Knowledge recommendation method, device and equipment based on aerospace remote sensing map and medium

Abstract

The invention provides a knowledge recommendation method, device and equipment based on an air-space remote sensing map and a medium, and relates to the technical field of artificial intelligence, the method comprises the following steps: rewriting and splitting a query question according to a question template, and extracting keywords from the rewritten and split query question; determining a first target node matched with the keyword in an air-space remote sensing map; the space and air remote sensing map is constructed according to a preset ontology framework based on entities, relationships and attributes extracted from theses related to remote sensing data and in combination with supplementary attributes of regional entities extracted from multi-source geographic data; obtaining a query template matched with the keyword, and filling the first target node into the query template to generate a query statement; and acquiring context data related to the query statement from the air-space remote sensing map, and inputting the context data and the query question into the generative large model as prompt information to obtain a knowledge answer. According to the method, accurate answers conforming to knowledge characteristics in the remote sensing field can be output, and efficient recommendation of remote sensing knowledge is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a knowledge recommendation method and device based on space remote sensing graph, equipment and medium. BACKGROUND

[0002] After years of development, the remote sensing field has accumulated a large amount of data and rich knowledge. Among them, the data includes original remote sensing image data and various products made based on these data; the knowledge covers the basic theory of remote sensing field, and the innovative achievements formed by the application of machine learning, deep learning and other artificial intelligence technologies in the remote sensing field. These remote sensing technologies not only continue to deepen in the field, but also are widely used in many other fields, and have produced a large amount of knowledge achievements in the form of papers. These achievements are reflected in the research results of the papers on the one hand, which can be directly used as a supplement to the knowledge of a specific research area; on the other hand, the research methods used in the papers have strong transferability and can be used for reference and application in other areas, thereby promoting the wide application and transformation of space information knowledge.

[0003] Although large language models have the ability to understand and process unstructured text, when using remote sensing related papers as a knowledge base, they face more complex situations than processing general documents. Specifically, such papers are usually long, diverse in structure, varied in language, and highly knowledge-intensive. In addition, a large number of papers have cross, scattered and redundant problems in research topics, data usage and method selection. These characteristics not only increase the difficulty of text processing, but also reduce the accuracy of space application knowledge recommendation. SUMMARY

[0004] The present application provides a knowledge recommendation method and device based on space remote sensing graph, equipment and medium, to solve the problem that when using remote sensing related papers as a knowledge base, due to the characteristics of long length, diverse structure, varied language and high knowledge intensity of remote sensing related papers, the accuracy of space application knowledge recommendation is low, and to realize efficient recommendation of remote sensing knowledge.

[0005] The present application provides a knowledge recommendation method based on space remote sensing graph, comprising the following steps: In response to a query question input by a user, rewriting and splitting the query question according to a question template, and extracting at least one keyword from the rewritten and split query question; determining a first target node matching the at least one keyword in the space remote sensing graph; the space remote sensing graph is a knowledge graph constructed based on entities, relationships and attributes extracted from remote sensing data related papers, and supplemented attributes of regional entities extracted from multi-source geographic data according to a preset ontology framework; obtaining a query template matched with the at least one keyword, filling the first target node into the query template to generate a query statement; obtaining context data related to the query statement from the space remote sensing atlas, and inputting the context data and the query question as prompt information into a generative large model to obtain a knowledge answer output by the generative large model.

[0006] According to the knowledge recommendation method based on the space remote sensing atlas provided by the application, the first target node matched with the at least one keyword is determined in the space remote sensing atlas, which comprises: For each keyword, the first target node matched with the keyword is determined based on the embedding similarity between the first embedding vector of the keyword and the second embedding vector of each first node in the space remote sensing atlas; the first node is a node of the same entity type as the keyword; In the case that the entity type corresponding to the keyword is a task, the first target node matched with the keyword is determined based on the embedding similarity between the first embedding vector of the keyword and the second embedding vector of each second node in the space remote sensing atlas; the second node includes a task node and a process node.

[0007] According to the knowledge recommendation method based on the space remote sensing atlas provided by the application, the query template matched with the at least one keyword is obtained, which comprises: In the case that the at least one keyword is a single keyword and there is no upper-level keyword for the at least one keyword, a first query template matched with the at least one keyword is obtained; In the case that the at least one keyword is a single keyword and there is an upper-level keyword for the at least one keyword, a second query template matched with the at least one keyword and a third query template matched with the upper-level keyword of the at least one keyword are obtained.

[0008] According to the knowledge recommendation method based on the space remote sensing atlas provided by the application, the query template matched with the at least one keyword is obtained, which comprises: In the case that the at least one keyword is a first keyword pair, a fourth query template, a fifth query template and a sixth query template matched with the at least one keyword pair are obtained; the first keyword pair includes a target area and a task, the fourth query template is used to query a paper matched with the target area and the task, the fifth query template is used to query a scheme in the paper, and the sixth query template is used to find a scheme in a paper of a similar area to the target area in the case that there is no paper belonging to the target area; In a case where the at least one keyword is a second keyword pair, a seventh query template, an eighth query template and a ninth query template matched with the at least one keyword pair are acquired; the second keyword pair includes a target region and data, the seventh query template is used to query numerical data in the target region, the eighth query template is used to query whether a spatial range corresponding to spatial data contains the target region, and the ninth query template is used to query whether there is an inclusion relationship between regions.

[0009] According to the method, the space remote sensing graph is constructed by the following method: A plurality of papers related to remote sensing data are acquired, and the text format of the papers is unified; Each paper with a unified text format and first prompt information are input into a large language model to obtain a paper summary corresponding to each paper output by the large language model; the first prompt information includes the preset ontology framework and a remote sensing application methodology logical chain; The paper summary corresponding to each paper and second prompt information are input into the large language model to obtain entities, relationships and attributes in each paper output by the large language model; the second prompt information includes the preset ontology framework and structured knowledge extraction prompt information; For a region entity in the entities, supplementary attributes of the region entity are extracted in combination with multi-source geographic data; Based on the entities, relationships, attributes in each paper and the supplementary attributes of the region entity, a space remote sensing graph is constructed.

[0010] According to the method, the space remote sensing graph is constructed by the following method: For each entity in the paper, a label and a name related to the entity are created; Based on the label and the name related to the entity, it is determined whether a second target node identical to the entity exists in a current graph database; In a case where the second target node does not exist in the current graph database, a new node corresponding to the entity is created; In a case where the second target node exists in the current graph database, it is determined whether an attribute corresponding to the entity exists in the second target node; In a case where the attribute corresponding to the entity does not exist in the second target node, the attribute corresponding to the entity is assigned to the second target node; In a case where the attribute corresponding to the entity exists in the second target node, attribute values are merged; For each relationship in the paper, the head node and the tail node corresponding to the relationship are determined, an edge between the head node and the tail node is established, and the attribute corresponding to the relationship is assigned to the edge between the head node and the tail node.

[0011] According to the knowledge recommendation method based on the space remote sensing graph provided by the application, the preset ontology framework is obtained in the following manner: The paper is determined as a first core entity type, and the paper associated author and the paper associated region are created. The task is determined as a second core entity type, the paper associated task is created, and the task associated process, the process associated data processing tool and the data processing tool associated parameter are created.

[0012] The application further provides a knowledge recommendation device based on the space remote sensing graph, comprising: The first knowledge recommendation module is used for rewriting and splitting the query question input by the user according to the question template, and extracting at least one keyword from the rewritten and split query question. The second knowledge recommendation module is used for determining a first target node matched with the at least one keyword in the space remote sensing graph. The third knowledge recommendation module is used for obtaining a query template matched with the at least one keyword, filling the first target node into the query template, and generating a query statement. The fourth knowledge recommendation module is used for obtaining context data related to the query statement from the space remote sensing graph, and inputting the context data and the query question as prompt information into the generative large model to obtain a knowledge answer output by the generative large model.

[0013] The application further provides an electronic device comprising a memory, a processor and a computer program stored on the memory and running on the processor.

[0014] The application further provides a non-transitory computer readable storage medium having a computer program stored thereon, which is executed by a processor to implement the knowledge recommendation method based on the space remote sensing graph.

[0015] The application further provides a computer program product comprising a computer program which, when executed by a processor, implements the knowledge recommendation method based on the space remote sensing graph as described above.

[0016] The knowledge recommendation method, device, equipment and medium based on the space remote sensing graph provided by the application accurately locate the core of the user's question by rewriting and splitting the query question through the question template and extracting keywords. Secondly, the construction of the space remote sensing graph integrates entities, relationships, attributes and multi-source geographic data in remote sensing papers, so that the originally cross, scattered and redundant knowledge can be structured and presented, facilitating the rapid and accurate determination of the first target node matched with the keywords. Finally, the first target node is filled into the query template to generate a query statement, further clarifying the query direction, while the context data related to the query statement obtained from the space remote sensing graph provides rich background knowledge for the generative large model, enabling it to output accurate answers that meet the characteristics of the remote sensing field knowledge, and realizing efficient recommendation of remote sensing knowledge. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0018] Figure 1 is a flowchart of the knowledge recommendation method based on the space remote sensing graph provided by the application.

[0019] Figure 2 is a relationship diagram of the ontology framework provided by the application.

[0020] Figure 3 is a structural diagram of the knowledge recommendation device based on the space remote sensing graph provided by the application.

[0021] Figure 4 is a structural diagram of the electronic equipment provided by the application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the application clearer, the technical solutions in the application will be described clearly and completely in combination with the drawings in the application. Obviously, the described embodiments are some embodiments of the application, not all embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the application.

[0023] The knowledge recommendation method based on space remote sensing atlas of the embodiment of the application comprises steps 110, 120, 130 and 140. Figure 1

[0024] In step 110, in response to a query question input by a user, the query question is rewritten and split according to a question template, and at least one keyword is extracted from the rewritten and split query question.

[0025] It should be understood that the query question input by the user often has variability and ambiguity, and directly processing these query questions may lead to an inability to accurately match the nodes in the knowledge graph. Therefore, the query question input by the user needs to be rewritten and split first, so as to correspond to one or more predefined and more structured question templates.

[0026] In this step, the query question input by the user is converted into a more standardized expression, and then the rewritten query question is further split into more fine-grained questions to conform to the predefined question template, so as to be processed respectively, and finally the keywords are identified from the rewritten and split questions.

[0027] In this embodiment, five types of keywords are defined for extracting keywords from the query question: Task: the main goal or problem of the research; Algorithm model or software tool: a specific method or tool for solving the problem; Data: data resources used in the research.

[0028] Region + task: research task in a specific region; Region + data: data resources in a specific region.

[0029] Here, the predefined question template is to standardize the query question of the user, so that it can be more easily mapped to the nodes and edges in the knowledge graph.

[0030] The problems to be dealt with in this embodiment are around the keywords of task, data, algorithm model, software tool and region. Therefore, two types of question templates are defined: The first type is a question template without region limitation, such as an implementation scheme of a certain task, a use and input data of a certain algorithm model or software tool, a source and use of a certain data; The second type is an implementation scheme of a certain task in a certain region, a source and use of a certain data.

[0031] ​Step 120, determining a first target node matching the at least one keyword in the aerospace remote sensing graph; the aerospace remote sensing graph is a knowledge graph constructed based on entities, relationships and attributes extracted from papers related to remote sensing data, combined with supplementary attributes of regional entities extracted from multi-source geographic data, and according to a preset ontology framework.

[0032] After extracting the keywords, the same or similar target nodes are found in the aerospace remote sensing graph.

[0033] Here, the preset ontology framework refers to a formal framework used to define and standardize concepts, entities, relationships and attributes in a field when constructing a knowledge graph. It provides a standardized template for the construction of a knowledge graph, ensuring consistency and logicality in the organization and storage of knowledge.

[0034] In one example, the preset ontology framework is designed as follows: Determine the paper as the first core entity type, create the paper associated author and the paper associated region; Determine the task as the second core entity type, create the paper associated task, and create the task associated process, the process associated data processing tool, and the data processing tool associated parameter; wherein the data processing tool includes at least one of an algorithm model or a software tool.

[0035] Firstly, the paper is the first core entity type; The paper is associated with the author, and the works of the same author are usually systematic and coherent; The paper is associated with the region, indicating the research scope of the paper. The description of the research scope is various, from global to any local region. The "region" and the "region" are connected through "belongs to" or "contains". A paper may use one or more administrative regions as an experimental area. The experimental area may not be a standard administrative region (for example, "five central regions of city a"), but it has a "contains" relationship with a specific administrative region. For regions outside country A, the value of the entity name attribute remains in the original language. In addition, the experimental area introduction and other parts of the paper will usually describe the area, geographical location, natural conditions (climate, soil, etc.), social conditions (economy, population, etc.) and other characteristics, which are very important for the transfer of the scheme. In order to avoid errors, it is also necessary to distinguish whether such "characteristics" are for the entire experimental area, for a part of the experimental area, or for the administrative region to which the experimental area belongs. If the scope of such characteristics cannot be accurately identified, errors may occur; Secondly, the task is the second core entity type. The task is the research topic of the paper, and usually multiple papers are directed to the same task. A "task" is associated with a "process", which refers to a series of steps to achieve the "task". A "process" can be further divided, and the closer to the bottom layer, the more likely to be shared by multiple "tasks"; A "process" is associated with an "algorithm model" or a "software tool". An "algorithm model" includes exponential method, machine learning model, deep learning model, forest fire spread model, power generation calculation formula, etc. A "software tool" refers to geographic information system software or software for specific purposes; An "algorithm model" and a "software tool" are associated with "parameters", "remote sensing images", "bands", "general data", etc. "Parameters" are input parameters of the algorithm model. "Remote sensing images" refer to Google images, high-resolution series images, Sentinel images, unmanned aerial vehicle images, LiDAR maps, etc. "Bands" refer to red bands, blue bands, near-infrared bands, etc. "General data" refers to population data, building vector data, road vector data, solar radiation data, GDP data, DEM data, DSM data, etc. These data have spatial resolution, coverage, etc.

[0036] For ease of understanding, refer to Figure 2 , as shown in Figure 2 A relationship diagram of an ontology framework provided by the present embodiment is shown, which shows the main entity types in the ontology framework and their relationships.

[0037] It should be understood that an entity type is a basic unit in a knowledge graph, used to represent different concepts or objects. For the storage of entity types, each row is a list containing two elements, the first element represents the specified entity type, and the second element is a sub-list containing some examples of the entity type. A relationship represents the connection between entities. For the storage of relationships, each row is a list containing at least three elements, the first and third elements are the specified entity types, and the second element is the relationship that may exist between the two types of entities, and the rest are the attributes of the relationship. An attribute is a specific feature of an entity. For the storage of attributes, each row is a list containing two elements, the first element is an attribute, and the second element is an entity type that may have the attribute.

[0038] In this step, we first collect papers related to remote sensing data. These papers are the main data source for building the knowledge graph. Then, we extract key entities, relationships between entities, and attributes of entities or relationships from the papers. In addition, for "region" type entities, in addition to the regional information in the papers, we can also supplement and improve it from multi-source geographic data. In this way, we combine this multi-source geographic data with the regional information extracted from the papers to build a more comprehensive attribute system for "region" type entities. Finally, the extracted entities, relationships, and attributes are organized according to the preset ontology framework to form a structured aerospace remote sensing atlas. This aerospace remote sensing atlas can be stored in a graph database to facilitate subsequent query and analysis.

[0039] Step 130: Obtain a query template matching the at least one keyword, fill the first target node into the query template, and generate a query statement. In this embodiment, different query templates are designed for different keyword types to achieve accurate retrieval of the graph database.

[0040] In this step, select the corresponding query template based on the identified keyword type. Fill the query template with the identified keyword value (such as task name, region name, etc.) as a parameter to generate a query statement.

[0041] Step 140: Obtain context data related to the query statement from the aerospace remote sensing atlas, and input the context data and the query question into the generative model as prompt information to obtain the knowledge answer output by the generative model.

[0042] It should be understood that the main function of the generative large model is to generate new and coherent text content based on the input prompt. The generative large model used in this embodiment is consistent with the existing technology and is not limited to this.

[0043] In this step, the query statement is first executed, and relevant data retrieved from the graph database's aerospace remote sensing atlas serves as context data. This data is retrieved from the graph database's aerospace remote sensing atlas based on the keywords and query template in the user's query. The retrieved context data is then merged with the user's original query to form a complete prompt. Finally, this merged prompt is input into the generative macro model, which leverages its language generation capabilities to generate a reasonable knowledge-based answer based on the prompt.

[0044] The knowledge recommendation method based on the space remote sensing graph provided by the embodiment of the application rewrites and splits the query question through a question template and extracts keywords to accurately locate the core of the user question. Secondly, the construction of the space remote sensing graph integrates entities, relationships, attributes and multi-source geographic data in remote sensing papers, so that originally cross, scattered and redundant knowledge can be structured and presented, facilitating quick and accurate determination of the first target node matched with the keywords. Finally, the first target node is filled into the query template to generate a query statement, further clarifying the query direction, and the context data related to the query statement obtained from the space remote sensing graph provides rich background knowledge for the generative large model, enabling it to output accurate answers that meet the characteristics of remote sensing field knowledge, and achieving efficient recommendation of remote sensing knowledge.

[0045] It should be noted that each embodiment of the present application can be freely combined, the order can be changed or executed alone, and does not need to rely on or depend on a fixed execution order.

[0046] In some embodiments, the space remote sensing graph is constructed by the following method: Obtain multiple papers related to remote sensing data, and unify the text format of the papers; input each paper with a unified text format and first prompt information into a large language model to obtain a paper summary corresponding to each paper output by the large language model; the first prompt information includes the preset ontology framework and remote sensing application methodology logical chain; input the paper summary corresponding to each paper and second prompt information into the large language model to obtain entities, relationships and attributes in each paper output by the large language model; the second prompt information includes the preset ontology framework and structured knowledge extraction prompt information; For regional entities in the entities, supplementary attributes of the regional entities are extracted in combination with multi-source geographic data; Based on the entities, relationships, attributes and supplementary attributes of the regional entities in each paper, a space remote sensing graph is constructed.

[0047] It should be understood that the large language model is a model constructed based on deep learning technology, mainly used for understanding and generating natural language text. The large language model used in this embodiment is consistent with the prior art, and is not limited.

[0048] In this embodiment, papers related to remote sensing data are collected as an external knowledge base. Then the papers are converted into a simple and lightweight unified text format to eliminate format differences between multiple texts. In addition, to improve efficiency, the literature review, thanks, references and other useless parts can also be removed.

[0049] As the length of the paper is usually longer and the knowledge is more intensive, in this embodiment, the prompt words are designed according to the remote sensing application methodology logical chain. In addition, in order to preserve as many details as possible, the preset ontology framework is also input as a prompt to the large language model. The large language model generates a paper summary for each paper according to the remote sensing application methodology logical chain and the ontology framework, including the title, author, research background, research topic, research scope and characteristics, specific process and algorithm model, software tool and input data, and research conclusion.

[0050] Here, the remote sensing application methodology logical chain refers to the logical steps from the definition of the research task to the specific implementation in remote sensing application research. This logical chain describes how to start from a research goal, gradually decompose into specific implementation steps, select appropriate methods and tools, and the required data resources, and finally complete the research task. In an example, the remote sensing application methodology logical chain is "task-process-algorithm model / software tool-data".

[0051] After extracting the paper summary of each paper based on the first prompt information, the entity, relationship and attribute in each paper are extracted from the paper summary of each paper based on the designed structured knowledge extraction prompt information and the preset ontology framework.

[0052] Here, the structured knowledge extraction prompt information refers to a series of guiding information designed to guide the model to generate output that meets specific format and requirements when using a large language model for text processing. These information help the model understand the specific requirements of the task and generate structured output. In an example, the structured knowledge extraction prompt information includes but is not limited to: Overall task: Clearly define the main task that the model needs to complete; Specific requirements: Detailed description of the specific requirements of the task, including the types of entities, relationship types and attributes that need to be extracted; Output format: Specify the output format of the model to ensure the structured and consistent output information; Examples: Provide specific examples to help the model understand the specific requirements and output format of the task.

[0053] Further, as the basic geographic entity, the attribute information of the region is not limited to the introduction of the region in the academic literature of the paper. Another part of the attributes of the region, such as the calculation results of some indicators or the findings or suggestions in some aspects, will also be included in the conclusion part of the paper. These are the most direct knowledge generated by the paper and should be mined as the attributes of the region. Similarly, it should also be distinguished whether these attributes are for the entire experimental area or for the local area. In addition, it can also be supplemented and improved by various external auxiliary geographic data. The administrative division data can provide the basic geographic information such as the accurate spatial range, area data and inclusion relationship between administrative divisions of each level region. In addition, the natural environment elements and social economic indicators based on the grid data can be associated to the corresponding regional unit through spatial aggregation method, so as to construct a comprehensive and accurate multi-dimensional attribute feature system for the region entity. This multi-source geographic data fusion method can improve the completeness and accuracy of the description of the region entity.

[0054] Therefore, in the embodiment, after the entities, relationships and attributes in each paper are extracted from the paper summary based on the second prompt information, the supplementary attributes of the region entity are extracted from the multi-source geographic data and supplemented to the attributes of the region entity.

[0055] Specifically, all the extracted entities, relationships, attributes and supplementary attributes of the region entity are constructed into the space remote sensing atlas and stored in the graph database.

[0056] The knowledge recommendation method based on the space remote sensing atlas provided by the embodiment of the application collects papers related to remote sensing data as an external knowledge base. Then, a large language model is called, prompt word technology is adopted, and a paper summary is generated for each paper according to the specified ontology framework. The entities, relationships and attributes are extracted from the paper summary, and the attributes of the region are supplemented from the multi-source geographic data. Therefore, the originally crossed, scattered and redundant knowledge can be structured and presented, and the space remote sensing atlas is constructed.

[0057] In some embodiments, the space remote sensing atlas is constructed based on the entities, relationships and attributes in each paper, including: For each entity in the paper, a label and a name related to the entity are created; Based on the label and the name related to the entity, it is determined whether the same second target node exists in the current graph database; In the case that the second target node does not exist in the current graph database, a new node corresponding to the entity is created; In the case that the second target node exists in the current graph database, it is determined whether the second target node has an attribute corresponding to the entity; In the case that the second target node does not have the attribute corresponding to the entity, the attribute corresponding to the entity is assigned to the second target node; In the case that the second target node exists, the attribute value merging is performed. For each relationship in the paper, the head node and the tail node corresponding to the relationship are determined, an edge between the head node and the tail node is established, and the attribute corresponding to the relationship is assigned to the edge between the head node and the tail node.

[0058] In the space remote sensing graph, a node is a basic unit of a graph and is used to represent an entity. Each node usually has the following two key attributes: a label and a name. The label represents the type of entity to which the node belongs, and the name represents the specific content or name of the node. Under the same label, the name is used to distinguish different nodes. For example, the titles of different papers and the names of different authors.

[0059] When creating a node, it is necessary to determine whether a node with the same label and name already exists in the graph database. Specifically, for each entity, first, the label and name related to the entity are created. Then it is determined whether a second target node with the same label and name already exists in the graph database. If not, a new node is directly created. If a second target node with the same label and name already exists, it is determined whether the attribute corresponding to the current entity exists in the second target node. If some attributes do not exist, the non-existing attributes are assigned to the second target node. If the attributes already exist, the new attribute value is merged with the old attribute value, and the merged value is stored as a list.

[0060] When creating an edge, the head node and the tail node are first read, and it is determined whether the head node and the tail node corresponding to the current relationship both already exist in the graph database. If the head node and the tail node corresponding to the current relationship both already exist in the graph database, an edge between them is established, and the attributes of the edge are stored in the database.

[0061] Further, each "paper" is a core node that contains specific research content, methods, and results. A "task" is the research topic of a "paper" and is also a core node, representing a specific problem to be solved or a research direction. Taking "papers" and "tasks" as core nodes, each "paper" generates a subgraph. This subgraph contains all information related to the "paper", such as authors, regions, methods, data, etc. A "task" node can be connected to multiple "paper" nodes, i.e., multiple "papers" can research the same "task", forming a multi-angle exploration of the "task". Each "paper" can contain multiple "process" nodes, representing the research methods and steps described in the "paper". When multiple "papers" share the same "task" node, the "task" node will connect "process" nodes from multiple "papers". This can cause confusion among "process" nodes, making it difficult to distinguish which "process" belongs to which "paper".

[0062] Based on this, two key attributes are also designed in this embodiment: source and sequence number. Among them, the source is used to identify which paper the node belongs to. Through the source attribute, the ownership of each node can be clearly identified to avoid confusion. The sequence number is used to identify the order of the node in the paper. For example, for the "process" node, through the source attribute, it can be clear that the "process" node comes from which paper, and through the sequence number attribute, it can be clear that the "process" node is in the order in the paper, which is convenient for understanding the research method and step of the paper.

[0063] In some embodiments, the determining, in the aerospace remote sensing graph, a first target node matching the at least one keyword comprises: For each keyword, based on the embedding similarity between the first embedding vector of the keyword and the second embedding vector of each first node in the aerospace remote sensing graph, a first target node matching the keyword is determined; the first node is a node of the same entity type as the keyword. Wherein, in the case that the entity type corresponding to the keyword is a task, based on the embedding similarity between the first embedding vector of the keyword and the second embedding vector of each second node in the aerospace remote sensing graph, a first target node matching the keyword is determined; the second node includes a task node and a process node.

[0064] It should be understood that five types of keywords are defined in this embodiment for extracting keywords from query questions: Task: the main goal or problem of the research; Algorithm model or software tool: specific method or tool for solving the problem; Data: data resources used in the research.

[0065] Region + task: research task in a specific region; Region + data: data resources in a specific region.

[0066] Due to the diversity of language in the paper, phrases with the same meaning often have multiple expressions, which may not completely match the nodes in the graph database. Therefore, this embodiment adopts embedding similarity-based keyword matching.

[0067] Specifically, the first embedding vector of the identified keyword is calculated, and the second embedding vector of each node in the graph database is generated in advance. Then, by calculating the embedding similarity between the embedding vectors, the first target node closest in semantics to the identified keyword is found.

[0068] Further, when the entity type corresponding to the keyword is a task, since the "process" node can be regarded as a low-level decomposition of the "task" node, they are specific steps to implement the "task". Therefore, not only the "task" node directly matching the "task" is searched, but also the "process" node related to the "task" is considered.

[0069] The knowledge recommendation method based on space remote sensing atlas provided by the embodiment of the application can accurately find the first target node matching the keyword by embedding similarity. In addition, when the entity type corresponding to the keyword is a task, not only the "task" node directly matching the "task" is searched, but also the "process" node related to the "task" is considered, so as to improve the comprehensiveness of the search result.

[0070] In some embodiments, the query template matching the at least one keyword is obtained, including: In the case that the at least one keyword is a single keyword and there is no upper-level keyword of the at least one keyword, a first query template matching the at least one keyword is obtained; In the case that the at least one keyword is a single keyword and there is an upper-level keyword of the at least one keyword, a second query template matching the at least one keyword and a third query template matching the upper-level keyword of the at least one keyword are obtained.

[0071] Here, the single keyword refers to a single independent keyword in a query task, which is not used in combination with other keywords and does not depend on other keywords to determine its specific meaning and query range. It is an independent and explicit query object used to directly locate the content related to it in the database. For example: task, algorithm model, software tool, data.

[0072] The upper-level keyword of the keyword refers to a more extensive category to which the current keyword belongs in the knowledge graph. It is usually used to describe the attribution relationship of the current keyword and help more accurately locate and understand the context of the current keyword. For example: for the keywords of algorithm model and software tool, they need to be summarized from the upper-level process and task, so the upper-level keywords of such keywords are process and task. For data, it needs to be summarized from the upper-level algorithm model and software tool, so the upper-level keywords of such keywords are algorithm model and software tool.

[0073] In an example, for the task keyword, a first query template can be constructed to query the process, data, algorithm model and software tool implementing the task. For example, the query steps designed in the first query template are as follows: search for a task node with the name "M"; find all process nodes directly contained by the task; Recursive find all sub-processes of these processes; Merge all processes and sub-processes, and expand after deduplication; For each process, find the tools (algorithm models or software tools) it needs to use; For each tool, find the data it needs to input; The return result includes: task name, process name, tool name, tool type (algorithm model or software tool), and data used by the tool.

[0074] For algorithm model and software tool keywords, a second query template can be constructed to query the use introduction of the algorithm model and software tool. For example, the query steps designed in the second query template are as follows: Find the tool node (type: algorithm model or software tool) with the name "N algorithm"; Directly return the "use introduction" attribute of the tool node.

[0075] For algorithm model and software tool keywords, a third query template can be constructed to query the process and task to which the algorithm model and software tool belong. For example, the query steps designed in the third query template are as follows: Find the tool node with the name "X"; Find all process nodes that use the tool; Find the task nodes to which these processes belong; The return result includes: process name and task name.

[0076] For data keywords, a second query template can be constructed to query the source or coverage range of the data. For example, the query steps designed in the second query template are as follows: Find the data node with the name "S"; Directly return the "data source" attribute of the data node.

[0077] For data keywords, a third query template can be constructed to query which tools will generate or use the data. For example, the query steps designed in the third query template are as follows: Find the data node with the name "S"; Find all tool nodes (algorithm models or software tools) that can output the data; The return result includes: tool name and tool type.

[0078] The knowledge recommendation method based on space remote sensing atlas provided in the embodiment of the application can, for a single keyword, determine whether the single keyword has an upper-level keyword, flexibly obtain a query template corresponding to the keyword, and thus improve the accuracy and comprehensiveness of the query result.

[0079] In some embodiments, the query template matched by the at least one keyword is obtained, comprising: In the case that the at least one keyword is a first keyword pair, a fourth query template, a fifth query template and a sixth query template matched by the at least one keyword pair are obtained; the first keyword pair comprises a target area and a task, the fourth query template is used to query a paper matching the target area and the task, the fifth query template is used to query a scheme in the paper, and the sixth query template is used to find a scheme in a paper of a similar area to the target area in the case that there is no paper of the target area. In the case that the at least one keyword is a second keyword pair, a seventh query template, an eighth query template and a ninth query template matched by the at least one keyword pair are obtained; the second keyword pair comprises a target area and data, the seventh query template is used to query numerical data in the target area, the eighth query template is used to query whether a spatial range corresponding to spatial data contains the target area, and the ninth query template is used to query whether there is an inclusion relationship between areas.

[0080] Here, the keyword pair refers to a query unit combined by two associated keywords in a query task, which is used to more accurately locate and retrieve information. In this embodiment, the keyword pair is divided into two types: a keyword pair of a target area and a task, and a keyword pair of a target area and data.

[0081] For the keyword pair of the target area and the task, it indicates that a scheme for carrying out a certain task in a specified target area is given. Therefore, the fourth query template, the fifth query template and the sixth query template are designed in this embodiment.

[0082] The fourth query template is used to query an existing research paper of a certain task in the target area. For example, the query steps designed in the fourth query template are as follows: Find a paper node with a research range of “T area” and a research theme of “P”.

[0083] The return result contains: paper title, research area, research theme.

[0084] The fifth query template is used to query a scheme in the paper, i.e. to query a process, data, algorithm model and software tool for implementing the task in the paper. Here, the fifth query template is the same as the first query template in the above, which will not be described in detail here.

[0085] The sixth query template is used for similar area substitution, i.e. when there is no paper directly researching the task in the target area, a scheme for researching the task in a paper of a similar area to the target area is recommended. For example, the query steps designed in the fourth query template are as follows: Find the task node with name "M"; Find all the paper nodes researching this task; Find the research area nodes of these papers; Return the result containing: research topic, paper title, research area.

[0086] Based on this, the research area of the target area can be recommended based on the research area queried by the fourth query template.

[0087] Here, the area similarity includes but is not limited to: similar area, same or similar administrative level, similar geographical location, similar environment.

[0088] For the keyword pair of the target area and the data, it indicates that the data in the specified target area is required. It should be understood that the data is divided into numerical data and spatial data. Numerical data refers to the value of the building roof area or photovoltaic power generation potential of the region, and spatial data refers to remote sensing images or solar radiation data with spatial coverage.

[0089] Based on this, different query templates are designed for different types of data in this embodiment. Specifically, for numerical data, the seventh query template is used to query the numerical data in the target area. For example, the query steps designed in the seventh query template are as follows: Find the area node with name "T area"; Return the name of the area, whether there is a "G" attribute, and the value of the attribute.

[0090] For spatial data, the eighth query template is used to query whether the spatial range corresponding to the spatial data contains the target area. For example, the query steps designed in the seventh query template are as follows: Find the data node with name "S".

[0091] Return the name and coverage attribute of the data.

[0092] Further, the eighth query template is also designed in this embodiment to query whether there is a containing relationship between regions. For example, the query steps designed in the eighth query template are as follows: Find whether there is a containing relationship path from "T area" to "Y area"; Return the result containing: whether there is a containing relationship, and the level depth.

[0093] The knowledge recommendation method based on space remote sensing graph provided by the embodiment of the application can realize multi-level association query of the keyword pair through the above designed multi-level query template, thereby improving the accuracy of the query result.

[0094] The knowledge recommendation device based on the space remote sensing graph provided by the present application is described below. The knowledge recommendation device based on the space remote sensing graph described below can be correspondingly referred to the knowledge recommendation method based on the space remote sensing graph described above.

[0095] The knowledge recommendation device based on the space remote sensing graph of the embodiment of the present application, as shown in Figure 3 includes the following modules: The first knowledge recommendation module 310 is configured to rewrite and split the query question input by the user according to the question template, and extract at least one keyword from the rewritten and split query question. The second knowledge recommendation module 320 is configured to determine a first target node matched with the at least one keyword in the space remote sensing graph. The space remote sensing graph is a knowledge graph constructed based on the entities, relationships and attributes extracted from the papers related to remote sensing data, and the supplementary attributes of regional entities extracted in combination with multi-source geographic data according to a preset ontology framework. The third knowledge recommendation module 330 is configured to obtain a query template matched with the at least one keyword, fill the first target node into the query template, and generate a query statement. The fourth knowledge recommendation module 340 is configured to obtain context data related to the query statement from the space remote sensing graph, and input the context data and the query question as prompt information into the generative large model to obtain the knowledge answer output by the generative large model.

[0096] The knowledge recommendation device based on the space remote sensing graph of the present embodiment can rewrite and split the query question according to the question template and extract the keyword, accurately positioning the core of the user question. Secondly, the construction of the space remote sensing graph integrates the entities, relationships, attributes in the remote sensing papers and multi-source geographic data, so that the originally crossed, scattered and redundant knowledge can be structured and presented, which facilitates the accurate determination of the first target node matched with the keyword. Finally, the first target node is filled into the query template to generate the query statement, which further clarifies the query direction. The context data related to the query statement obtained from the space remote sensing graph provides rich background knowledge for the generative large model, so that it can output accurate answers that meet the characteristics of the remote sensing field knowledge, and realizes the efficient recommendation of remote sensing knowledge.

[0097] Figure 4 An example of an entity structure diagram of an electronic device is shown in Figure 4As shown, the electronic device can include a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 complete mutual communication through the communications bus 440. The processor 410 can invoke a logical instruction in the memory 430 to execute a knowledge recommendation method based on a space remote sensing graph, which includes: In response to a query question input by a user, rewriting and splitting the query question according to a question template, and extracting at least one keyword from the rewritten and split query question; Determining a first target node matching the at least one keyword in a space remote sensing graph; the space remote sensing graph is a knowledge graph constructed according to a preset ontology framework based on entities, relationships, and attributes extracted from papers related to remote sensing data, and supplementary attributes of regional entities extracted in combination with multi-source geographic data; Obtaining a query template matching the at least one keyword, filling the first target node into the query template, and generating a query statement; Obtaining context data related to the query statement from the space remote sensing graph, and inputting the context data and the query question as prompt information to a generative large model to obtain a knowledge answer output by the generative large model.

[0098] In addition, the logical instruction in the memory 430 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and each kind of medium that can store program codes.

[0099] On the other hand, the present application also provides a computer program product, the computer program product includes a computer program, the computer program can be stored on a non-transitory computer readable storage medium, when the computer program is executed by a processor, the computer can execute the knowledge recommendation method based on the space remote sensing graph provided by each method, the method includes: In response to a query question input by a user, rewriting and splitting the query question according to a question template, and extracting at least one keyword from the query question after rewriting and splitting; determining a first target node matched with the at least one keyword in an aerospace remote sensing graph; the aerospace remote sensing graph is a knowledge graph constructed according to a preset ontology framework based on entities, relationships and attributes extracted from papers related to remote sensing data, and supplementary attributes of regional entities extracted in combination with multi-source geographic data; obtaining a query template matched with the at least one keyword, filling the first target node into the query template, and generating a query statement; obtaining context data related to the query statement from the aerospace remote sensing graph, and inputting the context data and the query question as prompt information into a generative large model to obtain a knowledge answer output by the generative large model.

[0100] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the knowledge recommendation method based on the aerospace remote sensing graph provided by each of the above methods, the method comprising: In response to a query question input by a user, rewriting and splitting the query question according to a question template, and extracting at least one keyword from the query question after rewriting and splitting; determining a first target node matched with the at least one keyword in an aerospace remote sensing graph; the aerospace remote sensing graph is a knowledge graph constructed according to a preset ontology framework based on entities, relationships and attributes extracted from papers related to remote sensing data, and supplementary attributes of regional entities extracted in combination with multi-source geographic data; obtaining a query template matched with the at least one keyword, filling the first target node into the query template, and generating a query statement; obtaining context data related to the query statement from the aerospace remote sensing graph, and inputting the context data and the query question as prompt information into a generative large model to obtain a knowledge answer output by the generative large model.

[0101] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement without creative labor.

[0102] Those skilled in the art can clearly understand the technical solutions of the present application from the above description of the embodiments, and the technical solutions can be implemented by means of software necessary for a general hardware platform, or by hardware. Based on such understanding, the technical solutions described above, which are essential or contribute to the prior art, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0103] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in each of the foregoing embodiments, or make equivalent replacement for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A knowledge recommendation method based on aerospace remote sensing atlas, characterized in that: include: In response to a query question input by a user, rewrite and split the query question according to a question template, and extract at least one keyword from the rewritten and split query question; Determining a first target node matching the at least one keyword in the aerospace remote sensing atlas; The aerospace remote sensing atlas is a knowledge graph constructed based on the entities, relationships and attributes extracted from papers related to remote sensing data, and supplementary attributes of regional entities extracted from multi-source geographic data, according to a preset ontology framework. Obtaining a query template matching the at least one keyword, filling the first target node into the query template, and generating a query statement; Context data related to the query statement is obtained from the aerospace remote sensing atlas, and the context data and the query question are input into the generative model as prompt information to obtain the knowledge answer output by the generative model.

2. The knowledge recommendation method based on aerospace remote sensing atlas according to claim 1, characterized in that: The determining of a first target node matching the at least one keyword in the aerospace remote sensing atlas includes: For each keyword, based on the embedding similarity between the first embedding vector of the keyword and the second embedding vector of each first node in the aerospace remote sensing atlas, a first target node matching the keyword is determined; the first node is a node of the same entity type as the keyword; Among them, when the entity type corresponding to the keyword is a task, the first target node matched by the keyword is determined based on the embedding similarity between the first embedding vector of the keyword and the second embedding vector of each second node in the aerospace remote sensing atlas; the second node includes a task node and a process node.

3. The knowledge recommendation method based on aerospace remote sensing atlas according to claim 1 is characterized in that: Obtaining a query template matching the at least one keyword includes: When the at least one keyword is a single keyword and there is no upper-level keyword for the at least one keyword, obtaining a first query template that matches the at least one keyword; When the at least one keyword is a single keyword and a parent keyword exists for the at least one keyword, a second query template matching the at least one keyword and a third query template matching the parent keyword of the at least one keyword are obtained.

4. The knowledge recommendation method based on aerospace remote sensing atlas according to claim 1, characterized in that: Obtaining a query template matching the at least one keyword includes: In a case where the at least one keyword is a first keyword pair, a fourth query template, a fifth query template, and a sixth query template that match the at least one keyword pair are obtained; the first keyword pair includes a target area and a task, the fourth query template is used to query papers that match the target area and the task, the fifth query template is used to query solutions in the papers, and the sixth query template is used to search for solutions in papers in an area similar to the target area when no papers belonging to the target area exist; In the case where the at least one keyword is a second keyword pair, a seventh query template, an eighth query template and a ninth query template that match the at least one keyword pair are obtained; the second keyword pair includes a target area and data, the seventh query template is used to query numerical data in the target area, the eighth query template is used to query whether the spatial range corresponding to the spatial data includes the target area, and the ninth query template is used to query whether there is an inclusion relationship between areas.

5. The knowledge recommendation method based on aerospace remote sensing atlas according to any one of claims 1 to 4, characterized in that: The aerospace remote sensing atlas is constructed in the following way: Obtain multiple papers related to remote sensing data and unify the text format of the papers; Inputting each paper in a unified text format and the first prompt information into the large language model to obtain a paper summary corresponding to each paper output by the large language model; the first prompt information includes the preset ontology framework and the remote sensing application methodology logic chain; Inputting the paper summary and the second prompt information corresponding to each paper into the large language model to obtain the entities, relationships, and attributes in each paper output by the large language model; The second prompt information includes the preset ontology framework and structured knowledge extraction prompt information; For the regional entity among the entities, extract the supplementary attributes of the regional entity by combining multi-source geographic data; Based on the entities, relationships, attributes in each paper and the supplementary attributes of the regional entities, an aerospace remote sensing map is constructed.

6. The knowledge recommendation method based on aerospace remote sensing atlas according to claim 5 is characterized in that: Based on the entities, relationships and attributes in each paper, an aerospace remote sensing map is constructed, including: For each entity in the paper, create entity-related labels and names; Based on the entity-related labels and names, determine whether the same second target node exists in the current graph database; If the second target node does not exist in the current graph database, create a new node corresponding to the entity; If the second target node exists in the current graph database, determining whether the second target node has an attribute corresponding to the entity; If the second target node does not have an attribute corresponding to the entity, assign the attribute corresponding to the entity to the second target node; If the second target node has an attribute corresponding to the entity, merge the attribute values; For each relationship in the paper, determine the head node and tail node corresponding to the relationship, establish an edge between the head node and the tail node, and assign the attribute corresponding to the relationship to the edge between the head node and the tail node.

7. The knowledge recommendation method based on aerospace remote sensing atlas according to any one of claims 1 to 6, characterized in that: The preset ontology framework is designed in the following way: Determine the paper as the first core entity type, and create paper-related authors and paper-related areas; Determine that the task is a second core entity type, create a paper-related task, and create a task-related process, a process-related data processing tool, and a data processing tool-related parameter; wherein the data processing tool includes at least one of an algorithm model or a software tool.

8. A knowledge recommendation device based on aerospace remote sensing atlas, characterized in that: include: A first knowledge recommendation module is configured to, in response to a query question input by a user, rewrite and split the query question according to a question template, and extract at least one keyword from the rewritten and split query question; A second knowledge recommendation module is configured to determine a first target node matching the at least one keyword in the aerospace remote sensing atlas; The aerospace remote sensing atlas is a knowledge graph constructed according to a preset ontology framework based on entities, relationships, and attributes extracted from papers related to remote sensing data, combined with regional attributes supplemented by multi-source geographic data; a third knowledge recommendation module, configured to obtain a query template matching the at least one keyword, fill the first target node into the query template, and generate a query statement; The fourth knowledge recommendation module is used to obtain context data related to the query statement from the aerospace remote sensing atlas, and input the context data and the query question as prompt information into the generative large model to obtain the knowledge answer output by the generative large model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the knowledge recommendation method based on aerospace remote sensing atlas according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the knowledge recommendation method based on aerospace remote sensing atlas according to any one of claims 1 to 7 is implemented.

Images