device based on a large language model, the device comprising:

[0046] The acquisition module is used to obtain the current question text of the target user;

[0047] An identification module is used to perform intent identification on the current question text, obtain a target intent type corresponding to the current question text, and target intent identification information under the target intent type;

[0048] A determination module, used to determine a target large language model matching the target intent type from a plurality of pre-trained large language models according to the target intent type; the plurality of pre-trained large language models are respectively: large language models of a plurality of different model frameworks pre-trained in the field of emergency rescue;

[0049] A generation module adopts the target large language model in English to generate target emergency rescue auxiliary information corresponding to the current question text according to the target intention recognition information and a preset emergency knowledge base.

[0050] In the third aspect, another embodiment of the present application provides a computer device, comprising: a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to perform the steps of the emergency rescue assistance method based on a large language model as described in any of the first aspects above.

[0051] In a fourth aspect, another embodiment of the present application provides a storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the emergency rescue assistance method based on a large language model as described in any one of the first aspects above are executed.

[0052] The beneficial effects of this application are:

[0053] The present application provides an emergency rescue auxiliary method and computer equipment based on a large language model, which obtains the current question text of the target user, performs intent recognition on the current question text, obtains the target intent type corresponding to the current question text, and the target intent recognition information under the target intent type, and determines the target large language model that matches the target intent type from multiple pre-trained large language models according to the target intent type, and uses the target large language model to generate the target emergency rescue auxiliary information corresponding to the current question text according to the target intent recognition information and a preset emergency knowledge base. The method in the present application can determine the corresponding target emergency rescue auxiliary information for the current question text of the target user, so that the generated target emergency rescue auxiliary information is more matched with the needs of the target user, more targeted, and does not require manual analysis, thereby improving the efficiency and accuracy of generating target emergency rescue auxiliary information. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 A flowchart of an emergency rescue assistance method based on a large language model provided in an embodiment of the present application;

[0056] Figure 2 A schematic diagram of a process for determining target intention recognition information in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application;

[0057] Figure 3 A schematic diagram of a process of deploying a large language model in an emergency rescue assistance method based on a large language model provided in an embodiment of the present application;

[0058] Figure 4 A schematic diagram of a process for determining a model file in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application;

[0059] Figure 5 A schematic diagram of a process of determining a target large language model in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application;

[0060] Figure 6 A schematic diagram of a process for determining a question-answer knowledge pair in an emergency rescue assistance method based on a large language model provided in an embodiment of the present application;

[0061] Figure 7 A schematic diagram of a process for determining a target semantic vector in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application;

[0062] Figure 8 A schematic diagram of a process of updating a target large language model in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application;

[0063] Fig. 9 A schematic diagram of the structure of an emergency rescue auxiliary device based on a large language model provided in an embodiment of the present application;

[0064] Fig.10 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0065] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of explanation and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn in real proportion. The flowchart used in this application shows the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowchart can be implemented out of sequence, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art can add one or more other operations to the flowchart under the guidance of the content of the present application, or remove one or more operations from the flowchart.

[0066] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0067] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the existence of the features declared thereafter, but does not exclude the addition of other features.

[0068] Task allocation in traditional emergency rescue auxiliary methods and products still relies on manual operation, lacks intelligent system support, and cannot be optimally configured according to the actual capabilities and resource conditions of the staff. It only has the management function of preset emergency plans, which is not very practical in real emergency handling scenarios, and does not accurately recommend customer usage problems based on customer usage habits, and cannot select the most suitable large language model product based on the user's usage scenario and task type, resulting in the target emergency rescue auxiliary information obtained being not flexible enough. To this end, the present application provides an emergency rescue auxiliary method based on a large language model, which determines the corresponding target emergency rescue auxiliary information based on the current question text of the target user, so that the generated target emergency rescue auxiliary information is more matched with the needs of the target user, more targeted, and does not require manual analysis, thereby improving the efficiency and accuracy of generating target emergency rescue auxiliary information. The method provided in the embodiment of the present application is described below in conjunction with multiple drawings.

[0069] Figure 1 A flowchart of an emergency rescue assistance method based on a large language model is provided in an embodiment of the present application, such as Figure 1 As shown, the method includes:

[0070] Step 101: Obtain the current question text of the target user.

[0071] The target user may be any person such as emergency management staff, grassroots staff, and the general public, and the present embodiment of the application does not limit this. The current question text may be a text in any format such as a natural language text or a structured text, and the present embodiment of the application does not limit this.

[0072] Optionally, the current question text input by the target user through the current page is obtained, or the voice content input by the target user is converted into the current question text, or the current question text is generated according to the current business scenario of the target user, or any other method is used, and the embodiments of the present application are not limited to this.

[0073] Step 102: perform intent recognition on the current question text to obtain the target intent type corresponding to the current question text and the target intent recognition information under the target intent type.

[0074] Among them, the intent type is used to identify the specific needs or goals of the current target question text. The intent type can be in the categories of mathematics and science, language ability, moral responsibility, industry ability and comprehensive ability. The target intent recognition information is the specific information of the current question text obtained by identifying the corresponding target intent type.

[0075] Optionally, intent recognition is performed on the current question text to obtain a target intent type corresponding to the current question text, and target intent recognition information under the target intent type is determined based on the current question text and the target intent type corresponding to the current question text.

[0076] Step 103: According to the target intent type, determine a target large language model that matches the target intent type from multiple pre-trained large language models.

[0077] Among them, the multiple pre-trained large language models are: pre-trained large language models with multiple different model frameworks in the field of emergency rescue.

[0078] Optionally, according to the target intent type, multiple large language models of different model frameworks that match the target intent type are determined from multiple pre-trained large language models, and the target large language model is determined according to the sorting of the large language models of the multiple different model frameworks. The sorting of the large language models of the multiple different model frameworks can be sorting according to the loss function, or sorting according to the objective score of the model, etc., which is not limited in the embodiments of the present application.

[0079] Step 104: using the target large language model, according to the target intention recognition information and the preset emergency knowledge base, generate target emergency rescue auxiliary information corresponding to the current question text.

[0080] Among them, the preset emergency knowledge base includes: multiple emergency rescue related documents, such as emergency legal and regulatory documents, emergency plan documents, emergency management instructions and skills documents, legal and regulatory documents, etc., which are not limited in the embodiments of the present application. The target emergency rescue auxiliary information may include: emergency rescue documents and at least one content in emergency rescue knowledge, which is specifically determined according to the current question text of the target user.

[0081] Optionally, the target intention recognition information and the preset emergency knowledge base are used as the input of the target large language model, and the target emergency rescue auxiliary information corresponding to the target intention recognition information is determined from the preset emergency knowledge base through the target large language model according to the target intention recognition information.

[0082] The present application provides an emergency rescue auxiliary method based on a large language model, which obtains the current question text of the target user, performs intent recognition on the current question text, obtains the target intent type corresponding to the current question text, and the target intent recognition information under the target intent type, and determines the target large language model that matches the target intent type from multiple pre-trained large language models according to the target intent type, and uses the target large language model to generate the target emergency rescue auxiliary information corresponding to the current question text according to the target intent recognition information and a preset emergency knowledge base. The method in the present application can determine the corresponding target emergency rescue auxiliary information for the current question text of the target user, so that the generated target emergency rescue auxiliary information is more matched with the needs of the target user, more targeted, and does not require manual analysis, thereby improving the efficiency and accuracy of generating target emergency rescue auxiliary information.

[0083] Based on the above embodiments, the present application also provides a process for determining target intention recognition information in an emergency rescue auxiliary method based on a large language model. Figure 2 A schematic diagram of a process for determining target intention recognition information in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application, such as Figure 2 As shown, in the above step 102, the current question text is subjected to intent recognition to obtain the target intent type corresponding to the current question text and the target intent recognition information under the target intent type, including:

[0084] Step 201: Generate a first prompt word according to the current question text and a field branch list of the emergency rescue field.

[0085] Among them, the field branch list includes: information on multiple field branches under multiple emergency rescue fields. Multiple emergency rescue fields can be natural disaster rescue fields, accident disaster rescue fields, public health incident rescue fields, social security incident rescue fields, etc., and the embodiments of the present application do not limit this. Multiple field branches under multiple emergency rescue fields are a clearer division of the emergency rescue fields. For example, when the emergency rescue field is a natural disaster rescue field, the multiple field branches can be an earthquake rescue field, a flood rescue field, a typhoon rescue field, etc.

[0086] The first prompt word determines the target domain corresponding to the current question text from the domain branch list of the emergency rescue field.

[0087] Optionally, a first prompt word is generated according to the current question text and the field branch list of the emergency rescue field, and is used to determine the target field corresponding to the current question text from the field branch list of the emergency rescue field.

[0088] Step 202: Using a preset large language model, according to the first prompt word, perform domain identification on the current question text to determine a target domain branch that matches the current question text from multiple domain branches.

[0089] The preset large language model may be a GPT (Generative Pre-trained Transformer) model, a BERT (Bidirectional Encoder Representations from Transformers) model, or the like. The preset large language model is a model that has been pre-trained using question text and prompt words.

[0090] Optionally, a preset large language model is used, and the first prompt word and the current question text are taken as input content. The current question text is subjected to domain identification according to the first prompt word to determine a target domain branch that matches the current question text from multiple domain branches.

[0091] Step 203: Generate a second prompt word according to the intention type list under the target domain branch and the current question text.

[0092] The intent type list includes: information of multiple preset intent types. The preset intent types may be mathematical science, language ability, moral responsibility, industry ability, and comprehensive ability. The second prompt word is used to determine the intent type corresponding to the current question text from the intent type list.

[0093] Optionally, based on the list of intention types under the target domain branch and the current question text, a second prompt word is generated to determine the intention type of the current question text from the list of intention types under the target domain branch.

[0094] Step 204: Use the preset large language model to perform intent recognition on the current question text according to the second prompt word to determine the target intent type that matches the current question text from multiple preset intent types.

[0095] Optionally, a preset large language model is used, and the second prompt word and the current question text are used as input content. The current question text is identified according to the second prompt word to determine the target intent type that matches the current question text from multiple preset intent types. If the user instruction is not in the given intent list, an empty list is output and the error handler is turned to. That is, when the target intent type cannot be obtained through the preset large language model, it can be processed manually to obtain the target intent type.

[0096] Step 205: Generate a third prompt word according to the current question text and the word slot list corresponding to the target intent type.

[0097] The word slot list includes: information of a plurality of preset word slots. The third prompt word is used to extract specific information corresponding to the current question from the target intention type.

[0098] Optionally, a third prompt word is generated based on the current question text and the word slot list corresponding to the target intent type, and specific information corresponding to the current question text is determined from the word slot list corresponding to the target intent type through the third prompt word.

[0099] Step 206: Using a preset large language model, perform semantic recognition on the current question text according to the third prompt word, and obtain entity recognition data of multiple preset word slots as target intent recognition information.

[0100] Optionally, a preset large language model is used, and the third prompt word and the current question text are used as input content. Semantic recognition is performed on the current question text according to the third prompt word to obtain entity recognition data of multiple preset word slots as target intent recognition information.

[0101] For example, the target intent identification information can be {"domain identification":"domain name","intent identification":"intent type","slot extraction":{"entity type 1":"entity content 1","entity type 2":"entity content 2","entity type 3":"entity content 3"}}, that is, the target intent identification information includes information on target domain branches, feasible target intents, and multiple slots.

[0102] The present application sets multiple prompt words to respectively determine the target domain branch, target intent type and target intent recognition information of the current question text, so as to make the recognition of the current question text more accurate, thereby improving the accuracy of the target intent recognition information determined according to the current question text.

[0103] Based on the above embodiments, the present application also provides a process for deploying a large language model in an emergency rescue auxiliary method based on a large language model. Figure 3 A flow chart of deploying a large language model in an emergency rescue assistance method based on a large language model provided in an embodiment of the present application is as follows: Figure 3 As shown, before the above step 103 determines a target large language model matching the target intent type from a plurality of pre-trained large language models according to the target intent type, the method further includes:

[0104] Step 301: extract knowledge pairs from a preset emergency knowledge base to obtain multiple question-answer knowledge pairs in the field of emergency rescue.

[0105] Among them, question-answering knowledge pairs are used to train the large language model, so that the large language model can correctly understand the questions and give corresponding answers.

[0106] Step 302: According to multiple question-answer knowledge pairs, a preset model fine-tuning mode is used to adjust the parameters of multiple basic large language models to generate model files of multiple pre-trained large language models.

[0107] Multiple basic large language models correspond to multiple different model frameworks.

[0108] Among them, the preset model fine-tuning mode can be one or both of the local high-efficiency fine-tuning mode and the ring local aggregation mode.

[0109] Optionally, multiple question-answer knowledge pairs are converted into the formats required by multiple basic large language models, and the multiple basic large language models are fine-tuned using a preset model according to the training results to adjust the parameters of the multiple basic large language models, and generate model files of multiple pre-trained large language models. Parameter adjustment includes adjusting the model layers, loss functions, optimizers, and learning rate scheduling of the multiple basic large language models, and this application does not impose any restrictions on this.

[0110] Step 303: Deploy multiple pre-trained large language models respectively according to the model files of the multiple pre-trained large language models.

[0111] Optionally, according to the model files of the multiple pre-trained large language models, the model files of the pre-trained large language models are loaded using a corresponding deep learning framework, thereby respectively deploying the multiple pre-trained large language models.

[0112] In the embodiment of the present application, multiple basic large language models are adjusted by using a preset model fine-tuning mode through multiple question-answer knowledge pairs, and model files of multiple pre-trained large language models are generated, thereby realizing the deployment of multiple pre-trained large language models. This makes the multiple pre-trained large language models more compatible with the current application scenario, thereby improving the accuracy of the model output content and the processing efficiency of the model.

[0113] Based on the above embodiment, the preset model fine-tuning mode is: local efficient fine-tuning mode. To this end, the present application also provides a process for determining a model file in an emergency rescue auxiliary method based on a large language model. Figure 4 A schematic diagram of a flow chart of determining a model file in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application, such as Figure 4 As shown, in the above step 302, according to multiple question-answer knowledge pairs, a preset model fine-tuning mode is used to adjust the parameters of multiple basic large language models to generate model files of multiple pre-trained large language models, including:

[0114] Step 401: Based on multiple question-answer knowledge pairs, a local efficient fine-tuning mode is used to perform initial parameter adjustment on each basic large language model to obtain an initial model file, and a preset model pruning algorithm is used to prune the initial model file.

[0115] Among them, the local efficient fine-tuning mode is a mode that adjusts the parameters of multiple basic large language models in the local environment while ensuring model performance. The model pruning algorithm is used to reduce the model size and computational complexity while ensuring model performance. The model pruning algorithm can remove unimportant weights or neurons in the model to obtain a more compact and efficient model.

[0116] Optionally, for each basic large language model, according to multiple question-answer knowledge pairs, and according to the training results, a local efficient fine-tuning mode is used to perform initial parameter adjustment on each basic large language model to obtain an initial model file. At the same time, a preset model pruning algorithm is used to prune the initial model file to reduce the model size and computational complexity while ensuring model performance.

[0117] Step 402: Use a local efficient fine-tuning mode to fine-tune the model file after the pruning process to generate a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0118] Optionally, a local efficient fine-tuning mode is used to fine-tune the pruned model file so that the model file meets the needs of the current application scenario, and the adjusted model file is used as a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0119] In the embodiment of the present application, according to multiple question-answer knowledge pairs, a local efficient fine-tuning mode is adopted to perform initial parameter adjustment on each basic large language model to obtain an initial model file, and a preset model pruning algorithm is used to prune the initial model file, thereby adopting a local efficient fine-tuning mode to fine-tune the pruned model file to generate a model file of a pre-trained large language model corresponding to each preset basic large language model. The generated model file can maintain the characteristics of the large model, while significantly reducing the size of the model, reducing the cost of storing and transmitting the model, and accelerating the reasoning speed of the model.

[0120] On the basis of the above embodiment, the preset model fine-tuning mode is: ring local aggregation mode. To this end, the embodiment of the present application also provides another process for determining the model file in the emergency rescue auxiliary method based on the large language model. In the above step 302, according to multiple question-answer knowledge pairs, the preset model fine-tuning mode is used to adjust the parameters of multiple basic large language models to generate model files of multiple pre-trained large language models, including:

[0121] According to multiple question-answer knowledge pairs, a circular local aggregation mode is adopted to adjust the parameters of each preset basic large language model at multiple local nodes to obtain multiple partial model files, and then the multiple partial model files are aggregated to obtain a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0122] Among them, the ring local aggregation mode is a closed-loop distributed system in which each computing node only exchanges gradient information with its adjacent nodes. By locally calculating and gradually aggregating the gradients from each node in the ring, the model parameters can be updated efficiently to achieve optimization of the model on specific tasks.

[0123] In an embodiment of the present application, a ring-shaped local aggregation mode is adopted to adjust the parameters of each preset basic large language model at multiple local nodes to obtain a model file, thereby improving the fault tolerance and stability of the large language model and ensuring the connectivity of the large language model.

[0124] Based on the above embodiments, the present application also provides a process for determining a target large language model in an emergency rescue auxiliary method based on a large language model. Figure 5 A schematic diagram of a process of determining a target large language model in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application. Based on the above steps 301 to 303, the method further includes:

[0125] Step 501: Evaluate each pre-trained large language model in the evaluation dimensions of multiple preset intent types to obtain evaluation parameter values ​​of each pre-trained large language model in the evaluation dimensions of multiple preset intent types.

[0126] Among them, multiple preset intent types can be mathematical science, language ability, moral responsibility, industry ability and comprehensive ability. When the preset intent type is mathematical science, the evaluation dimensions can be mathematics, common sense, science, logical reasoning and other dimensions; when the preset intent type is language ability, the evaluation dimensions can be analogy transfer, text generation, intention understanding, role playing, open logical reasoning and other dimensions; when the preset intent type is moral responsibility, the evaluation dimensions can be risk information, model robustness, biased information, knowledge illusion and other dimensions; when the preset intent type is industry ability, the evaluation dimensions can be financial industry, government industry, e-commerce industry, medical industry, other industries and other dimensions; when the preset intent type is comprehensive ability, the evaluation dimensions can be artificial intelligence Chinese agent, professional text, super long text and other dimensions. The evaluation method can be objective evaluation and subjective evaluation. The objective evaluation can be evaluated according to the preset large language model specification and the actual situation of the large language model. The subjective evaluation can include customer evaluation and expert evaluation. Customer evaluation is the evaluation by customers based on actual usage, and expert evaluation is the evaluation by people with professional knowledge and experience based on the usage of the large language model.

[0127] Optionally, each pre-trained large language model is evaluated in the evaluation dimensions of multiple preset intent types through subjective evaluation and objective evaluation, and the subjective evaluation and the objective evaluation are weighted to obtain the evaluation parameter values ​​of each pre-trained large language model in the evaluation dimensions of multiple preset intent types.

[0128] In the above step 103, according to the target intent type, a target large language model matching the target intent type is determined from multiple pre-trained large language models, including:

[0129] Step 502: According to the target intent type, determine from multiple pre-trained large language models a pre-trained large language model with the highest evaluation parameter value under the evaluation dimension corresponding to the target intent as the target large language model.

[0130] Optionally, multiple pre-trained large language models under evaluation dimensions corresponding to the target intent are determined from multiple pre-trained large language models according to the target intent type, and a pre-trained large language model with the highest evaluation parameter value among the multiple pre-trained large language models is obtained as the target large language model.

[0131] In an embodiment of the present application, each pre-trained large language model is evaluated to obtain evaluation parameter values ​​of each pre-trained large language model under evaluation dimensions of multiple preset intent types. And according to the intent type, the pre-trained large language model with the highest evaluation parameter value under the evaluation dimension corresponding to the target intent is determined from multiple pre-trained large language models as the target large language model. The present application can ensure that the target large language model better meets the needs of users and improves user satisfaction, while providing more relevant responses, ensuring that resources are effectively utilized, and reducing resource waste.

[0132] Based on the above embodiments, the present application also provides a process for determining question-answer knowledge pairs in an emergency rescue assistance method based on a large language model. Figure 6 A flowchart of determining a question-answer knowledge pair in an emergency rescue assistance method based on a large language model provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, in the above step 301, knowledge pairs are extracted from the preset emergency knowledge base to obtain multiple question-answer knowledge pairs in the field of emergency rescue, including:

[0133] Step 601: Generate a fourth prompt word based on information in the field of emergency rescue and a preset emergency knowledge base.

[0134] Among them, the preset emergency knowledge base includes: multiple emergency rescue related documents, for example, emergency legal and regulatory documents, emergency plan documents, emergency management instructions and skills documents, legal and regulatory documents, etc., which are not limited in this embodiment of the present application. Emergency rescue field information can be in the field of natural disaster rescue, accident disaster rescue, public health incident rescue, social security incident rescue, etc., which are not limited in this embodiment of the present application. The fourth prompt word is used to extract question-answer knowledge pairs, and the fourth prompt word includes emergency rescue field information.

[0135] Optionally, a fourth prompt word is generated according to information in the field of emergency rescue and a preset emergency knowledge base, and the fourth prompt word is used to extract question-answer knowledge pairs, and the fourth prompt word includes information in the field of emergency rescue.

[0136] Step 602: Using a preset large language model, and based on the fourth prompt word, extract knowledge from a preset emergency knowledge base to obtain a plurality of question-answer knowledge pairs.

[0137] Among them, the preset large language model can be a Qwen2-7B-Instruct model or a Baichuan2-7B-Chat model, etc., and the embodiment of the present application does not limit this.

[0138] For example, by presetting the fourth prompt word, a preset large language model can be used to extract knowledge pairs from the preset emergency knowledge base to obtain multiple question-answer knowledge pairs in the field of emergency rescue. The result of the question-answer knowledge pair can be {"conversation":[{"system":"You are a knowledge expert in the field of emergency management. You are very familiar with relevant coal mine safety regulations, procedures, systems, technologies and other documents. Please answer the coal mine safety-related questions that users want to ask professionally and correctly.","input":"……","output":"……"}]}, where system is used to indicate the preset fourth prompt word, the content corresponding to input is the question, and the question is generated based on the fourth prompt word, and the content corresponding to output is the answer to the question, and the answer to the question is the content extracted from the preset emergency knowledge base.

[0139] In the embodiment of the present application, a fourth prompt word is generated based on information in the field of emergency rescue and a preset emergency knowledge base, and a preset large language model is used to extract knowledge from the preset emergency knowledge base based on the fourth prompt word to obtain multiple question-answer knowledge pairs. The present application can more effectively retrieve question-answer knowledge pairs related to specific situations from a large number of emergency knowledge bases, provide more accurate question-answer knowledge pairs, and reduce the interference of irrelevant information.

[0140] Based on the above embodiments, the present application also provides a process for determining a target semantic vector in an emergency rescue auxiliary method based on a large language model. Figure 7 A flowchart of determining a target semantic vector in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, before the above step 101 obtains the current question text of the target user, the method further includes:

[0141] Step 701: Build a user profile of the target user based on the historical interaction data of the target user.

[0142] Among them, the historical interaction data includes the historical question text of the target user and the emergency rescue auxiliary information corresponding to the historical question text. The historical interaction data of the target user also includes: basic information of the target user, for example, the target user's occupational name, the target user's occupational level, the target user's work unit and other target user-related information, which is not limited in the embodiment of the present application. User portraits are used to abstract various user information into analyzable models to better understand and predict user behavior and needs.

[0143] Optionally, the usage preferences of the target user are obtained based on the historical interaction data of the target user, and a user profile of the target user is constructed based on the usage preferences of the target user.

[0144] Step 702: Based on the user portrait and the historical large language model corresponding to the historical natural language text, the preset knowledge graph embedding technology is used to map the structured semantic information of multiple preset emergency questions into a low-dimensional vector space to obtain target semantic vectors of multiple preset emergency questions.

[0145] Among them, knowledge graph embedding technology is used to convert entities and relations in the knowledge graph into continuous vectors.

[0146] Optionally, based on the user portrait and the historical large language model corresponding to the historical natural language text, multiple preset emergency questions are constructed, and the structured semantic information of the multiple preset emergency questions is mapped to a low-dimensional vector space using a preset knowledge graph embedding technology to obtain target semantic vectors of the multiple preset emergency questions.

[0147] Step 703: Determine at least one recommended emergency question for the target user based on the user portrait and the target semantic vectors of the plurality of preset emergency questions.

[0148] Among them, the recommended emergency questions can be nine, ten or any number, which is not limited in the embodiment of the present application and is determined according to the actual usage scenario and target users.

[0149] Optionally, key features of the target user are extracted based on the user portrait, and semantic vectors of the key features of the target user are determined. Based on the semantic vectors of the key features of the target user and target semantic vectors of multiple preset emergency questions, at least one recommended emergency question for the target user is determined.

[0150] Step 704: Display at least one recommended emergency question.

[0151] In the above step 101, the current question text of the target user is obtained, including:

[0152] Step 705: Obtain a recommended emergency question selected by the target user from at least one recommended emergency question as the current question text.

[0153] Optionally, a recommended emergency question selected by the target user from at least one recommended emergency question is obtained as the current question text. The current question text can be determined according to the current application scenario, or according to the target user's selection, or in any other way. The embodiment of the present application is not limited to this.

[0154] In the embodiment of the present application, the user profile of the target user is used to determine at least one recommended emergency question for the target user, and the current question text is obtained according to the at least one recommended emergency question. The present application can provide a more personalized service that meets user expectations, thereby improving user satisfaction and experience.

[0155] Based on the above embodiments, the present application also provides a process for updating a target large language model in an emergency rescue auxiliary method based on a large language model. Figure 8 A schematic diagram of a flow chart of updating a target large language model in an emergency rescue auxiliary method based on a large language model provided in an embodiment of the present application. Based on the above steps 101 to 104, the method further includes:

[0156] Step 801: Obtain feedback information from a target user regarding a target emergency rescue plan.

[0157] Among them, feedback information includes positive feedback and negative feedback. Positive feedback may include accurate and professional answers, clear and easy-to-understand answers, and fast response speed. Negative feedback may include inaccurate generated content, monotonous content lacking in innovation, semantic and grammatical errors, incoherent logic, slow generation of content involving privacy and security, and lengthy / too concise generated content.

[0158] Optionally, the target user provides feedback on the target emergency rescue plan to obtain feedback information of the target user. The feedback of the target user can be provided in the form of a score combined with description information. The feedback information of the target user is determined according to the score and the description information.

[0159] Step 802: Update the parameters of the target large language model according to the feedback information.

[0160] Optionally, when the feedback information is positive feedback, there is no need to update the parameters of the target large language model; when the feedback information is negative feedback, the parameters of the target large language model are adjusted according to the description information in the target user feedback information, so that the target large language model meets the user's needs.

[0161] In the embodiment of the present application, by obtaining the target user's feedback information on the target emergency rescue plan, the parameters of the target large language model are updated according to the feedback information. The present application can adjust the model parameters in a targeted manner, improve the performance of the model on specific tasks, and enable the model to provide more personalized and accurate solutions.

[0162] Based on the same inventive concept, the embodiments of the present application also provide an emergency rescue assistance device based on a large language model corresponding to the emergency rescue assistance method based on a large language model. Since the principle of solving the problem by the device in the embodiments of the present application is similar to the above-mentioned emergency rescue assistance method based on a large language model in the embodiments of the present application, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be repeated.

[0163] Fig. 9 A schematic diagram of the structure of an emergency rescue auxiliary device based on a large language model provided in an embodiment of the present application, the device includes:

[0164] An acquisition module 901 is used to acquire the current question text of the target user;

[0165] The recognition module 902 is used to perform intent recognition on the current question text, obtain the target intent type corresponding to the current question text, and the target intent recognition information under the target intent type;

[0166] The determination module 903 is used to determine a target large language model matching the target intent type from a plurality of pre-trained large language models according to the target intent type; the plurality of pre-trained large language models are respectively: large language models of a plurality of different model frameworks pre-trained in the field of emergency rescue;

[0167] The generation module 904 uses the target large language model in English to generate target emergency rescue auxiliary information corresponding to the current question text according to the target intention recognition information and the preset emergency knowledge base.

[0168] Optionally, the identification module 902 is specifically used to: generate a first prompt word according to the current question text and a field branch list of the emergency rescue field; the field branch list includes: information of multiple field branches under multiple emergency rescue fields;

[0169] Using a preset large language model, according to the first prompt word, the current question text is subjected to domain recognition, so as to determine a target domain branch that matches the current question text from among multiple domain branches;

[0170] Generate a second prompt word according to the intention type list under the target domain branch and the current question text; the intention type list includes: information of multiple preset intention types;

[0171] Using a preset large language model, according to the second prompt word, the intent of the current question text is identified to determine the target intent type that matches the current question text from multiple preset intent types;

[0172] Generate a third prompt word according to the current question text and the word slot list corresponding to the target intention type; the word slot list includes: information of multiple preset word slots;

[0173] Using a preset large language model, semantic recognition is performed on the current question text according to the third prompt word, and entity recognition data of multiple preset word slots are obtained as target intent recognition information.

[0174] Optionally, the determination module 903 is further used to: extract knowledge pairs from a preset emergency knowledge base to obtain a plurality of question-answer knowledge pairs in the field of emergency rescue;

[0175] According to multiple question-answer knowledge pairs, a preset model fine-tuning mode is used to adjust the parameters of multiple basic large language models to generate model files of multiple pre-trained large language models; the multiple basic large language models correspond to multiple different model frameworks respectively;

[0176] According to the model files of multiple pre-trained large language models, deploy multiple pre-trained large language models respectively.

[0177] Optionally, the preset model fine-tuning mode is: local efficient fine-tuning mode, and the determination module 903 is specifically used to: according to multiple question-answer knowledge pairs, use the local efficient fine-tuning mode to perform initial parameter adjustment on each basic large language model to obtain an initial model file, and use the preset model pruning algorithm to prune the initial model file;

[0178] The local efficient fine-tuning mode is used to fine-tune the pruned model file to generate a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0179] Optionally, the preset model fine-tuning mode is: a circular local aggregation mode, and the determination module 903 is specifically used for: according to multiple question-answer knowledge pairs, adopting a circular local aggregation mode, adjusting parameters of each preset basic large language model at multiple local nodes to obtain multiple partial model files, and aggregating the multiple partial model files to obtain a model file of a pre-trained large language model corresponding to each preset basic large language model.

[0180] Optionally, the determination module 903 is further used to: evaluate each pre-trained large language model in the evaluation dimensions of multiple preset intent types to obtain evaluation parameter values ​​of each pre-trained large language model in the evaluation dimensions of multiple preset intent types;

[0181] According to the target intent type, a target large language model matching the target intent type is determined from multiple pre-trained large language models, including:

[0182] According to the target intent type, a pre-trained large language model with the highest evaluation parameter value under the evaluation dimension corresponding to the target intent is determined from multiple pre-trained large language models as the target large language model.

[0183] Optionally, the determination module 903 is specifically configured to: generate a fourth prompt word according to information in the field of emergency rescue and a preset emergency knowledge base;

[0184] Using a preset large language model, knowledge is extracted from a preset emergency knowledge base according to the fourth prompt word to obtain multiple question-answer knowledge pairs.

[0185] Optionally, the device further includes: a construction module, specifically used to: construct a user portrait of the target user according to the historical interaction data of the target user;

[0186] According to the user portrait and the historical large language model corresponding to the historical natural language text, the preset knowledge graph embedding technology is used to map the structured semantic information of multiple preset emergency questions into a low-dimensional vector space to obtain the target semantic vectors of multiple preset emergency questions;

[0187] Determine at least one recommended emergency question for a target user according to the user portrait and target semantic vectors of multiple preset emergency questions;

[0188] Demonstrate at least one recommended emergency question;

[0189] Optionally, the acquisition module 901 is specifically used to: acquire a recommended emergency question selected by the target user from at least one recommended emergency question as the current question text.

[0190] Optionally, the device further includes: a feedback module, specifically used to: obtain feedback information of the target user on the target emergency rescue plan;

[0191] Based on the feedback information, the parameters of the target large language model are updated.

[0192] For descriptions of the processing flow of each module in the device and the interaction flow between each module, reference may be made to the relevant descriptions in the above method embodiment, which will not be described in detail here.

[0193] The present application also provides a computer device, Fig.10 A schematic diagram of the structure of a computer device provided in an embodiment of the present application includes: a processor 1001, a memory 1002, and optionally, a bus 1003. The memory 1002 stores machine-readable instructions executable by the processor 1001. When the computer device is running, the processor 1001 communicates with the memory 1002 through the bus 1003. When the machine-readable instructions are executed by the processor 1001, the steps of the emergency rescue auxiliary method based on the large language model are executed.

[0194] An embodiment of the present application also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned emergency rescue assistance method based on a large language model are executed.

[0195] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, the specific working process of the system and device described above can refer to the corresponding process in the method embodiment, and will not be repeated in this application. In the several embodiments provided in this application, it should be understood that the disclosed system, device and method can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation. For example, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0196] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or part of the technical solution that contributes to the prior art or 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 several 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, RandomAccess Memory), disk or optical disk and other media that can store program code.

[0197] The above are only specific implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be covered by the protection scope of the present application.

Claims

1. An emergency rescue assistance method based on a large language model, characterized in that: The method comprises: Get the current question text of the target user; Performing intent recognition on the current question text to obtain a target intent type corresponding to the current question text and target intent recognition information under the target intent type; According to the target intent type, a target large language model matching the target intent type is determined from a plurality of pre-trained large language models; the plurality of pre-trained large language models are respectively: large language models of a plurality of different model frameworks pre-trained in the field of emergency rescue; The target large language model is adopted to generate target emergency rescue auxiliary information corresponding to the current question text according to the target intention recognition information and a preset emergency knowledge base.

2. The method according to claim 1, characterized in that The performing of intent recognition on the current question text to obtain a target intent type corresponding to the current question text and target intent recognition information under the target intent type includes: Generate a first prompt word according to the current question text and the field branch list of the emergency rescue field; the field branch list includes: information of multiple field branches under the multiple emergency rescue fields; Using a preset large language model, according to the first prompt word, the current question text is subjected to domain identification, so as to determine a target domain branch that matches the current question text from among the multiple domain branches; Generate a second prompt word according to the intention type list under the target domain branch and the current question text; the intention type list includes: information of multiple preset intention types; Using the preset large language model, according to the second prompt word, the intent recognition is performed on the current question text, so as to determine the target intent type matched by the current question text from the multiple preset intent types; Generate a third prompt word according to the current question text and the word slot list corresponding to the target intention type; the word slot list includes: information of multiple preset word slots; The preset large language model is used to perform semantic recognition on the current question text according to the third prompt word, and entity recognition data of the multiple preset word slots are obtained as the target intent recognition information.

3. The method according to claim 1, characterized in that Before determining, according to the target intent type, a target large language model matching the target intent type from a plurality of pre-trained large language models, the method further includes: Extracting knowledge pairs from the preset emergency knowledge base to obtain a plurality of question-answer knowledge pairs in the field of emergency rescue; According to the multiple question-answer knowledge pairs, a preset model fine-tuning mode is used to adjust the parameters of multiple basic large language models to generate model files of the multiple pre-trained large language models; the multiple basic large language models correspond to multiple different model frameworks respectively; According to the model files of the multiple pre-trained large language models, the multiple pre-trained large language models are deployed respectively.

4. The method according to claim 3, characterized in that The preset model fine-tuning mode is a local efficient fine-tuning mode, wherein the preset model fine-tuning mode is used to train multiple basic large language models according to the multiple question-answer knowledge pairs, so as to adjust the parameters of the multiple preset basic large language models and generate model files of the multiple pre-trained large language models, including: According to the multiple question-answer knowledge pairs, the local efficient fine-tuning mode is used to perform initial parameter adjustment on each basic large language model to obtain an initial model file, and a preset model pruning algorithm is used to prune the initial model file; The local efficient fine-tuning mode is adopted to fine-tune the model file after the pruning process to generate a model file of a pre-trained large language model corresponding to each preset basic large language model.

5. The method according to claim 3, characterized in that: The preset model fine-tuning mode is a circular local aggregation mode, wherein the preset model fine-tuning mode is used to train multiple preset basic large language models according to the multiple question-answer knowledge pairs, so as to adjust the parameters of the multiple preset basic large language models and generate model files of the multiple pre-trained large language models, including: According to the multiple question-answer knowledge pairs, the circular local aggregation mode is adopted to adjust the parameters of each preset basic large language model at multiple local nodes to obtain multiple partial model files, and the multiple partial model files are aggregated to obtain a model file of a pre-trained large language model corresponding to each preset basic large language model.

6. The method according to claim 3, characterized in that The method further comprises: Evaluate each pre-trained large language model in the evaluation dimensions of multiple preset intent types to obtain evaluation parameter values ​​of each pre-trained large language model in the evaluation dimensions of the multiple preset intent types; The step of determining, according to the target intent type, a target large language model matching the target intent type from a plurality of pre-trained large language models comprises: According to the target intent type, a pretrained large language model with the highest evaluation parameter value under the evaluation dimension corresponding to the target intent is determined from the multiple pretrained large language models as the target large language model.

7. The method according to claim 3, characterized in that The extracting of knowledge pairs from the preset emergency knowledge base to obtain a plurality of question-answer knowledge pairs in the field of emergency rescue includes: Generate a fourth prompt word according to the information in the emergency rescue field and the preset emergency knowledge base; Using a preset large language model, knowledge is extracted from the preset emergency knowledge base according to the fourth prompt word to obtain the multiple question-answer knowledge pairs.

8. The method according to claim 1, characterized in that Before obtaining the current question text of the target user, the method further includes: Building a user profile of the target user based on the historical interaction data of the target user; According to the user portrait and the historical large language model corresponding to the historical natural language text, a preset knowledge graph embedding technology is used to map the structured semantic information of multiple preset emergency questions into a low-dimensional vector space to obtain target semantic vectors of the multiple preset emergency questions; Determining at least one recommended emergency question for the target user according to the user portrait and the target semantic vectors of the plurality of preset emergency questions; presenting the at least one recommended emergency question; The step of obtaining the current question text of the target user includes: A recommended emergency question selected by the target user from the at least one recommended emergency question is obtained as the current question text.

9. The method according to claim 1, characterized in that: The method further comprises: Obtaining feedback information from the target user regarding the target emergency rescue plan; According to the feedback information, parameters of the target large language model are updated.

10. An electronic device, characterized in that: include: A processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor executes the machine-readable instructions to perform the steps of the emergency rescue assistance method based on a large language model as described in any one of claims 1 to 9.

Images