A method and system for enhancing meteorological map retrieval and generation

Abstract

This invention relates to the field of electronic digital data processing technology, and particularly to a method and system for enhancing meteorological map retrieval and generation. The method includes: processing meteorological text data to be processed to obtain candidate entities, candidate relationships, candidate attributes, and new schema types; performing semantic similarity verification and processing on the new schema types to form a hierarchical schema; generating a meteorological knowledge graph; performing multi-granularity community detection to generate community summaries, relationship path hints, and key attribute information; processing the user's natural language question and performing a first-stage meteorological domain-level retrieval to obtain and evaluate an initial answer; performing a second-stage iterative precise retrieval to generate and evaluate a final answer; and continuing until a satisfactory answer result is obtained or a preset number of iterations is reached. This invention improves knowledge organization capabilities, retrieval accuracy, evidence traceability, and reasoning reliability in complex question-answering tasks in the meteorological domain.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to a method and system for enhancing the generation of meteorological maps for retrieval. Background Technology

[0002] The large amount of specialized text data generated during meteorological operations is characterized by dense technical terminology, complex entity types and hierarchical structures, long event and response chains, and significant cross-system correlations of the same issue. When troubleshooting, analyzing the scope of impact, finding the basis for handling, or reviewing operational anomalies, operations and maintenance personnel often need to simultaneously understand the relationships between equipment, systems, products, alarms, fault phenomena, handling measures, and operational impacts. Large-scale language models demonstrate strong capabilities in natural language understanding and generation tasks; however, in meteorological operations and maintenance and meteorological operational support scenarios, relying solely on model parametric knowledge can easily lead to inaccurate understanding of meteorological concepts, untraceable factual basis, inconsistencies between generated content and current operational procedures, and a lack of evidence supporting fault handling recommendations. Retrieval-enhanced generation (RAG) technology, by introducing external knowledge bases, provides the model with updatable and verifiable contextual information, which can improve the factuality and reliability of question-and-answer results to some extent. However, traditional RAGs are mostly based on text fragment vector similarity recall, which is insufficient to fully express the multi-hop relationships between equipment topology, alarm causality, handling processes, product dependencies, and operational impacts in meteorological operational texts. Graph Retrieval Enhancement Generation (GraphRAG) introduces knowledge graphs into the retrieval enhancement generation framework. By explicitly modeling entities, relationships, and their semantic dependencies, it provides structured knowledge support for complex question answering, multi-hop reasoning, and evidence aggregation in meteorological operations and maintenance. However, the effectiveness of GraphRAG is highly dependent on the standardization and structural quality of the underlying knowledge graph. While commonly used open information extraction methods can automatically extract entities and relationships from meteorological operations and maintenance documents, fault tickets, and business specification texts, their results often suffer from inconsistent type granularity, ambiguous semantic boundaries, and significant relational noise, making it difficult to provide stable and reliable structural support for complex reasoning tasks in meteorological operations. Therefore, in meteorological question answering and multi-hop reasoning scenarios, it is necessary to introduce schemas to constrain the knowledge graph construction process, thereby standardizing meteorological entity types, relationship categories, and their structural mapping methods. Existing GraphRAG methods often use flat, static, predefined schemas, which are difficult to adapt to the dynamic changes in the distribution of meteorological business knowledge and operations and maintenance data. On the one hand, meteorological operational systems, observation equipment, monitoring indicators, alarm types, and meteorological products are constantly being updated. Fixed schemas struggle to keep up with new equipment models, operational subsystems, fault categories, and handling types, easily compressing specific semantics into coarse-grained types and causing information loss. On the other hand, schema expansion and maintenance typically rely on meteorological operational experts and IT operations and maintenance experts, which is costly and inconsistent. Furthermore, existing schemas often lack explicit hierarchical structures, making it difficult to simultaneously express abstract domain types and fine-grained entity or event types, limiting the overall effectiveness of meteorological knowledge graph construction, community organization, retrieval path selection, and subsequent reasoning generation. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a meteorological map retrieval enhancement generation method and system, which can adaptively evolve a hierarchical schema based on the distribution of meteorological text data for meteorological operation and maintenance and meteorological business support texts, and use the schema to guide the entire process of GraphRAG graph construction, community detection, retrieval and reasoning generation, thereby improving the knowledge organization ability, retrieval accuracy, evidence traceability and reasoning reliability in complex question answering tasks in the meteorological field.

[0004] This invention is achieved through the following technical solution: A meteorological map retrieval enhancement generation method includes the following steps: S1: Segment and preprocess the meteorological text data to be processed, and identify it under the current hierarchical schema constraints to obtain candidate entities, candidate relationships, candidate attributes and new schema types; S2: Perform semantic similarity verification between the new schema type and the existing schema type, and perform diversion, buffering, merging or promotion of the new schema type to complete the adaptive hierarchicalization of the schema type and form a hierarchical schema; perform type mapping on candidate entities, candidate relationships and candidate attributes according to the hierarchical schema to generate a meteorological knowledge graph with meteorological domain level or meteorological fine-grained level type annotation; S3: Perform multi-granularity community detection on the meteorological knowledge graph based on hierarchical schema to form meteorological domain-level communities, fine-grained type communities and semantic communities, and generate community summaries, relation path hints and key attribute information based on meteorological domain-level communities, fine-grained type communities and semantic communities; S4: When receiving a natural language question from a user regarding meteorological operation and maintenance or meteorological service support, the question is semantically parsed and decomposed into sub-questions. The meteorological domain layer type, meteorological fine-grained layer type, relationship path, and key attributes involved in the question are identified. Based on the community summary, the first stage of meteorological domain layer retrieval is performed to obtain an initial candidate evidence set and an initial answer. It is then determined whether the initial answer fully answers the user's question. If the initial answer meets the answer requirements, the answer is directly output. If the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, the process proceeds to the next step. S5: Perform the second-stage iterative precise retrieval based on the meteorological fine-grained layer type to obtain a supplementary evidence set. Merge, deduplicatize, and rearrange the initial candidate evidence set and the supplementary evidence set to form a structured meteorological knowledge evidence set. Generate the final answer based on the structured meteorological knowledge evidence set until the final answer meets the answer requirements or the iteration reaches the set number of times.

[0005] The optimized step S1, under the current hierarchical schema constraints, includes meteorological entity identification, meteorological business relationship extraction, attribute identification, and candidate type identification.

[0006] The optimized meteorological text data in step S1 includes meteorological equipment operation and maintenance manuals, observation station or radar station maintenance records, business system configuration instructions, monitoring and alarm records, fault work orders, handling plans, forecast and early warning service specifications, and meteorological product description texts.

[0007] Furthermore, the method for obtaining candidate entities, candidate relationships, candidate attributes, and new schema types in step S1 is as follows: S111: Perform text cleaning, format unification, and text segmentation on the meteorological text data to be processed. Divide each document in the meteorological text data into several text segments and retain the source mapping relationship between the text segments and the original document. S112: When performing identification under the current hierarchical schema constraints, the current hierarchical schema is modeled as a hierarchical semantic space that evolves with the data, and the first... The hierarchical schema for round extraction is represented by equation (1): (1); in: Indicates the first Hierarchical schema during round extraction Indicates the first The collection of abstract layer types during round extraction. Indicates the first The set of meteorological domain layer types during round extraction. Indicates the first The set of fine-grained meteorological types extracted in rounds. Indicates the first The set of meteorological business relationship types extracted in rounds. Indicates the first The set of attribute types extracted in round-robin; S113: Based on the text fragment and the current hierarchical schema, execute the extraction function according to equation (2) to obtain candidate entities, candidate relations, candidate attributes, and new schema types: (2); in: This represents the extraction function. Indicates that the document is The first segment obtained A text fragment, Indicates from text fragments The candidate entity set extracted from the data. Indicates from text fragments The candidate relation set extracted from the data. Indicates from text fragments The candidate attribute set extracted from the data. Indicates from text fragments The set of candidate new types extracted from the data.

[0008] Furthermore, in step S2, the adaptive hierarchicalization of the schema type is completed, and the method for forming a hierarchical schema is as follows: S211: First, determine whether the new Schema type generated in step S1 already exists in the current Schema. If the new Schema type already exists in the current Schema, then do not add the new Schema again. If the new Schema type does not exist in the current Schema, then calculate the maximum semantic similarity between the new Schema type and all current meteorological domain-level types according to equation (3): (3); in: Indicates the new schema type Maximum semantic similarity with all current meteorological domain-level types. Indicates an embedded function. Represents cosine similarity. Indicates the first The set of meteorological domain layer types during round extraction. express Elements in; S212: Compare the maximum semantic similarity with the adaptive threshold, and based on the comparison results, combine the three-pool type management and frequency-driven promotion mechanism to perform diversion, buffering, merging or promotion processing on the new schema type, progressively confirm the candidate type, complete the adaptive hierarchicalization of the schema type, and form a hierarchical schema.

[0009] Furthermore, in step S2, hierarchical type labels are assigned according to equation (4) to complete the type mapping, and the constructed knowledge graph is represented by equation (5): (4); (5); in: Indicates candidate entities Hierarchical type annotation, Representing entities L1 domain layer type, Representing entities L2 fine-grained type, Indicates the first Knowledge graph during round extraction Indicates the first The set of entity nodes during round extraction. Indicates the first The set of relationship edges between entities during round extraction. Indicates the first Hierarchical schema during round extraction.

[0010] Furthermore, the meteorological domain-level community formed in step S3 is represented by equation (6), and the generated community summary is represented by equation (7): (6); (7); in: Indicates that the L1 domain layer type is Domain-level community Representing entities L1 domain layer type annotation, Represents a set of entities. This represents a community summary. Indicates community The entity set in the middle, Represents a set of relationships within or related to a community. This represents the set of attributes corresponding to entities in the community. This represents the function that generates the summary.

[0011] Furthermore, the method for performing the first stage of meteorological domain-level retrieval based on community summaries in step S4 is as follows: S411: Based on the L1 meteorological domain layer type set, the first stage of retrieval is performed in the Schema type index to locate the candidate entity set as shown in equation (8): (8); in: This represents the set of candidate entities obtained in the first stage of retrieval. Representation type The corresponding set of entity nodes, This represents the set of meteorological domain layer types obtained in the first stage of retrieval. Represents the union; S412: Perform entity relation retrieval and one-hop neighbor expansion around the candidate entities to obtain relation triple evidence related to the candidate entities. At the same time, recall relevant original text fragments based on the semantic similarity between the question vector and the text fragment vector to form the initial candidate evidence set retrieved in the first stage as shown in Equation (9): (9); in: Represents the initial set of candidate evidence. This represents graph structure evidence obtained based on entity relationship paths. This indicates evidence from a triplet. This represents evidence from text fragments; S413: The initial set of candidate evidence is organized together with the L1 community summary, L1 relation path hints and key attribute information to form the initial retrieval context input to the large language model, generating the initial answer; S414: Determine whether the initial answer has fully answered the user's question. If the initial answer meets the answer requirements, output the initial answer directly. If the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, proceed to the next step.

[0012] Furthermore, the method for performing the second-stage iterative precise retrieval based on the meteorological fine-grained layer type in step S5 is as follows: S511: Perform the second-stage iterative precise retrieval based on the meteorological fine-grained layer type set, and locate the precise candidate entity set in the schema type index as shown in equation (10): (10); in: Represents the precise set of candidate entities. This represents the set of fine-grained meteorological types obtained in the second stage of retrieval; S512: Entity relation retrieval, triplet retrieval, and text fragment retrieval are performed around the precise candidate entity set. Combined with L2 community summaries, L2 relation path hints, and key attribute information, missing evidence is supplemented to obtain the supplementary evidence set as shown in Equation (11), until sufficient evidence is obtained, a reliable answer is generated, or the preset maximum number of iterations is reached. (11); in: This indicates a supplementary collection of evidence. This indicates the graph structure evidence obtained in the second phase of retrieval. This indicates the triplet evidence obtained in the second phase of the search. This refers to the text fragment evidence obtained during the second phase of retrieval.

[0013] A meteorological map retrieval enhancement generation system is used to execute a meteorological map retrieval enhancement generation method as described in any one of the above, which includes a data acquisition and initial identification module, a meteorological knowledge graph acquisition module, a community summary, a relationship path prompt and key attribute information acquisition module, a first-stage retrieval and evaluation module, and a second-stage precise retrieval and evaluation module. The data acquisition and initial identification module is used to segment and preprocess the meteorological text data to be processed, and to identify candidate entities, candidate relationships, candidate attributes and new schema types under the current hierarchical schema constraints. The meteorological knowledge graph acquisition module is used to perform semantic similarity verification between the new schema type and the existing schema type, and to perform diversion, buffering, merging or promotion of the new schema type to complete the adaptive hierarchicalization of the schema type and form a hierarchical schema; based on the hierarchical schema, the module performs type mapping on candidate entities, candidate relationships and candidate attributes to generate a meteorological knowledge graph with meteorological domain-level or meteorological fine-grained-level type annotations. The community summary, relationship path hint, and key attribute information acquisition module performs multi-granularity community detection on the meteorological knowledge graph based on a hierarchical schema, forming meteorological domain-level communities, fine-grained type communities, and semantic communities, and generates community summaries, relationship path hints, and key attribute information based on the meteorological domain-level communities, fine-grained type communities, and semantic communities; The first-stage retrieval and evaluation module is used to perform semantic parsing and sub-question decomposition on user natural language questions about meteorological operation and maintenance or meteorological business support. It identifies the meteorological domain layer type, fine-grained meteorological layer type, relationship path, and key attributes involved in the question. Based on the community summary, it performs the first-stage meteorological domain layer retrieval to obtain an initial set of candidate evidence and an initial answer. It then determines whether the initial answer adequately answers the user's question. If the initial answer meets the answer requirements, it directly outputs the answer; otherwise, if the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, the second-stage retrieval is performed. The second-stage precise retrieval and evaluation module performs a second-stage iterative precise retrieval based on the meteorological fine-grained layer type to obtain a supplementary evidence set. The initial candidate evidence set and the supplementary evidence set are then fused, deduplicated, and rearranged to form a structured meteorological knowledge evidence set. The final answer is generated based on the structured meteorological knowledge evidence set until the final answer meets the answer requirements or the iteration reaches the set number of times.

[0014] Beneficial effects of the invention: The meteorological map retrieval enhancement generation method and system provided by this invention have the following advantages: 1. Through adaptive hierarchical schema evolution, the generation of meteorological map retrieval enhancement no longer depends on static, flat predefined schemas. It can dynamically discover new types based on the data distribution of meteorological operation and maintenance and meteorological business support texts, and form a multi-level type system, thereby improving the ability of meteorological knowledge graph to express new equipment, new business systems, new alarm types, new meteorological products and fine-grained fault semantics.

[0015] 2. By adopting a strategy that combines initial broad-coverage retrieval at the meteorological domain level with iterative precise retrieval at the meteorological fine-grained level, we can first ensure the evidence coverage of complex meteorological issues, and then conduct more precise in-depth retrieval around specific equipment, alarms, faults, products or handling types when the answers are insufficient, thereby achieving a balance between recall scope, retrieval accuracy and inference efficiency.

[0016] 3. By combining community summaries, relationship path prompts, key attribute information, and a two-stage, two-evaluation mechanism, the risk of invalid retrieval and erroneous reasoning in meteorological operation and maintenance Q&A can be reduced, and the completeness, interpretability, and generation reliability of answers in fault root cause analysis, business impact assessment, handling process query, and forecast and early warning service knowledge Q&A can be improved. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0018] A meteorological map retrieval enhancement generation method, the overall flowchart of which is shown in Figure 1, specifically includes the following steps: S1: Segment and preprocess the meteorological text data to be processed, and identify it under the current hierarchical schema constraints to obtain candidate entities, candidate relationships, candidate attributes and new schema types; Specifically, the meteorological text data to be processed includes meteorological equipment operation and maintenance manuals, observation station or radar station maintenance records, business system configuration instructions, monitoring and alarm records, fault work orders, emergency response plans, forecast and early warning service specifications, and meteorological product description texts.

[0019] Specifically, identification under the current hierarchical schema constraints includes meteorological entity identification, meteorological business relationship extraction, attribute identification, and candidate type identification.

[0020] The system aims to extract knowledge from the aforementioned meteorological texts without requiring manual pre-setting of a complete meteorological schema, and to provide candidate entities, candidate relationships, candidate attributes, and candidate type information for subsequent adaptive hierarchical schema evolution and meteorological knowledge graph construction.

[0021] Specifically, the methods for obtaining candidate entities, candidate relationships, candidate attributes, and new schema types are as follows: S111: Perform text cleaning, format unification, and text segmentation on the meteorological text data to be processed. Divide each document in the meteorological text data into several text segments and retain the source mapping relationship between the text segments and the original document. During text segmentation, the system retains the source mapping relationship between text fragments and the original document, which can ensure that subsequent question and answer results can be traced back to the original evidence.

[0022] S112: When performing identification under the current hierarchical schema constraints, the current hierarchical schema is modeled as a hierarchical semantic space that evolves with the data, and the first... The hierarchical schema for round extraction is represented by equation (1): (1); in: Indicates the first Hierarchical schema during round extraction Indicates the first The abstract layer type set used during round extraction to represent the highest-level general categories; Indicates the first The set of meteorological domain types extracted in each round is relatively stable. Indicates the first The set of fine-grained meteorological types extracted in rounds. The specific types are gradually discovered from the data. Meteorological fine-grained types are linked to their corresponding meteorological domain-level types through parent-child relationships. The meteorological domain-level types are then linked to the abstraction layer, thus forming a hierarchical schema structure. Indicates the first The set of meteorological business relationship types extracted in rounds. Indicates the first The set of attribute types extracted in round-robin; When performing identification under the current hierarchical schema constraints, the system does not treat the schema as a static extraction template, but models the current schema as a hierarchical semantic space that evolves with the data, and can dynamically discover new types based on the data distribution of meteorological operation and maintenance and meteorological business support texts.

[0023] The system primarily utilizes L1 meteorological domain-level types and L2 meteorological fine-grained types to hierarchically label entities. The L1 meteorological domain-level types include observation equipment, operational systems, monitoring indicators, alarm events, fault phenomena, handling measures, meteorological products, and operational impacts. The L2 meteorological fine-grained types include radar station link anomalies, automatic station sensor failures, numerical forecast product delays, data transmission interruptions, alarm threshold violations, and emergency response steps. This provides unified constraints for subsequent knowledge representation, community organization, and hierarchical retrieval.

[0024] S113: Based on the text fragment and the current hierarchical schema, execute the extraction function according to equation (2) to obtain candidate entities, candidate relations, candidate attributes, and new schema types: (2); in: This represents the extraction function. Indicates that the document is The first segment obtained A text fragment, Indicates from text fragments The candidate entity set extracted from the data. Indicates from text fragments The candidate relation set extracted from the data. Indicates from text fragments The candidate attribute set extracted from the data. Indicates from text fragments The set of candidate new types extracted from the data.

[0025] Specifically, candidate entities can be represented as: ;in: Represents an entity, Representing entities L1 domain layer type, Representing entities L2 fine-grained type, Indicates the location of the text source corresponding to the entity.

[0026] The candidate relation can be represented as: ,in: Indicates a candidate relationship. Indicates the head entity. Indicates the tail entity. Indicates semantic relations. Indicates the text source corresponding to the relationship.

[0027] Candidate attributes can be represented as: ,in: Indicates the attribute name. Indicates the text source corresponding to the attribute.

[0028] By simultaneously obtaining candidate entities, candidate relationships, candidate attributes, and new schema types during text segmentation and knowledge extraction in step S1, the extraction results can not only be used for knowledge graph construction, but also provide data-driven input for subsequent adaptive hierarchical schema evolution.

[0029] S2: Perform semantic similarity verification between the new schema type and the existing schema type, and perform diversion, buffering, merging or promotion of the new schema type to complete the adaptive hierarchicalization of the schema type and form a hierarchical schema; perform type mapping on candidate entities, candidate relationships and candidate attributes according to the hierarchical schema to generate a meteorological knowledge graph with meteorological domain level or meteorological fine-grained level type annotation; Specifically, the method for achieving adaptive hierarchicalization of schema types and forming a hierarchical schema is as follows: S211: First, determine whether the new Schema type generated in step S1 already exists in the current Schema. If the new Schema type already exists in the current Schema, then do not add the new Schema again. If the new Schema type does not exist in the current Schema, then calculate the maximum semantic similarity between the new Schema type and all current meteorological domain-level types according to equation (3): (3); in: Indicates the new schema type Maximum semantic similarity with all current meteorological domain-level types. Indicates an embedded function. Represents cosine similarity. Indicates the first The set of meteorological domain layer types during round extraction. express Elements in; Similarity metrics can provide a unified semantic foundation for subsequent candidate type validation, schema evolution, community building, and hierarchical retrieval.

[0030] S212: Compare the maximum semantic similarity with the adaptive threshold, and based on the comparison results, combine the three-pool type management and frequency-driven promotion mechanism to perform diversion, buffering, merging or promotion processing on the new schema type, progressively confirm the candidate type, complete the adaptive hierarchicalization of the schema type, and form a hierarchical schema.

[0031] The three-pool type consists of a core pool, a candidate pool, and a rejection pool. The core pool stores validated and stable schema types; the candidate pool temporarily stores new types that have not yet been fully validated; and the rejection pool caches types that are judged to be semantically duplicated or nearly redundant. The candidate pool further includes an L2 candidate pool and an L1 candidate pool, corresponding to potential fine-grained types and potential new domain types, respectively.

[0032] Specifically, the methods for splitting, buffering, merging, or promoting are as follows: When a candidate type is highly similar to an existing L1 type, the system considers the candidate type to be a duplicate or near-synonymous expression of the existing type, de-lists it and writes it into the rejection pool, and merges the type labels in the current extraction results into the most similar existing type. This process can avoid schema type inflation caused by the language model outputting synonyms, aliases, or expression variants.

[0033] When a candidate type has moderate semantic similarity to an existing L1 type, the system considers it a potential fine-grained subtype under that L1 type and writes it into the L2 candidate pool. At this point, the system records the suggested parent type, frequency of occurrence, confidence level, and the times of first and most recent occurrence of the candidate type. The candidate type does not immediately become a formal L2 type; instead, it needs to be buffered and repeatedly appear in subsequent extraction processes, meeting the promotion criteria before being merged and promoted into the core pool.

[0034] When a candidate type has low similarity to all existing L1 meteorological domain types, the system considers that the type may represent a new meteorological operational semantic category and adds it to the L1 candidate pool. Compared with L2 candidate types, L1 candidate types are used to discover new meteorological domain layer categories. Their promotion criteria mainly rely on the frequency of occurrence across samples to avoid occasional noise being incorrectly solidified as new domain types.

[0035] The three-pool splitting of candidate types can be formally represented as: Let ,in Indicates candidate type Maximum semantic similarity with existing L1 types; when At that time, Write to the rejection pool; when At that time, Write to the L2 candidate pool; when At that time, Write to the L1 candidate pool, where Indicates the rejection threshold. Indicates the candidate threshold, and .

[0036] The threshold is not fixed, but can be adaptively updated based on the accumulated similarity distribution during the extraction process to adapt to different datasets and semantic densities in different fields.

[0037] Specifically, during the initial cold start phase, the system does not directly use manually preset candidate and rejection thresholds. Instead, during each candidate schema type verification, it calculates the maximum semantic similarity between the candidate type and existing L1 meteorological domain layer types, and writes the maximum semantic similarity as an unfiltered sample into the similarity sample set. During the cold start period, candidate types are temporarily stored in the candidate pool, and no immediate rejection or formal acceptance judgment is made based on fixed thresholds, in order to avoid the initial thresholds biasing subsequent learning results.

[0038] Once the accumulated number of similar samples reaches the preset cold start window, the system adaptively models the distribution of similar samples. Specifically, it first uses KMeans to initialize the clustering of one-dimensional similar samples, then uses the KMeans cluster centers as the initial mean of the Gaussian Mixture Model (GMM) to perform probabilistic clustering of similar samples, and selects the number of clusters with better clustering performance based on model evaluation metrics. The system semantically labels clusters according to their mean values ​​from low to high, where low similarity clusters correspond to novel types, low-to-medium similarity clusters correspond to weakly related types, medium similarity clusters correspond to potential parent-child hierarchical types, high similarity clusters correspond to co-occurring or near-synonymous types, and extremely high similarity clusters correspond to repetitive types.

[0039] After obtaining the semantic clustering results, an adaptive threshold is determined based on the boundaries of adjacent clusters. The candidate threshold is used to distinguish between novel types and weakly related types, and its value can be determined by the boundary between novel type clusters and weakly related type clusters. The rejection threshold is used to distinguish between potential parent-child hierarchical types and synonymous or repetitive types, and its value can be determined by the boundary between parent-child hierarchical type clusters and highly similar type clusters. The promotion interval is used to determine whether the candidate L2 type and the suggested parent L1 type constitute a reasonable parent-child semantic relationship, and its upper and lower bounds can be determined by the statistical boundaries of parent-child hierarchical type clusters.

[0040] If the clustering confidence of the Gaussian mixture model is lower than a preset confidence threshold, a data-driven threshold can be generated based on quantiles. Specifically, higher quantiles of similar samples can be used as rejection thresholds, lower quantiles as candidate thresholds, and the middle quantile interval as promotion intervals. For example, the 85th percentile can be used as the rejection threshold, the 20th percentile as the candidate threshold, and the interval from the 20th to the 85th percentile as the confidence interval for promoting a candidate type to L2 type. These quantile thresholds are derived from the current data distribution, rather than fixed empirical values.

[0041] After the cold start is complete, the system enters the normal adaptive update phase. Each time a new candidate type completes validation, the system records the similarity and triage result for that candidate type. When the cumulative number of validations reaches the preset update frequency, the system relearns new candidate thresholds, rejection thresholds, and promotion intervals based on historical similarity samples. To avoid sudden changes in thresholds due to local sample fluctuations, a smooth update method can be used to update the thresholds. This involves weighting and fusing the newly learned threshold with the currently used threshold according to a smoothing coefficient to obtain the effective threshold for the next round.

[0042] For candidate types entering the L2 candidate pool, the system further performs a local deduplication check. Specifically, the system determines its most similar L1 type as the suggested parent type and calculates the maximum similarity between the candidate type and existing L2 subtypes under the parent type. When a candidate type is highly similar to an existing L2 subtype, it indicates that it may be a duplicate expression of an existing fine-grained type, and the system writes it into the rejection pool; otherwise, it is retained in the L2 candidate pool to await subsequent frequency-driven promotion.

[0043] When a candidate type reappears during subsequent extractions, the system updates its candidate pool record, including the number of occurrences, the most recent occurrence time, confidence level, or maximum similarity. For L2 candidate types, when their occurrence count reaches a preset or adaptively determined frequency threshold, and their semantic consistency with the suggested parent type is within a set range, the system promotes them to a core L2 type and establishes a parent-child relationship between them and the corresponding L1 parent type. For L1 candidate types, when their occurrence count meets the L1 promotion criteria, the system promotes them to a core L1 type and attaches them to the L0 abstraction layer.

[0044] Furthermore, in step S2, hierarchical type labels are assigned according to equation (4) to complete the type mapping, and the constructed knowledge graph is represented by equation (5): (4); (5); in: Indicates candidate entities Hierarchical type annotation, Representing entities L1 domain layer type, Representing entities L2 fine-grained type, Indicates the first Knowledge graph during round extraction Indicates the first The set of entity nodes during round extraction. Indicates the first The set of relationship edges between entities during round extraction. Indicates the first Hierarchical schema during round extraction.

[0045] The entity node set records its name, source text fragment, and L1 or L2 type annotation; the relation edge records the semantic relationship between entities; the attribute node or attribute field records the attribute information corresponding to the entity.

[0046] If the type corresponding to the entity can only be mapped to L1, then It can be empty; if the candidate type corresponding to the entity is rejected, the system maps it to the most similar existing type; if the candidate type is in the candidate pool, the system can retain its candidate fine-grained label in the current graph construction and further confirm it in the subsequent schema stabilization process.

[0047] The core of this step is that the system does not immediately solidify the new types that appear during the extraction process into formal schema types. Instead, it gradually confirms the candidate types through semantic similarity verification, three-pool type management, and frequency-driven promotion mechanism, thereby suppressing the growth of noisy and repetitive types while maintaining the open extensibility of the schema.

[0048] S3: Perform multi-granularity community detection on the meteorological knowledge graph based on hierarchical schema to form meteorological domain-level communities, fine-grained type communities and semantic communities, and generate community summaries, relation path hints and key attribute information based on meteorological domain-level communities, fine-grained type communities and semantic communities; The system forms meteorological domain-level communities based on the L1 meteorological domain layer type of entity nodes. Entity nodes with the same or related L1 meteorological domain layer types are aggregated into the same meteorological domain-level community, such as observation equipment community, business system community, alarm event community, fault handling community, meteorological product community, and business impact community, to provide a wider range of background semantics and candidate search scope.

[0049] After obtaining the meteorological domain-level community, the system further constructs fine-grained type communities based on the L2 meteorological fine-grained layer type of the entity nodes. Entity nodes with the same L2 type or belonging to related L2 types under the same L1 parent type are organized into more fine-grained sub-communities, such as radar station link anomaly community, automatic weather station sensor failure community, numerical weather prediction product delay community, and early warning issuance process community, to support accurate retrieval in subsequent iterative inference stages. Compared with L1 communities, L2 communities have a smaller retrieval scope and higher semantic consistency, making them suitable for deep localization of specific equipment roles, failure event categories, business products, or key attribute types in complex meteorological problems.

[0050] Furthermore, the system can also perform semantic community partitioning by combining the topological structure and semantic relationships in the knowledge graph. Semantic communities do not rely entirely on node types, but rather comprehensively consider the relationship edges between entities, connection density, path structure, and semantic similarity, aggregating closely related or semantically relevant entities into communities. Through the combination of domain-level communities, fine-grained type communities, and semantic communities, the system can simultaneously preserve schema-level constraints and real-world association patterns in the graph structure.

[0051] Specifically, the meteorological domain-level community is represented by equation (6), and the generated community summary is represented by equation (7): (6); (7); in: Indicates that the L1 domain layer type is Domain-level community Representing entities L1 domain layer type annotation, Represents a set of entities. This represents a community summary. Indicates community The entity set in the middle, Represents a set of relationships within or related to a community. This represents the set of attributes corresponding to entities in the community. This represents the function that generates the summary.

[0052] Community summaries are used to summarize the main entities, core relationships, typical attributes, and semantic themes in the community, which can be injected into the large language model as contextual information in subsequent retrieval stages.

[0053] Simultaneously, the system generates relational path hints based on the relationships within and between communities. These hints describe frequently occurring or inferentially valuable relational patterns within a specific L1 or L2 type, such as which relationships typically connect a certain type of entity to another. This information assists in selecting inference paths during subsequent complex question answering, reducing the probability of accessing irrelevant relationships during the retrieval phase.

[0054] The system also extracts key attribute information from the community. Key attributes can be high-frequency attributes, attributes highly relevant to the question type, or attributes that contribute significantly to entity differentiation and answer generation. This type of attribute information can be used as cue words to enhance content in subsequent question-answering stages, guiding the large language model to focus on key evidence such as time, location, identity, quantity, and category.

[0055] This step further organizes the hierarchical knowledge graph obtained in step S2 into a multi-granularity index structure oriented towards retrieval and reasoning. This structure retains the semantic constraints brought by the L1 or L2 schema hierarchy, while leveraging community summaries, relational path hints, and key attribute information to enhance the contextual organization capabilities of the subsequent question-answering stage, providing support for two-stage retrieval and multi-hop reasoning.

[0056] S4: When receiving a natural language question from a user regarding meteorological operation and maintenance or meteorological service support, the question is semantically parsed and decomposed into sub-questions. The meteorological domain layer type, meteorological fine-grained layer type, relationship path, and key attributes involved in the question are identified. Based on the community summary, the first stage of meteorological domain layer retrieval is performed to obtain an initial candidate evidence set and an initial answer. It is then determined whether the initial answer fully answers the user's question. If the initial answer meets the answer requirements, the answer is directly output. If the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, the process proceeds to the next step. When a user asks a natural language question about meteorological operations and maintenance or meteorological service support, the question is semantically parsed and decomposed into sub-questions to obtain a structured question parsing result. for For complex multi-hop problems, the system decomposes the original problem into several subproblems. ,in: This represents the set of subproblems obtained by decomposing the original problem. Represents a set of meteorological domain types. Represents a set of meteorological fine-grained layer types. Represents a set of relation types. Represents a set of key attributes. Indicates the first The issue of height. Indicates the number of subproblems.

[0057] Furthermore, the method for performing the first stage of meteorological domain-level retrieval based on community summaries in step S4 is as follows: S411: Based on the L1 meteorological domain layer type set, the first stage of retrieval is performed in the Schema type index to locate the candidate entity set as shown in equation (8): (8); in: This represents the set of candidate entities obtained in the first stage of retrieval. Representation type The corresponding set of entity nodes, This represents the set of meteorological domain layer types obtained in the first stage of retrieval. Represents the union; S412: Perform entity relation retrieval and one-hop neighbor expansion around the candidate entities to obtain relation triple evidence related to the candidate entities. At the same time, recall relevant original text fragments based on the semantic similarity between the question vector and the text fragment vector to form the initial candidate evidence set retrieved in the first stage as shown in Equation (9): (9); in: Represents the initial set of candidate evidence. This represents graph structure evidence obtained based on entity relationship paths. This indicates evidence from a triplet. This represents evidence from text fragments; The one-hop neighbor expansion means starting from the candidate entity that has already been found, taking only one step along the relation edges in the knowledge graph, and finding the entities and relations that it is directly connected to.

[0058] S413: The initial set of candidate evidence is organized together with the L1 community summary, L1 relation path hints and key attribute information to form the initial retrieval context input to the large language model, generating the initial answer; S414: Determine whether the initial answer has fully answered the user's question. If the initial answer meets the answer requirements, output the initial answer directly. If the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, proceed to the next step.

[0059] In the first stage of retrieval, the system performs a broad-coverage retrieval based on the L1 meteorological domain layer type set obtained from question parsing. This stage recalls entities, relations, triples, and text fragments related to the question from the meteorological knowledge graph. Since L1 types represent a broad range of semantic categories in the meteorological domain, this stage can cover a large candidate range, reducing the risk of missing key evidence in the initial retrieval of complex meteorological questions.

[0060] S5: Perform the second-stage iterative precise retrieval based on the meteorological fine-grained layer type to obtain a supplementary evidence set. Merge, deduplicatize, and rearrange the initial candidate evidence set and the supplementary evidence set to form a structured meteorological knowledge evidence set. Generate the final answer based on the structured meteorological knowledge evidence set until the final answer meets the answer requirements or the iteration reaches the set number of times.

[0061] Furthermore, the method for performing the second-stage iterative precise retrieval based on the meteorological fine-grained layer type in step S5 is as follows: S511: Perform the second-stage iterative precise retrieval based on the meteorological fine-grained layer type set, and locate the precise candidate entity set in the schema type index as shown in equation (10): (10); in: Represents the precise set of candidate entities. This represents the set of fine-grained meteorological types obtained in the second stage of retrieval; S512: Entity relation retrieval, triplet retrieval, and text fragment retrieval are performed around the precise candidate entity set. Combined with L2 community summaries, L2 relation path hints, and key attribute information, missing evidence is supplemented to obtain the supplementary evidence set as shown in Equation (11), until sufficient evidence is obtained, a reliable answer is generated, or the preset maximum number of iterations is reached. (11); in: This indicates a supplementary collection of evidence. This indicates the graph structure evidence obtained in the second phase of retrieval. This indicates the triplet evidence obtained in the second phase of the search. This refers to the text fragment evidence obtained during the second phase of retrieval.

[0062] During the iterative reasoning process, if the model determines that more information is still needed, the system generates a new retrieval query based on the current reasoning state and continues to execute the second stage of iterative precise retrieval. This process can be repeated until sufficient evidence is obtained, a reliable answer is generated, or the preset maximum number of iterations is reached.

[0063] The system selects an answer generation strategy based on the execution of iterative reasoning. If the iterative reasoning process has already generated an answer with a final answer identifier, the system extracts candidate answers from the iterative reasoning results and performs a confirmation evaluation on the candidate answers. The confirmation evaluation is used to determine whether the candidate answers have uncorrected logical errors, numerical errors, comparison errors, contradictory evidence, or incomplete conclusions.

[0064] When the evaluation result is confirmed as passed, the system outputs the candidate answers obtained during the iterative reasoning process as the final answer. When the evaluation result is confirmed as failed, or the system fails to extract valid candidate answers from the iterative reasoning process, the system reconstructs the prompt words based on the final rearranged reasoning context and calls the large language model to regenerate the answer. If the iterative reasoning reaches the preset maximum number of rounds, the system also generates and outputs the current optimal answer based on the final rearranged reasoning context.

[0065] In the second stage of retrieval, the system performs precise retrieval based on the L2 meteorological fine-grained layer type set obtained from question parsing. Unlike the first stage, L2 types represent more specific meteorological equipment, business systems, alarm events, fault categories, meteorological products, or handling procedures. Therefore, this stage can narrow the search scope and improve the matching degree between candidate evidence and question semantics.

[0066] By performing a two-stage retrieval process in steps S4 and S5, and evaluating each step separately, a two-stage hierarchical retrieval workflow was formed, prioritizing the meteorological domain layer and deepening the retrieval at the fine-grained meteorological layer. The first stage ensures evidence coverage through L1 meteorological domain layer types, while the second stage improves retrieval accuracy through L2 meteorological fine-grained layer types, and enhances the reasoning context by incorporating community summaries, relational path hints, and key attribute information. This approach effectively supports complex problems and multi-hop reasoning tasks in meteorological operations and maintenance and meteorological service support, reduces interference from irrelevant evidence, and improves the reliability of subsequent answer generation.

[0067] This approach allows for direct output of results when the initial answer is sufficient, and iterative retrieval to accumulate meteorological textual and graph structural evidence when the initial answer is insufficient. A final confirmation and evaluation process determines whether to use iterative reasoning or regenerate the answer. This mechanism reduces unnecessary computational overhead and mitigates the impact of erroneous reasoning, insufficient evidence, or inaccurate citations of operational procedures on the final answer in meteorological question-and-answer scenarios. This improves the accuracy, completeness, and reliability of answer generation in root cause analysis, operational impact assessment, handling process queries, and meteorological product and service question-and-answer scenarios.

[0068] A meteorological map retrieval enhancement generation system is used to execute a meteorological map retrieval enhancement generation method as described in any one of the above, which includes a data acquisition and initial identification module, a meteorological knowledge graph acquisition module, a community summary, a relationship path prompt and key attribute information acquisition module, a first-stage retrieval and evaluation module, and a second-stage precise retrieval and evaluation module. The data acquisition and initial identification module is used to segment and preprocess the meteorological text data to be processed, and to identify candidate entities, candidate relationships, candidate attributes and new schema types under the current hierarchical schema constraints. The meteorological knowledge graph acquisition module is used to perform semantic similarity verification between the new schema type and the existing schema type, and to perform diversion, buffering, merging or promotion of the new schema type to complete the adaptive hierarchicalization of the schema type and form a hierarchical schema; based on the hierarchical schema, the module performs type mapping on candidate entities, candidate relationships and candidate attributes to generate a meteorological knowledge graph with meteorological domain-level or meteorological fine-grained-level type annotations. The community summary, relationship path hint, and key attribute information acquisition module performs multi-granularity community detection on the meteorological knowledge graph based on a hierarchical schema, forming meteorological domain-level communities, fine-grained type communities, and semantic communities, and generates community summaries, relationship path hints, and key attribute information based on the meteorological domain-level communities, fine-grained type communities, and semantic communities; The first-stage retrieval and evaluation module is used to perform semantic parsing and sub-question decomposition on user natural language questions about meteorological operation and maintenance or meteorological business support. It identifies the meteorological domain layer type, fine-grained meteorological layer type, relationship path, and key attributes involved in the question. Based on the community summary, it performs the first-stage meteorological domain layer retrieval to obtain an initial set of candidate evidence and an initial answer. It then determines whether the initial answer adequately answers the user's question. If the initial answer meets the answer requirements, it directly outputs the answer; otherwise, if the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, the second-stage retrieval is performed. The second-stage precise retrieval and evaluation module performs a second-stage iterative precise retrieval based on the meteorological fine-grained layer type to obtain a supplementary evidence set. The initial candidate evidence set and the supplementary evidence set are then fused, deduplicated, and rearranged to form a structured meteorological knowledge evidence set. The final answer is generated based on the structured meteorological knowledge evidence set until the final answer meets the answer requirements or the iteration reaches the set number of times.

[0069] In summary, the meteorological map retrieval enhancement generation method and system provided by this invention not only reduces unnecessary computational overhead but also mitigates the impact of erroneous reasoning, insufficient evidence, or inaccurate citation of business procedures on the final answer in meteorological question-and-answer scenarios. This improves the accuracy, completeness, and reliability of answer generation in scenarios such as root cause analysis of faults, assessment of business impacts, query of handling procedures, and question-and-answer of meteorological products and services.

[0070] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for enhancing the generation of meteorological maps for retrieval, characterized in that: Includes the following steps: S1: Segment and preprocess the meteorological text data to be processed, and identify it under the current hierarchical schema constraints to obtain candidate entities, candidate relationships, candidate attributes and new schema types; S2: Perform semantic similarity verification between the new schema type and the existing schema type, and perform traffic splitting, buffering, merging or promotion of the new schema type to complete the adaptive hierarchicalization of the schema type and form a hierarchical schema; Based on the hierarchical schema, candidate entities, candidate relationships, and candidate attributes are type-mapped to generate a meteorological knowledge graph with meteorological domain-level or meteorological fine-grained-level type annotations. S3: Perform multi-granularity community detection on the meteorological knowledge graph based on hierarchical schema to form meteorological domain-level communities, fine-grained type communities and semantic communities, and generate community summaries, relation path hints and key attribute information based on meteorological domain-level communities, fine-grained type communities and semantic communities; S4: When receiving a natural language question from a user regarding meteorological operation and maintenance or meteorological service support, the question is semantically parsed and decomposed into sub-questions. The meteorological domain layer type, meteorological fine-grained layer type, relationship path, and key attributes involved in the question are identified. Based on the community summary, the first stage of meteorological domain layer retrieval is performed to obtain an initial candidate evidence set and an initial answer. It is then determined whether the initial answer fully answers the user's question. If the initial answer meets the answer requirements, the answer is directly output. If the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, the process proceeds to the next step. S5: Perform the second-stage iterative precise retrieval based on the meteorological fine-grained layer type to obtain a supplementary evidence set. Merge, deduplicatize, and rearrange the initial candidate evidence set and the supplementary evidence set to form a structured meteorological knowledge evidence set. Generate the final answer based on the structured meteorological knowledge evidence set until the final answer meets the answer requirements or the iteration reaches the set number of times.

2. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: Step S1 involves identification under the current hierarchical schema constraints, including meteorological entity identification, meteorological business relationship extraction, attribute identification, and candidate type identification.

3. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: The meteorological text data mentioned in step S1 includes meteorological equipment operation and maintenance manuals, observation station or radar station maintenance records, business system configuration instructions, monitoring and alarm records, fault work orders, handling plans, forecast and early warning service specifications, and meteorological product description texts.

4. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: The method for obtaining candidate entities, candidate relationships, candidate attributes, and new schema types in step S1 is as follows: S111: Perform text cleaning, format unification, and text segmentation on the meteorological text data to be processed. Divide each document in the meteorological text data into several text segments and retain the source mapping relationship between the text segments and the original document. S112: When performing identification under the current hierarchical schema constraints, the current hierarchical schema is modeled as a hierarchical semantic space that evolves with the data, and the first... The hierarchical schema for round extraction is represented by equation (1): （1）； in: Indicates the first Hierarchical schema during round extraction Indicates the first The collection of abstract layer types during round extraction. Indicates the first The set of meteorological domain layer types during round extraction. Indicates the first The set of fine-grained meteorological types extracted in rounds. Indicates the first The set of meteorological business relationship types extracted in rounds. Indicates the first The set of attribute types extracted in round-robin; S113: Based on the text fragment and the current hierarchical schema, execute the extraction function according to equation (2) to obtain candidate entities, candidate relations, candidate attributes, and new schema types: （2）； in: This represents the extraction function. Indicates that the document is The first segment obtained A text fragment, Indicates from text fragments The candidate entity set extracted from the data. Indicates from text fragments The candidate relation set extracted from the data. Indicates from text fragments The candidate attribute set extracted from the data. Indicates from text fragments The set of candidate new types extracted from the data.

5. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: The method for completing the adaptive hierarchical transformation of the schema type in step S2, forming a hierarchical schema, is as follows: S211: First, determine whether the new Schema type generated in step S1 already exists in the current Schema. If the new Schema type already exists in the current Schema, then do not add the new Schema again. If the new Schema type does not exist in the current Schema, then calculate the maximum semantic similarity between the new Schema type and all current meteorological domain-level types according to equation (3): （3）； in: Indicates the new schema type Maximum semantic similarity with all current meteorological domain-level types. Indicates an embedded function. Represents cosine similarity. Indicates the first The set of meteorological domain layer types during round extraction. express Elements in; S212: Compare the maximum semantic similarity with the adaptive threshold, and based on the comparison results, combine the three-pool type management and frequency-driven promotion mechanism to perform diversion, buffering, merging or promotion processing on the new schema type, progressively confirm the candidate type, complete the adaptive hierarchicalization of the schema type, and form a hierarchical schema.

6. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: In step S2, hierarchical type labels are assigned according to equation (4) to complete the type mapping. The constructed knowledge graph is represented by equation (5): （4）； （5）； in: Indicates candidate entities Hierarchical type annotation, Representing entities L1 domain layer type, Representing entities L2 fine-grained type, Indicates the first Knowledge graph during round extraction Indicates the first The set of entity nodes extracted in round 2. Indicates the first The set of relationship edges between entities during round extraction. Indicates the first Hierarchical schema during round extraction.

7. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: The meteorological domain-level community formed in step S3 is represented by equation (6), and the generated community summary is represented by equation (7): （6）； （7）； in: Indicates that the L1 domain layer type is Domain-level community Representing entities L1 domain layer type annotation, Represents a set of entities. This represents a community summary. Indicates community The entity set in the middle, Represents a set of relationships within or related to a community. This represents the set of attributes corresponding to entities in the community. This represents the function that generates the summary.

8. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: The method for performing the first stage of meteorological domain-level retrieval based on community summaries in step S4 is as follows: S411: Based on the L1 meteorological domain layer type set, the first stage of retrieval is performed in the Schema type index to locate the candidate entity set as shown in equation (8): （8）； in: This represents the set of candidate entities obtained in the first stage of retrieval. Representation type The corresponding set of entity nodes, This represents the set of meteorological domain layer types obtained in the first stage of retrieval. Represents the union; S412: Perform entity relation retrieval and one-hop neighbor expansion around the candidate entities to obtain relation triple evidence related to the candidate entities. At the same time, recall relevant original text fragments based on the semantic similarity between the question vector and the text fragment vector to form the initial candidate evidence set retrieved in the first stage as shown in Equation (9): （9）； in: Represents the initial set of candidate evidence. This represents graph structure evidence obtained based on entity relationship paths. This indicates evidence from a triplet. This represents evidence from text fragments; S413: The initial set of candidate evidence is organized together with the L1 community summary, L1 relation path hints and key attribute information to form the initial retrieval context input to the large language model, generating the initial answer; S414: Determine whether the initial answer has fully answered the user's question. If the initial answer meets the answer requirements, output the initial answer directly. If the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, proceed to the next step.

9. The meteorological map retrieval enhancement generation method according to claim 1, characterized in that: The method for performing the second-stage iterative precise retrieval based on the meteorological fine-grained layer type in step S5 is as follows: S511: Perform the second-stage iterative precise retrieval based on the meteorological fine-grained layer type set, and locate the precise candidate entity set in the schema type index as shown in equation (10): (10)； in: Represents the precise set of candidate entities. This represents the set of fine-grained meteorological types obtained in the second stage of retrieval; S512: Entity relation retrieval, triplet retrieval, and text fragment retrieval are performed around the precise candidate entity set. Combined with L2 community summaries, L2 relation path hints, and key attribute information, missing evidence is supplemented to obtain the supplementary evidence set as shown in Equation (11), until sufficient evidence is obtained, a reliable answer is generated, or the preset maximum number of iterations is reached. （11）； in: This indicates a supplementary collection of evidence. This indicates the graph structure evidence obtained in the second phase of retrieval. This refers to the triplet evidence obtained in the second phase of the search. This refers to the text fragment evidence obtained during the second phase of retrieval.

10. A meteorological map retrieval enhancement generation system, characterized in that: The method for performing a meteorological map retrieval enhancement generation method as described in any one of claims 1 to 9 includes a data acquisition and initial identification module, a meteorological knowledge graph acquisition module, a community summary, a relationship path prompt and key attribute information acquisition module, a first-stage retrieval and evaluation module, and a second-stage precise retrieval and evaluation module. The data acquisition and initial identification module is used to segment and preprocess the meteorological text data to be processed, and to identify candidate entities, candidate relationships, candidate attributes and new schema types under the current hierarchical schema constraints. The meteorological knowledge graph acquisition module is used to perform semantic similarity verification between the new schema type and the existing schema type, and to perform diversion, buffering, merging or promotion of the new schema type to complete the adaptive hierarchicalization of the schema type and form a hierarchical schema. Based on the hierarchical schema, candidate entities, candidate relationships, and candidate attributes are type-mapped to generate a meteorological knowledge graph with meteorological domain-level or meteorological fine-grained-level type annotations. The community summary, relationship path hint, and key attribute information acquisition module performs multi-granularity community detection on the meteorological knowledge graph based on a hierarchical schema, forming meteorological domain-level communities, fine-grained type communities, and semantic communities, and generates community summaries, relationship path hints, and key attribute information based on the meteorological domain-level communities, fine-grained type communities, and semantic communities; The first-stage retrieval and evaluation module is used to perform semantic parsing and sub-question decomposition on user natural language questions about meteorological operation and maintenance or meteorological business support. It identifies the meteorological domain layer type, fine-grained meteorological layer type, relationship path, and key attributes involved in the question. Based on the community summary, it performs the first-stage meteorological domain layer retrieval to obtain an initial set of candidate evidence and an initial answer. It then determines whether the initial answer adequately answers the user's question. If the initial answer meets the answer requirements, it directly outputs the answer; otherwise, if the initial answer has insufficient information, missing evidence, incomplete reasoning, or logical uncertainty, the second-stage retrieval is performed. The second-stage precise retrieval and evaluation module performs a second-stage iterative precise retrieval based on the meteorological fine-grained layer type to obtain a supplementary evidence set. The initial candidate evidence set and the supplementary evidence set are then fused, deduplicated, and rearranged to form a structured meteorological knowledge evidence set. The final answer is generated based on the structured meteorological knowledge evidence set until the final answer meets the answer requirements or the iteration reaches the set number of times.

Images