Disaster reporting method, electronic equipment and computer readable storage medium

Abstract

The invention discloses a disaster reporting method, electronic equipment and a computer readable storage medium, and relates to the field of disaster information collection, and the method comprises the steps: collecting original dialogue information through a dialogue agent; the dialogue agent preprocesses the original dialogue information to obtain target dialogue information; the information management agent is controlled to receive the target dialogue information in real time, feature extraction is carried out on the target dialogue information, and disaster information is obtained; respectively storing the disaster information and all the original dialogue information in a database; and controlling the information management agent to generate a report according to the disaster information. According to the application, by introducing the user agent and the disaster information management agent, multi-modal data acquisition, information extraction and multi-round interaction are realized. The system can automatically identify the type, the place and the severity of the disaster situation, and quickly screen and integrate key elements. By means of a multi-agent framework, a dialogue strategy and a data analysis process can be dynamically adjusted, the integrity and accuracy of information are ensured, and support is provided for disaster reporting and rescue decision making.

Description

Technical Field

[0001] The present application relates to the field of disaster information collection, and in particular to a disaster reporting method, an electronic device, and a computer-readable storage medium. Background Art

[0002] Climate change has led to an increase in extreme weather events, such as floods, hurricanes, earthquakes, and forest fires. These disasters often cause serious casualties and economic losses, so it is crucial to quickly obtain and process disaster information. When a disaster occurs, the public will spontaneously generate a large number of help messages or publish appeals, which include real-time reports from witnesses, resource requirements and supply conditions, etc. The real-time nature and rich diversity of this information provide valuable intelligence for emergency response and decision-making. However, how to effectively screen, analyze, and utilize this large amount of disorganized data has become a major challenge.

[0003] However, many places currently rely on traditional manual methods for disaster reporting. Typically, on-site personnel need to observe and collect information, and then report the information through telephone, social media tools, etc.; or report the disaster to relevant departments through hotlines and broadcasting systems staffed by dedicated personnel. This traditional method is still effective in some cases, but its efficiency is often limited by various factors. Manual reporting may cause delays due to the long chain of information transmission, or inaccurate information due to human errors. In addition, when the disaster is large in scale or involves multiple locations, the burden and pressure of manual information processing will also increase significantly.

[0004] In this context, the introduction of automated information collection and reporting systems is particularly critical. Some systems are based on sensor networks, drones, and satellite remote sensing technology, which can obtain disaster data as soon as a disaster occurs. For example, drones can quickly cover a wide range of disaster-stricken areas, capture high-resolution images and videos, and help identify the scope and extent of the disaster. Sensors can monitor key environmental indicators such as water levels, temperature, and air quality in real time, and upload data directly to cloud platforms for analysis and sharing. However, this type of deployment of sensing equipment has limited coverage and flexibility, making it difficult to respond to real-time, sudden disasters. Summary of the invention

[0005] The present application aims to solve at least one of the technical problems existing in the prior art. To this end, the present application proposes a disaster reporting method, an electronic device and a computer-readable storage medium, which can autonomously interact with users in multiple rounds, thereby comprehensively and accurately collecting disaster information.

[0006] In a first aspect, the present application provides a disaster reporting method, comprising: Collecting original dialogue information through the dialogue agent; Controlling the dialogue agent to pre-process the original dialogue information to obtain target dialogue information, and sending the target dialogue information to the information management agent in real time; Controlling the information management agent to receive the target dialogue information in real time, and extracting features from the target dialogue information to obtain disaster information; storing the disaster information and all the original conversation information in a first database and a second database respectively; Control the information management agent to generate a report based on the disaster information and send the report.

[0007] According to the disaster reporting method of the embodiment of the present application, at least the following beneficial effects are achieved: by setting up a user dialogue agent and a disaster information management agent respectively, the dialogue interaction and information extraction are divided and conquered: the former focuses on the semantic understanding and guidance of multiple rounds of dialogues, ensuring timely capture of multimodal content such as voice, text, image and location information in user reports; the latter further summarizes and analyzes this information, automatically performs entity recognition, relationship extraction, data deduplication and structured processing, and then stores the processing results in a structured database or an unstructured database, thereby providing comprehensive and accurate data support for emergency management personnel or subsequent analysis. Through such close collaboration of multiple agents, on the one hand, it is possible to track and supplement the disaster elements that may be missing or ambiguous in the dialogue in real time; on the other hand, on the basis of dynamic dialogue management, the core tasks of the dialogue are kept concentrated, effectively avoiding dialogue digression and information redundancy. This adaptive, highly fault-tolerant, multimodal information acquisition and processing method can greatly improve the efficiency of emergency response and reduce casualties and property losses caused by disasters.

[0008] According to some embodiments of the present application, collecting original dialogue information through the dialogue agent includes: Controlling the dialog agent to generate question information using a preset large language model, and obtaining response information corresponding to the question information; Controlling the dialog agent to generate new question information using the large language model according to the previous response information; Control the dialogue agent to generate the original dialogue information in real time according to the question information and the response information.

[0009] According to some embodiments of the present application, the controlling the information management agent to receive the target dialogue information and extracting features from the target dialogue information to obtain disaster information, further comprising: The information management agent is controlled to send feedback information to the dialogue agent in real time, and the dialogue agent adjusts the response strategy and the question information in real time according to the feedback information; wherein the feedback information is obtained according to the disaster information.

[0010] According to some embodiments of the present application, controlling the dialogue agent to pre-process the original dialogue information to obtain target dialogue information, and sending the target dialogue information to the information management agent in real time includes: Controlling the dialog agent to calculate the repetition of the original dialog information and the relevance to the disaster report topic in real time using the large language model; Determine the original conversation information whose repetition degree is higher than a preset first threshold as repeated information, and determine the original conversation information whose relevance degree is lower than a preset second threshold as irrelevant information; Merging the repeated information in the original conversation information and removing the irrelevant information in the original conversation information to obtain target conversation information; The target dialogue information is sent to the information management agent in real time.

[0011] According to some embodiments of the present application, the controlling the information management agent to receive the target dialogue information and extracting features from the target dialogue information to obtain disaster information includes: Feature extraction of target dialogue information based on manually configured disaster templates to obtain event type, location, time, severity, and resource demand information; Disaster information is obtained by combining the event type, location, time, severity and resource requirements.

[0012] According to some embodiments of the present application, the controlling the information management agent to send feedback information to the dialogue agent in real time, and the dialogue agent adjusting the answering strategy and the question information in real time according to the feedback information, includes: Controlling the reflection module of the information management agent to compare the disaster information with a preset feature information template, determining whether the disaster information has missing information, repeated information, and potential conflicts, and obtaining a comparison result; The reflection module controlling the information management agent analyzes the missing elements and elements to be clarified in the current target dialogue information according to the comparison result, and generates feedback information; The feedback information is sent to the dialogue agent in real time, and the dialogue agent adjusts the response strategy and the question information for the user according to the feedback information.

[0013] According to some embodiments of the present application, the controlling the dialog agent to generate new question information using the large language model according to the previous response information includes: Controlling the dialog agent to receive the response information of the previous round, wherein the response information includes one or more of text, voice, picture, video and location information; Using the large language model, the text is semantically analyzed, the voice information is transcribed and semantically recognized, the image and video information is identified and features are extracted, and the location information is analyzed for geographic location, so as to obtain composite interaction data; The large language model is controlled to generate new question information according to the composite interaction data.

[0014] According to some embodiments of the present application, the training method of the large language model includes: Get the original dialogue dataset; De-noising, de-duplication and disaster scene matching screening are performed on the original dialogue data set to obtain a first data set; Annotating the first data set to obtain annotation information, and performing structured processing on the text, voice, picture, video and location information to obtain structured information, and merging the annotation information and the structured information into a second data set; Expanding the second data set to obtain an augmented data set; Using the quality scoring module of the large language model to perform quality assessment on the augmented data set, filter out low-quality data, and obtain a training data set; The training data set is input into a large language model, and its parameters are continuously optimized iteratively to obtain a large language model for disaster scenarios.

[0015] In a second aspect, the present application further provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the disaster reporting method described in the embodiment of the first aspect is implemented.

[0016] In a third aspect, the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and wherein when the computer program is executed by a processor, the disaster reporting method described in the embodiment of the first aspect is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present application is further described below with reference to the accompanying drawings and embodiments, wherein: Figure 1 The main flow chart of the disaster reporting method provided in the embodiment of the present application; Figure 2 for Figure 1 Flow chart of step S100; Figure 3 for Figure 1 Flow chart of step S200; Figure 4 for Figure 1 Flow chart of step S300; Figure 5 This is a flow chart of S600 in the disaster reporting method of an embodiment of the present application; Figure 6 for Figure 2 Flow chart of step S120; Figure 7 The training flow chart of the large language model of the embodiment; Figure 8 Schematic diagram of the structure of an electronic device according to an embodiment.

[0018] Reference numerals: Processor 101 ; memory 102 ; input / output interface 103 ; communication interface 104 ; bus 105 . DETAILED DESCRIPTION

[0019] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application, and cannot be understood as limiting the present application.

[0020] In the description of the present application, it should be understood that descriptions involving orientation, such as up, down, front, back, left, right, etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0021] In the description of this application, "several" means more than one, "more" means more than two, "greater than", "less than", "exceed", etc. are understood to exclude the number itself, and "above", "below", "within", etc. are understood to include the number itself. If there is a description of "first" or "second", it is only used for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0022] In the description of this application, unless otherwise clearly defined, terms such as setting, installing, connecting, etc. should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in this application based on the specific content of the technical solution.

[0023] In the description of the present application, the description with reference to the terms "one embodiment", "some embodiments", "illustrative embodiments", "examples", "specific examples", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples in a suitable manner.

[0024] In the first aspect, the disaster reporting method of the embodiment is applied to a disaster reporting system, which includes a dialogue agent and an information management agent. The dialogue agent is responsible for real-time interaction with the user to obtain original dialogue information related to the disaster; the information management agent analyzes and manages the obtained target dialogue information and outputs a disaster report. The system can be applied to emergency command centers in areas prone to natural disasters, and can also be deployed in the cloud to achieve cross-regional collaboration. Figure 1 As shown, the disaster reporting method of the embodiment includes: S100, collecting original dialogue information through a dialogue agent.

[0025] In step S100, after a disaster occurs, users (such as disaster victims, rescue workers, or on-site commanders) conduct real-time dialogues with the dialogue agent through the terminal. The large language model built into the dialogue agent can recognize and receive inputs in various forms, including text, voice, pictures, videos, and location information including GPS coordinates. For character or voice input, the dialogue agent will transcribe it and perform preliminary semantic analysis; for image and video input, the image recognition or video target detection module is called to extract key features, thereby obtaining valuable original dialogue information.

[0026] S200, controlling the dialogue agent to pre-process the original dialogue information to obtain target dialogue information, and sending the target dialogue information to the information management agent in real time.

[0027] In step S200, after obtaining multiple rounds of original dialogue information, the dialogue agent will first use the large language model to remove duplication and calculate the relevance of each round of dialogue. Optionally, if the repetition of some dialogue content is higher than a preset threshold, it will be judged as "duplicate information", and the dialogue agent will merge the content to avoid redundant storage; if the relevance of some dialogue content to the disaster topic is lower than another threshold, it will be judged as "irrelevant information" and eliminated. The retained information forms relatively concise and targeted target dialogue information, which is sent to the information management agent in real time through the system communication module for subsequent extraction and comprehensive analysis of disaster information.

[0028] S300, controlling the information management agent to receive the target dialogue information in real time, and extracting features from the target dialogue information to obtain disaster information.

[0029] In step S300, after receiving the target dialogue information, the information management agent will first perform feature recognition and extraction on the received content based on the pre-configured disaster template. For example, the information management agent will focus on extracting key information elements such as disaster type, location, time, severity, and resource requirements, and combine them to generate disaster information data related to the disaster.

[0030] S400, storing the disaster information and all original conversation information in the first database and the second database respectively.

[0031] In step S400, the first database mainly stores the extracted disaster information (i.e., structured data) to facilitate subsequent comparison, retrieval or visualization with other disaster information; the second database retains all original conversation information related to the disaster (i.e., multimodal conversation content before reduction) to facilitate subsequent inspection and supplementary analysis when tracing is required.

[0032] S500, controlling the information management agent to generate a report according to the disaster information and sending the report.

[0033] In step S500, after the information management agent completes the feature extraction and data integration of the disaster information, it will automatically generate a disaster report, such as a summary of the current disaster situation, rescue resource requirements, affected areas, etc. Finally, the information management agent pushes the report to the disaster management department or rescue command platform, so that relevant personnel can make further deployments and decisions based on the report content.

[0034] In step S100 to step S500, the disaster reporting method of the embodiment collects multimodal data including various forms through multiple rounds of interaction between the dialogue agent and the user, and uses a large language model to determine the repetition of information and the relevance of the disaster theme, so as to realize the merging of repeated information and the elimination of irrelevant information, which can significantly reduce the interference of meaningless or low-value data, making the subsequent disaster information extraction process more focused and accurate. Secondly, after receiving the filtered target dialogue information, the information management agent extracts and comprehensively analyzes the key elements (such as event type, location, time, severity and resource requirements), and automatically generates a disaster report. The collaborative work of the dialogue agent and the information management agent effectively avoids the understanding bias and omissions that may occur in a single model when dealing with complex disasters, and improves the accuracy and completeness of the report. Finally, the structured data of the disaster information and the unstructured data of all the original dialogue information are stored in two databases respectively, which not only ensures the complete retention of the original interaction context, facilitates subsequent tracing and verification, but also provides technical guarantee for the system to quickly query and accurately analyze key disaster elements. In summary, this application utilizes a multi-agent collaborative framework driven by a large language model to achieve effective collection and real-time processing of user multimodal interaction data, combined with high-quality disaster information extraction and report generation mechanism, to provide efficient and reliable technical support for command decision-making in emergency scenarios, and has significant practical value and promotion significance.

[0035] Reference Figure 2 Understandably, in order to effectively collect the original dialogue information and conduct multiple rounds of iterative question and answer, the dialogue agent is pre-installed with a large language model that supports multiple input and output formats, which can automatically generate and continuously optimize the dialogue flow according to the user's needs and response content in different disaster scenarios. Figure 2 As shown, step S100 may include but is not limited to steps S110 to S130: S110, controlling the dialogue agent to generate question information using a preset large language model, and obtaining response information corresponding to the question information.

[0036] In step S110, when the dialogue agent starts a dialogue with the user, it automatically generates the first round of "question information" using a large language model in accordance with a pre-configured guidance strategy or based on known scenario information (such as the first round of information sent by the user, the types of disasters commonly encountered in the current season and region, etc.), such as "Please tell me the exact location and time of the disaster?", "Are there any injured people at the scene and how many?", etc. The question information can include text, voice, or visual prompts to adapt to multiple input and output channels of the user terminal.

[0037] S120, controlling the dialogue agent to generate new question information based on the previous response information using the large language model.

[0038] In step S120, after receiving the first round of questions from the dialogue agent, the user can give corresponding response information through text, voice, picture or location sharing. Therefore, in the process of generating new question information, the large language model needs to process multiple inputs such as text, voice, picture, video, location information, etc., perform semantic analysis and feature extraction on them, and generate new question information based on the analyzed content.

[0039] For example, Figure 6 As shown, step S120 may include but is not limited to steps S121 to S123: S121, control the dialogue agent to receive the previous round of response information, the response information includes one or more of text, voice, picture, video and location information.

[0040] S122, using a large language model to perform semantic analysis on text, perform voice transcription and semantic recognition on voice information, perform target recognition and feature extraction on image and video information, and perform geographic location analysis on location information, to obtain composite interactive data.

[0041] S123, controlling the large language model to generate new question information according to the composite interaction data.

[0042] In steps S121 to S123, the dialogue agent receives multimodal information from the user environment during each round of interaction, for example: Text information: Directly use natural language processing components to perform word segmentation, entity recognition, and dependency analysis.

[0043] Voice information: It is transcribed into text through the voice recognition module, and the acoustic emotional characteristics can be recorded to assist in judging the degree of urgency or user status.

[0044] Image information: Use target detection or image recognition algorithms to identify whether there are disaster markers (such as water surface height references, fire points, traffic jams, etc.).

[0045] Video information: Extract key frames from the video and use image recognition and scene detection technology to extract features related to the disaster, such as the spread of fire and the expansion of floods.

[0046] Location information: Based on the geographic location information, the administrative area, longitude and latitude, or surrounding special terrain and landforms are analyzed to assess the possible impact range of the disaster.

[0047] The results obtained from the above processing flow will be integrated into "composite interaction data", which not only contains the text description or feature vector of the aforementioned multimodal recognition, but also comes with metadata such as timestamps and conversation context identifiers. The composite interaction data is input into the large language model for further semantic fusion and element recognition. In this process, the large language model will combine text context and non-text elements (such as key objects in the image recognition results, action scene descriptions in the video, etc.) to infer whether the current user is in an emergency state, whether there is still key information missing, etc. For example: if the image recognition results show signs of landslides, the model determines that the location of the mountain should be further inquired, and the scale of the landslide may not be clear; if the location information shows that it is in a flood-prone area, the current status of the dam or water area can be further explored. Through this multimodal semantic fusion, the large language model can more accurately identify the missing points of disaster elements (such as location details, time nodes, resource requirements, etc.), and generate the required follow-up inquiries based on this. Subsequently, based on element recognition, the large language model will generate new question information instructions based on the context and the missing key information in the composite interaction data.

[0048] S130, controlling the dialogue agent to generate original dialogue information in real time according to the question information and the answer information. During multiple rounds of interaction between the user and the dialogue agent, all question information and corresponding answer information will be recorded and merged to form original dialogue information.

[0049] It can be seen that in step S110 to step S130, the disaster reporting method of the embodiment conducts multiple rounds of dialogues between the dialogue agent and the user, avoiding the problems of incomplete information and untimely updates that may be caused by a single round of questions and answers. The dialogue agent presets a large language model, and thus uses the large language model to adjust the questioning strategy and language style in real time according to the user's feedback, which improves the degree of automation of the processing, significantly reduces the need for manual intervention, and ensures the real-time, flexibility and adaptability of the information collection stage.

[0050] Understandably, during the interaction between the dialogue agent and the user, the user may send a large amount of detailed, off-topic or repeated content. In order to reduce the pressure of the subsequent information management agent, the dialogue agent will pre-process the original dialogue information in real time to remove irrelevant and repeated content. Figure 3 Step S200 may include but is not limited to steps S210 to S240: S210, control the dialogue agent to use the large language model to calculate the repetition of the original dialogue information and the relevance to the disaster report topic in real time. After the dialogue agent collects multiple rounds of original dialogue information, it first calls the built-in large language model to calculate the repetition between the information and the content currently stored in the cache; at the same time, the dialogue agent uses the large language model to calculate the relevance between the original dialogue information and the disaster report topic.

[0051] S220, original dialogue information with a repetition degree higher than a preset first threshold is determined as repeated information, and original dialogue information with a relevance degree lower than a preset second threshold is determined as irrelevant information. When the repetition degree is higher than a preset threshold, the dialogue agent can determine that the information is similar or identical to the content described in the previous round of dialogue information, and classify it as "repeated information". At this time, the dialogue agent will merge the content with higher consistency to avoid redundant storage or repeated analysis in the subsequent process; at the same time, if the relevance of the original dialogue information to the disaster topic is lower than another preset threshold, it will be determined as "irrelevant information" and removed, and no longer sent to the information management agent, so as to avoid adding noise or interference in the subsequent disaster extraction.

[0052] S230, merge the duplicate information in the original dialogue information, and remove the irrelevant information in the original dialogue information to obtain the target dialogue information. After the above preprocessing operation, the retained information forms the "target dialogue information". The target dialogue information usually contains the content with high relevance and low redundancy in multiple rounds of dialogue, and is organized in chronological order or thematic module.

[0053] S240, sending the target dialogue information to the information management agent in real time. To ensure that the disaster information can be obtained and analyzed in a timely manner, the dialogue agent will send the target dialogue information to the information management agent in real time (such as through a message queue or Socket communication). This real-time feature helps the information management agent to extract features and further process the disaster information in the shortest time, thereby improving the efficiency of emergency response.

[0054] It can be seen that the dialogue agent uses a large language model to make a comprehensive judgment on the repetition and relevance of the original dialogue information, which greatly reduces the noise and interference of useless or low-value information on the system, ensuring that the content received by the information management agent is more targeted and of higher quality. At the same time, the target dialogue information is sent in real time, realizing the dynamic update of disaster data. Once the user provides new or revised disaster details, the information management agent can be synchronized in a very short time and conduct further analysis or generate reports.

[0055] Understandably, after receiving the newly arrived target dialogue information, the information management agent will immediately enter the feature extraction process. Figure 4Step S300 may include but is not limited to steps S310 to S320: S310, extracting features of the target dialogue information based on the manually configured disaster template to obtain information on the event type, location, time, severity, and resource requirements; S320, disaster information is obtained based on the event type, location, time, severity and resource requirements.

[0056] Among them, common elements of several disaster templates (such as floods, earthquakes, typhoons, fires, etc.) are pre-configured within the information management agent. Each template includes common disaster information elements, such as: Time type: occurrence time, duration or end time, etc.; Location: province, city / county, street, GPS coordinates, etc.; Time: Specific date or time period. If there is a vague description such as "last night", the natural language processing algorithm is used for further analysis; Severity: number of casualties, scale of housing damage, degree of traffic disruption, etc.; Resource requirements: emergency supplies, medical resources, food supplies, etc.

[0057] After receiving the target conversation information, the information management agent will first make a preliminary judgment on which disaster template or templates need to be matched based on the "event type" pointed to or displayed in the conversation. After determining the disaster template, the information management agent will extract features of the target conversation information in real time. For example, the information management agent can call the pre-trained natural language processing algorithm and the large language model to implement the following feature extraction steps: Time type extraction: Identify the time when the disaster occurred in the conversation information (such as "20 o'clock yesterday", "last weekend") and the update time of subsequent new information; Location extraction: Obtain province, city, district / county, and more precise street or GPS coordinates through keyword matching or word segmentation recognition; Severity extraction: The disaster severity is graded based on the description in the conversation (e.g., "the river water has flooded the first floor of the residential building", "the road is completely paralyzed", etc.); Resource demand identification: Structure the demand content based on the description of material or rescue needs in the conversation (such as "urgent need for water pumps", "need for drone inspections", "need for food supplies"); Supplement of special elements: If it involves important information such as the number of injured people, demand for temporary resettlement sites, disaster scope forecast, etc., it will be extracted into the disaster information and given a separate field.

[0058] After completing the identification of characteristic elements, the information management agent integrates this information into structured disaster information, which can be directly stored in the first database (structured database).

[0059] It is understandable that after the information management agent completes the above-mentioned disaster information extraction each time, if the information management agent detects that the current disaster information is incomplete or there are potential doubts (critical disaster information is seriously missing, location information only describes the county but omits the precise street, there is a conflict between the previous and subsequent descriptions of the disaster-affected area, etc.), the information management agent can immediately generate a "feedback information", which contains information such as elements that need to be further verified or supplemented, and send it to the dialogue agent to request additional information, or supervise and correct the behavior of the dialogue agent, so as to improve the efficiency and accuracy of the system response. Therefore, in some embodiments, after step S300, it also includes: S600, the control information management agent sends feedback information to the dialogue agent in real time, and the dialogue agent adjusts the response strategy and question information in real time according to the feedback information; wherein the feedback information is obtained according to the disaster information.

[0060] In step S600, the information management agent provides feedback to the dialogue agent in real time to request additional information, or monitor and correct the behavior of the dialogue agent to improve the efficiency and accuracy of the system response.

[0061] More specifically, Figure 5 As shown, in some embodiments, step S600 may include but is not limited to steps S610 to S630: S610, controlling the reflection module of the information management agent to compare the disaster information with a preset feature information template, determining whether the disaster information has missing information, repeated information, and potential conflicts, and obtaining a comparison result; S620, controlling the reflection module of the information management agent to analyze the missing elements and elements to be clarified in the current target dialogue information according to the comparison result, and generate feedback information; S630, sending the feedback information to the dialogue agent in real time, and the dialogue agent adjusts the response strategy and the question information for the user according to the feedback information.

[0062] Among them, the information management agent has a built-in "reflection module". After completing the disaster information extraction, it will automatically compare the currently obtained disaster elements with the pre-configured feature information template. The template usually contains a list of elements and logical rules for common disaster types. Through comparison, the reflection module can quickly detect whether there is a lack of key information (such as not giving the accurate time, location or resource requirements), whether there is an unreasonable conflict (such as inconsistent location descriptions before and after the conversation, etc.), and whether the information is redundant (multiple repeated content). If the reflection module finds that the currently extracted disaster information is missing or contradictory, it will automatically generate a comparison result record and conduct in-depth analysis of each element if necessary. For example: when it is detected that the location information conflicts with each other in the previous round and the current round of data, the reflection module will mark it as "location conflict" and indicate the conflict location in the comparison result; when it is detected that the severity and resource requirements are contradictory, such as the previous description that the flood has submerged a large amount of farmland, but the later text mentions that there is basically no impact on the farmland, it will be marked as "disaster severity conflict". After the reflection module makes a preliminary assessment of the comparison results, it will generate an executable "feedback message" that clearly states: which information elements are missing or insufficient to support subsequent decisions; which information is inconsistent or potentially conflicting and needs further verification; whether special resources or additional descriptions are required (such as the unknown quantity of medical supplies), etc. This feedback information will be sent to the dialogue agent in real time through the communication module of the information management agent, so that the dialogue agent can quickly ask more targeted questions to the user in the next conversation.

[0063] After receiving the feedback information, the dialogue agent will use the large language model to parse the missing information or conflict points, and quickly form one or more new question information. For example, when a location conflict is identified, the system will prompt the dialogue agent to ask the user, "Please confirm the specific street name or GPS coordinates where the disaster occurred in order to verify the accuracy of the previous and subsequent records." When it is detected that additional resource information is needed, the dialogue agent will ask the user, "Do you still need more emergency supplies, such as water pumps or generators, in the current disaster?" The new information provided by the user after the response is again pre-processed by the dialogue agent, analyzed by the information management agent, and verified by the reflection module, so that the disaster information can be continuously iterated and optimized, and the timeliness and accuracy of the information can be maximized. It can be seen that through the comparison and feedback mechanism of the reflection module, the system can quickly discover inconsistencies or missing points in the disaster information and improve the completeness of the data; multiple rounds of interaction allow the dialogue agent and the user to maintain continuous information updates and corrections, significantly reducing misjudgments or omissions caused by insufficient initial information collection; the disaster report finally generated is more timely and credible, providing reliable data support for management personnel and rescue operations.

[0064] The large language model used in this application in the disaster reporting scenario is used as the core algorithm support. Therefore, the training method of the large language model must ensure the effective integration and analysis of multimodal data, and have a high level of understanding and reasoning capabilities in the disaster field. The training process of the large language model can be divided into key stages such as data acquisition, data cleaning and screening, data labeling and structured processing, data expansion, quality assessment, and training iteration optimization. For example, Figure 7 The training method of the large language model preset in the disaster reporting system of the embodiment of the present application includes but is not limited to steps S710 to S760: S710, obtaining an original dialogue dataset.

[0065] In step S170, the original conversation dataset can be obtained from various channels such as historical chat records of the disaster emergency management platform, social media related topic data, simulation exercises and simulation data, covering multimodal forms such as text, voice, pictures, videos and location information.

[0066] S720, denoising, deduplication and disaster scenario matching screening are performed on the original dialogue data set to obtain a first data set.

[0067] In step S720, the original dialogue data set is quality screened. Specifically, invalid characters and redundant samples are removed, and samples with high relevance to disaster reports are retained based on the disaster theme matching degree (such as keywords such as "fire" and "earthquake". Based on the above screened data, the first data set is formed, which not only retains the diversity of various disaster information as much as possible, but also ensures its close integration with the disaster scene.

[0068] S730, annotate the first data set to obtain annotation information, perform structured processing on the text, voice, picture, video and location information to obtain structured information, and merge the annotation information and the structured information into a second data set.

[0069] In step S730, first, the disaster type, location, time point, severity, resource requirements and other elements in the text data are manually or semi-automatically annotated to obtain annotation information; for non-text data such as voice, pictures, videos and location information, the key information is extracted and mapped to pre-defined fields as annotation information. Then, the text, voice, pictures or videos are uniformly formatted to obtain result information. Finally, the annotated information and structured information are merged into a second data set. Preferably, in the process of annotating the first data set, it can be divided into two types: extraction of key disaster elements and summary, so that the model can efficiently identify disaster elements and summarize necessary information in different task scenarios.

[0070] S740: Expand the second data set to obtain an augmented data set.

[0071] For example, in step S740, for text data, the text data can be expanded by inserting synonymous phrases, adding or correcting typos, adjusting word order, etc., and colloquial and incomplete sentences can be added as needed to simulate the real situation of accent differences and expression confusion that may occur when reporting disaster areas; for voice data, background noise can be injected into existing voice samples, the speaking speed and pitch can be changed, or different accents can be simulated by synthesis to improve the robustness of large language models in noisy and multi-accent environments; for video data, pictures or videos can be flipped, scaled, adjusted in illumination, and background synthesized to supplement disaster scenes in various visual environments, and corresponding feature replacement can be performed for different situations such as flooding, landslides, and fires. The data expanded by the above-mentioned multiple processing methods are merged into the second data set, further enriching the distribution of disaster-related samples and obtaining a more diverse augmented data set.

[0072] S750, using the quality scoring module of the large language model to evaluate the quality of the augmented data set, filter out low-quality data, and obtain a training data set.

[0073] For example, in S750, pre-trained metrics or sub-models can be used to evaluate each sample in the augmented data set in terms of "readability", "consistency", "disaster relevance", etc. Combined with multimodal consistency detection, such as determining whether the text description is consistent with the disaster elements in the picture / video, and whether the voice content is consistent with the transcribed text, avoid obvious conflicts or false associations in the training set.

[0074] S760, input the training data set into the large language model, continuously iterate and optimize its parameters, and obtain the large language model for the disaster scenario.

[0075] In step S760, in the model initialization stage, the general pre-trained weights of the large language model are first loaded to inherit its ability to understand general natural language. Subsequently, the training data set is input into the model in batches, and the model parameters are continuously updated through the adaptive optimization algorithm. At the same time, multi-task joint training can be combined with different goals (such as multimodal element extraction, question-answer pair generation, or information missing detection). During the training process, fine-tuning can be performed for the specific sub-functions required by each agent, such as: for user agents, highlighting the recognition accuracy of text and voice input and the ability to generate real-time questions, as well as strengthening the multimodal fusion and key element extraction of pictures, videos and location information; for information management agents, enhancing the ability to reflect and analyze disaster information, which is convenient for subsequent report compilation and feedback. Finally, for the verification and testing of the large language model, an independent verification set can be used to evaluate the model, such as the accuracy of key element extraction, the rationality of dialogue guidance, the ability to identify missing information, etc. If there is any deficiency, return to the augmented data or fine-tuning strategy for targeted optimization, and finally obtain a large language model that can efficiently understand disaster information.

[0076] In a second aspect, an embodiment of the present application further provides an electronic device, the electronic device comprising a memory and a processor, the memory storing a computer program, and the processor implementing the above disaster reporting method when executing the computer program. The electronic device may be any intelligent terminal including a tablet computer, a vehicle-mounted computer, etc.

[0077] See also Figure 8 , Figure 8 The hardware structure of an electronic device of another embodiment is illustrated, and the electronic device includes: The processor 101 may be implemented by a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit (Application Specific Integrated Circuit, ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present application; The memory 102 can be implemented in the form of a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 102 can store an operating system and other application programs. When the technical solution provided in the embodiment of this specification is implemented by software or firmware, the relevant program code is stored in the memory 102, and the processor 101 calls and executes the disaster reporting method of the embodiment of this application; Input / output interface 103, used to implement information input and output; The communication interface 104 is used to realize the communication interaction between the device and other devices. The communication can be realized through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.); A bus 105 , which transmits information between various components of the device (e.g., the processor 101 , the memory 102 , the input / output interface 103 , and the communication interface 104 ); The processor 101 , the memory 102 , the input / output interface 103 and the communication interface 104 are connected to each other in communication within the device via the bus 105 .

[0078] In a third aspect, an embodiment of the present application further provides a computer-readable storage medium, which stores a computer program, and the computer program implements the above-mentioned disaster reporting method when executed by a processor.

[0079] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory, and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0080] The embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the purpose of the present application. In addition, the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

Claims

1. A disaster reporting method, characterized in that: Applied to a disaster reporting system, the disaster reporting system includes a dialogue agent and an information management agent, and the disaster reporting method includes: Collecting original dialogue information through the dialogue agent; Controlling the dialogue agent to pre-process the original dialogue information to obtain target dialogue information, and sending the target dialogue information to the information management agent in real time; Controlling the information management agent to receive the target dialogue information in real time, and extracting features from the target dialogue information to obtain disaster information; storing the disaster information and all the original conversation information in a first database and a second database respectively; Control the information management agent to generate and send a report based on the disaster information.

2. The disaster reporting method according to claim 1, characterized in that: The collecting of original dialogue information by the dialogue agent includes: Controlling the dialog agent to generate question information using a preset large language model, and obtaining response information corresponding to the question information; Controlling the dialog agent to generate new question information using the large language model according to the previous response information; Control the dialogue agent to generate the original dialogue information in real time according to the question information and the response information.

3. The disaster reporting method according to claim 2, characterized in that: The controlling the information management agent to receive the target dialogue information in real time and extracting features from the target dialogue information to obtain disaster information, further comprising: The information management agent is controlled to send feedback information to the dialogue agent in real time, and the dialogue agent adjusts the response strategy and the question information in real time according to the feedback information; wherein the feedback information is obtained according to the disaster information.

4. The disaster reporting method according to claim 2, characterized in that: The controlling the dialogue agent to pre-process the original dialogue information to obtain target dialogue information, and sending the target dialogue information to the information management agent in real time includes: Controlling the dialog agent to calculate the repetition of the original dialog information and the relevance to the disaster report topic in real time using the large language model; Determine the original conversation information whose repetition degree is higher than a preset first threshold as repeated information, and determine the original conversation information whose relevance degree is lower than a preset second threshold as irrelevant information; Merging the repeated information in the original conversation information and removing the irrelevant information in the original conversation information to obtain target conversation information; The target dialogue information is sent to the information management agent in real time.

5. The disaster reporting method according to claim 1, characterized in that: The controlling the information management agent to receive the target dialogue information and extracting features from the target dialogue information to obtain disaster information includes: Feature extraction of target dialogue information based on manually configured disaster templates to obtain event type, location, time, severity, and resource demand information; Disaster information is obtained by combining the event type, location, time, severity and resource requirements.

6. The disaster reporting method according to claim 3, characterized in that: The controlling the information management agent to send feedback information to the dialogue agent in real time, and the dialogue agent adjusting the answering strategy and the question information in real time according to the feedback information, comprises: Controlling the reflection module of the information management agent to compare the disaster information with a preset feature information template, determining whether the disaster information has missing information, repeated information, and potential conflicts, and obtaining a comparison result; The reflection module controlling the information management agent analyzes the missing elements and elements to be clarified in the current target dialogue information according to the comparison result, and generates feedback information; The feedback information is sent to the dialogue agent in real time, and the dialogue agent adjusts the response strategy and the question information for the user according to the feedback information.

7. The disaster reporting method according to claim 2, characterized in that: The controlling the dialog agent to generate new question information using the large language model according to the previous response information includes: Controlling the dialog agent to receive the response information of the previous round, wherein the response information includes at least one of text, voice, picture, video and location information; Using the large language model, the text is semantically analyzed, the voice information is transcribed and semantically recognized, the image and video information is identified and features are extracted, and the location information is analyzed for geographic location, so as to obtain composite interaction data; The large language model is controlled to generate new question information according to the composite interaction data.

8. The disaster reporting method according to claim 2, characterized in that: The training method of the large language model includes: Get the original conversation dataset; De-noising, de-duplication and disaster scene matching screening are performed on the original dialogue data set to obtain a first data set; Annotating the first data set to obtain annotation information, and performing structured processing on the text, voice, picture, video and location information to obtain structured information, and merging the annotation information and the structured information into a second data set; Expanding the second data set to obtain an augmented data set; Using the quality scoring module of the large language model to perform quality assessment on the augmented data set, filter out low-quality data, and obtain a training data set; The training data set is input into a large language model, and its parameters are continuously optimized iteratively to obtain a large language model for disaster scenarios.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, the memory stores a computer program, and the processor implements the disaster reporting method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the disaster reporting method according to any one of claims 1 to 8 is implemented.

Images